Artificial Intelligence, Education, and EntrepreneurshipOctober 26,
2020
Abstract
The scarcity of the human capital needed for R&D in Artificial
Intelligence (AI) created an unprecedented brain drain of AI
professors from universities in 2004-2018. We exploit this brain
drain as a source of variation in students’ domain-specific knowl-
edge and provide causal evidence that domain-specific knowledge is
important for startup formation, seed funding, round A funding, and
funding growth rate. The effect is the largest for top
universities, PhD students, and startups in the area of deep learn-
ing. These results contribute to the entrepreneurial finance
literature and challenge the current view that entrepreneurs are
jacks-of-all-trades, master of none (Lazear 2004).
Keywords: Artificial intelligence, startup financing,
entrepreneurship, human capital, innovation, education, deep
learning
∗We are grateful to Ron Kaniel, Adrien Matray, Cade Metz, Michael
Roach, Scott Stern, Zvi Wiener, participants at 2019 Stanford
HAI-AI Index Roundtable Workshop, and seminar participants at the
Hebrew University and the University of Alberta for their helpful
comments and discussions. Michael Gofman acknowledges the Alfred P.
Sloan Foundation for financial support (Grant number:
G-2020-14005). Zhao Jin acknowledges the Ewing Marion Kauffman
Foundation for financial support (Grant number: G-201806-4434) and
Crunchbase for access to its database. All errors are our
own.
[email protected], University of Rochester.
[email protected], Cheung Kong Graduate School of
Business.
1 Introduction
Startups and entrepreneurship are important for innovation,
employment, and economic growth (e.g., Schumpeter 1942; King and
Levine 1993). It is important to understand what are the main
determinants of entrepreneurship. In this paper, we study the role
of founders’ domain-specific knowledge on the formation and funding
of startups. Ex-ante, the role of domain-specific knowl- edge for
startup success is not clear. Some founders of the tech startups
that turned into the most valuable companies in the world are
college dropouts (e.g., Bill Gates, Steve Jobs, and Mark
Zuckerberg), while others took advanced computer science classes
and excelled in their academic studies (e.g., Jeff Bezos, Larry
Page, and Sergey Brin). The existing literature emphasizes the role
of general education because entrepreneurs need to be
jacks-of-all-trades, master of none (Lazear 2004).1 The main
contribution of this paper is to provide a causal evidence that
formal education that develops specialized knowledge helps students
to establish startups and to attract funding from VCs.
Our analysis is based on a novel sample of 177 AI startups founded
by 363 AI entrepreneurs. It is especially important for AI startups
to understand factors that increase entrepreneurship and help to
attract VC funding because of their exponential growth and high
potential for creative destruc- tion. According to Perrault et al.
(2019), investments in global AI startups reached a record of $40B
in 2018, representing a nearly thirty-fold growth from 2010 to 2018
with an average annual growth rate of over 48%. AI becomes one of
the most promising and disruptive General Purpose Technol- ogy
(GPT) with a potential to change every aspect of our lives and spur
economic growth (Aghion et al. 2017; Trajtenberg 2018; Acemoglu and
Restrepo 2018; Cockburn et al. 2018). The largest companies and
countries in the world fight for leadership in AI. Sundar Pichai,
CEO of Google, said in 2016, “In the next 10 years, we will shift
to a world that is AI-first.” In February 2019, the White House
issued an executive order titled, “Maintaining American Leadership
in AI”.2 Over a short period of time, AI has become “a key driver
of the Fourth Industrial Revolution”.3 We study startups that are
at the forefront of this revolution.
To identify a causal effect of domain-specific knowledge on the
probability of starting a com- pany and attracting funding is quite
challenging, mainly due to the difficulty in measuring the
variation in the domain-specific knowledge that students receive
during their studies at the univer-
1Education could also be purely a costly signaling device without
any intrinsic value (Spence 1973). 2https://www.whitehouse.gov/ai
3World Economic Forum
https://www.weforum.org/projects/reimagining-regulation-for-the-age-of-ai
1
sity. To overcome this challenge, we use an unprecedented brain
drain of AI faculty from academia to industry as a source of
variation in students’ AI-specific knowledge. The significant pay
gap be- tween industry and academia makes it virtually impossible
for universities to retain their best AI professors. Top AI
researchers, such as Geoffrey Hinton and Yann LeCun who won the
2018 Tur- ing Award, were hired by Google and Facebook respectively
to lead their AI projects. Geoffrey Hinton’s student, Ilya
Sutskever, got paid more than $1.9 million by OpenAI in 2016 (Metz
2018a). In addition to the unmatchable compensation package,
industry jobs attract AI professors due to superior computational
resources, big data, and the ability to deploy their intellectual
products at scale (Etzioni 2019).
We find that during 2004-2018, there are 221 AI faculty departures.
131 AI professors com- pletely left universities to pursue an
industry career or to establish their startups, and 90 AI pro-
fessors worked for a private company or opened his/her own startup
while still retaining affiliation with the university. We also
identify 160 AI faculty who moved to another North American uni-
versity during the same period. As a comparison, Azoulay et al.
(2017) document that for every life scientist who left for an
industry job from 1977 to 2006, 90 moved to another department in
academia. The high ratio of AI faculty who left for the industry
job to AI faculty who moved to another university is one of the
signs that North American universities have been experiencing an
unprecedented AI brain drain to the industry.
We use the vast AI brain drain as a shock to study the effect of
founders’ AI knowledge on startup formation and funding.4 The idea
for identification is that faculty departures to the indus- try
reduce the knowledge transfer to the students. If domain-specific
education is an important factor in the creation and the funding of
startups, we should expect that the reduction in student’s AI
knowledge due to faculty departures to the industry would result in
less startup formation by students and these startups would attract
less capital.5
We find that domain-specific knowledge is important for AI
entrepreneurship both on the ex- tensive and intensive margins. On
the extensive margin, we find that AI faculty departures have a
negative effect on the AI startups of students who graduated from
the professors’ former uni- versities. In particular, we find that
the departures by tenured professors’ during the time window
4Usually the term “brain drain” is used in the context of
immigration of highly skilled works to another country (Kwok and
Leland 1982), in this paper we use this term to describe the exodus
of AI professors to the industry.
5In a recent Bloomberg op-ed, Ariel Procaccia, a computer science
professor at CMU, says “If industry keeps hiring the cutting-edge
scholars, who will train the next generation of innovators in
artificial intelligence?” (Procaccia 2019).
2
[t− 6, t− 4] have the largest effect on students graduating in year
t. Specifically, for a given uni- versity, departure of one tenured
AI professor during the time window [t − 6, t − 4] on average
reduces the number of future AI entrepreneurs who graduate in year
t by about 11%. Untenured faculty departures do not show a
significant effect. Moreover, the effect is most pronounced for
entrepreneurs, who graduate from the top 10 universities, who hold
PhD degrees, who establish startups in the field of deep learning,6
and who start new firms within the first two years of their
graduation. We also find that when tenured AI professors leave
universities and they are replaced by faculty from lower-ranked
schools or by untenured AI professors, the negative effect of the
faculty departures on AI startups intensifies.
The extensive margin results suggest that gaining specialized
knowledge from professors is important for establishing a startup.
First, we see the effect mainly for departures prior to students’
enrollment. Second, tenured faculty, especially at the top 10
computer science departments, are more likely to be star professors
whose knowledge students can utilize in their startups, while
untenured professors can leave because they have not received
tenure and they are also less likely to supervise PhD students.
Third, deep learning is a novel field in machine learning that
drives many recent innovations. The fact that faculty departures
disproportionately affect deep learning startups suggests that the
source of knowledge in this advanced area of AI is coming from AI
professors. Any non-specialized skill transfer from professors to
students, such as leadership skills, should be unrelated to the
type of technology used by startups. Fourth, our finding that PhD
students are the most affected category suggests that advanced
knowledge of AI is important for startup success. PhD students make
the largest investment in specialized knowledge and also depend
more on interaction and learning from professors than undergraduate
or master students. Lastly, the fact that the negative effect of
tenured professors’ departures are intensified when they are
replaced by untenured professors or professors from lower-ranked
schools suggests the importance of high quality specialized
knowledge for startup formation.
On the intensive margin, we find that tenured AI faculty departures
at time [t− 6, t− 4] have a negative effect on the early-stage
funding of AI startups with founders who graduated in year t from
the affected universities.7 Specifically, the departure of one
tenured AI professor in time window [t−6, t−4] decreases, on
average, first-round and series-A round funding by,
respectively,
6Deep learning is a machine-learning technique that uses neural
networks with a very large number of layers (LeCun et al.
2015).
7Our focus is on early-stage startups because, as of the year 2018,
85% of the AI startups in our sample are still in the early stage
(i.e., no later than series A round).
3
$1.68 million and $6.6 million. Also, the funding growth rate of a
representative AI startup in our sample is reduced from 630% to
246% by departure of one tenured AI professor in time window window
[t− 6, t− 4]. More recent departures (i.e., one to three years
prior to the entrepreneur’s graduation) or untenured faculty
departures do not have a significant effect.
Our intensive margin results suggest that access to funding and the
funding growth rate depend on the specialized knowledge of
entrepreneurs. On average, AI entrepreneurs in our sample start new
AI startups over two years after graduation, receive seed funding
three years after graduation, and receive Series A funding four and
a half years after graduation. It means that a significant gap
exists between the time professors depart and the time when a
startup receives funding. This long-term effect of AI brain drain
on funding can be explained by the lack of knowledge transfer from
professors to students and by the importance of specialized
knowledge for VC funding. An alternative explanation is that
professors who leave for the industry might have VC connections,
especially in the same city, that benefit students. When professors
leave, students lose these con- nections. However, we do not find a
larger effect of AI brain drain on startups in the same location as
the university or on startups that do not have MBA co-founders.
Another alternative is that VCs who took a class from the same
professor as the founders might have a higher propensity to invest
in the startup. But, this alternative would not explain our results
for deep learning startups or why the negative effect is stronger
when professors are replaced by professors from lower-ranked
schools or by untenured professors.
We also study alternative explanations for our extensive margin
results. One explanation is that professors who leave for the
industry take their best students with them. Inconsistent with this
explanation, we see a strong negative effect for departures that
take place 4-6 years prior to students’ graduation (before
students’ enrollment) and no effect for departures 1-3 years prior
to graduation. Students selection is another important concern.
Maybe the best students enroll in a different university if they
see AI professors leaving for industry from their target school. In
this case, the negative effect should be stronger for departing
professors who do not keep university affiliation. We find the
opposite. The effect is driven by faculty who take industry jobs or
establish own startups but keep university affiliation. These are
likely to be the most accomplished pro- fessors because
universities agree to forego an option to replace them by letting
them keep their affiliation.
There could be unobservables not captured by university fixed
effects that cause a university to lose its AI faculty and to
produce fewer AI entrepreneurs in the future. These
unobservables
4
should also be affecting the amount of funding AI startups receive
a decade after faculty depar- tures. These are unlikely to be
university-specific shocks because we do not find a negative effect
on non-AI startups. We also rule out city-level shocks by showing
that there is no effect of AI faculty departures in one city on
local AI startups established by entrepreneurs who graduated from
universities located in other cities. To address residual
endogeneity concerns, we instrument AI faculty departures 6-4 years
prior to student’s graduation with AI professors’ average number of
citation received 8-6 years before student’s graduation.8 On the
one hand, companies are more likely to hire star faculty with a
high number of (recent) citations (Zucker et al. 2002). On the
other hand, citations are unlikely to directly affect the creation
of startups by the students who graduate six to eight years later,
and are highly unlikely to directly affect the financing of those
startups a decade later.9 The results from the 2SLS regression
further strengthen our causal interpretation of the negative effect
of AI brain drain on the number of startups by students and the
amount of funding.10
Related literature. Previous studies have found that financing
(Hall and Lerner 2010; Kaplan and Strömberg 2003; Kerr et al. 2011;
Corradin and Popov 2015; Bernstein et al. 2017; Ersahin et al.
2020), working experience in tech firms (Gompers et al. 2005;
Elfenbein et al. 2010), peer effects (Nanda and Sørensen 2010;
Lerner and Malmendier 2013), and age (Azoulay et al. 2020) are
important determinants for startup’s creation. Our main
contribution is show the importance of university-acquired
domain-specific knowledge both for startup creation and early-stage
financing.
Kaplan et al. (2009) study 50 VC-backed companies and emphasize the
importance of non- human assets by showing that business lines are
stable from inception to IPO, while management turnover is
substantial. On the other hand, Ewens and Marx (2018) show that VCs
add value through replacing the founders. Bernstein et al. (2017)
study the importance of the founding team in startups’ early-stage
funding. They design a randomized field experiment for 21
early-stage startups and show that the average investor pays more
attention to the founding team information rather than the
information about business ideas and current investors. In a recent
study, Bernstein et al. (2019) find that when local opportunities
arise, the firm formation is almost entirely driven by
8We also use two alternative instrumental variables, the cumulative
number of industry coauthors for all the AI professors at a given
university up to t− 6 and the average number of citations received
in [t− 8, t− 6] based on the papers coauthored with people working
in industry.
9Extensive analysis of the exclusion restriction appears in Section
5.2.2. 10All our regressions include university fixed-effects and a
dynamic ranking of computer science departments. Even
if citations provide an additional signal about the quality of the
department and its students, we would expect to see an increase in
the number of AI startups following an increase in citations, not a
decline.
5
skilled entrepreneurs. To the best of our knowledge, our paper is
the first to show that highly spe- cialized human capital acquired
at universities is a major determinant of entrepreneurship activity
both on the extensive and on the intensive margins.
Lazear (2005) uses data from Stanford MBA alumni and find that
those who end up being entrepreneurs study a more varied curriculum
when they are in the program than do those who end up working for
others, suggesting a university’s generalist, rather than
specialist, training is a more important determinant of its
graduates’ entrepreneurship. Wagner (2003) uses data from Germany
and finds a positive correlation between the diverse experience
across several domains and a probability to establish a new firm.
Our paper challenges the jack-of-all-trades, master of none
characterization of entrepreneurs (Lazear 2004). It is the first to
provide causal evidence that domain-specific knowledge is important
for startup formation and funding.
Our paper also contributes to the debate about the effect of tech
giants on entrepreneurship. On the one side, Babina and Howell
(2018) document a knowledge spillover effect showing that corporate
R&D leads to more startups by the employees. Jin (2020) uses
Amazon’s HQ2 finalists decision to show that the prospect of
getting funded or acquired by tech giants have a positive effects
on entrepreneurship. On the other side, Kamepalli et al. (2020)
argue that big tech platforms create a “kill zone” around their
business. They show that investments by VCs in startups drop after
Facebook and Google make major acquisitions in the area in which
these startups operate.11 Our paper shows that big tech firms can
also have a negative effects on entrepreneurship through their
poaching of university professors and reducing the diffusion of AI
knowledge.
More broadly, the paper contributes to a new literature that
studies AIâs impact on the economy and society. In a recent paper,
Acemoglu and Restrepo (2018) develop a framework in which AI can
have both a negative and a positive effect on demand for labor. The
negative effect is due to the displacement risk, and the positive
effect is due to improved productivity and higher capital
accumulation. Korinek and Stiglitz (2017) and Guerreiro et al.
(2017) discuss the channels through which AI can increase
inequality and the best way to address this concern. Aghion et al.
(2017) study the implications of AI on economic growth. To the best
of our knowledge, this is the first finance paper about AI.12 Our
first contribution to this literature is to show that scarcity of
AI
11Related, Cunningham et al. (2020) find that big pharma firms make
acquisitions to discontinue target’s innovation process in order to
preserve its current business.
12There is a growing literature that uses deep learning and machine
learning as a tool for finance applications (Heaton et al. 2017;
Chen et al. 2019; Aubry et al. 2019; Erel et al. 2018; Gu et al.
2020) but AI itself is not the focus of this literature.
6
human capital is important for innovation by AI startups. This
scarcity can also reduce labor- augmenting technical progress,
increase inequality, and negatively impact opportunities for AI-
driven economic growth. Our second contribution to this literature
is to analyze the obstacles of AI startup formation and funding.
Given the importance of innovation to the economy (Romer 1990),
these obstacles should not be overlooked.
2 Data Description and Summary Statistics
2.1 AI Entrepreneurs and AI Startups
Entrepreneurs’ and startups’ information is based on a sample from
the CrunchBase database.13
According to the Kauffman Foundation, CrunchBase is "the premier
data asset on the tech/startup world."14 One of the advantages of
CrunchBase over other commercial databases is the broad cov- erage
due to crowdsourcing and data coverage from TechCrunch. As a
result, CrunchBase sample includes startups that are not financed
by venture capitals, compared to the data from VentureX- pert.
Dalle et al. (2017) compare the coverage of CrunchBase with that of
OECD Entrepreneurship Financing Database and find that starting
from 2010, the start year of our sample, CrunchBase has broader
coverage in both the US and other countries.
More importantly, relative to other databases, CrunchBase provides
two unique features that are necessary for our paper. First,
CrunchBase provides technology-specific or product-specific
keywords that allow us to clearly identify AI startups. To classify
AI entrepreneurs/startups, we rely on the business categories
CrunchBase provides. AI entrepreneurs/startups have at least one of
the following categories: artificial intelligence, machine
learning, deep learning, neural net- works, robotics, face
recognition, image processing, computer vision, speech recognition,
natural language processing, autonomous driving, autonomous
vehicle, and the semantic web. Second, given that AI is still an
emerging technology and AI brain drain does not become prevalent at
most universities until 2012, it is important to use a database
that provides the most recent data available. CrunchBase allows us
to track newly established AI startups until 2018.
The sample we obtain from CrunchBase includes 363 AI entrepreneurs
and 985 Non-AI (IT) entrepreneurs who graduated from 69 North
American universities during 2010 - 2018. An AI
13CrunchBase would be a less than ideal database for studying
entrepreneurship in non-tech industries, but for our study, the
database’s tech industry focus works to our advantage.
14Source:
https://www.kauffman.org/microsites/state-of-the-field/topics/finance/equity/venture-capital
7
entrepreneur is identified if he or she starts an AI startup after
received the highest degree in our sample. The 69 universities are
selected so that at least one AI entrepreneur graduated from each
of those universities before 2010. In total, they founded or
co-founded 177 AI startups and 591 Non- AI (IT) startups. These AI
startups innovate in the areas of cybersecurity, retail, early
detection of breast cancer, solutions for the oil and gas industry,
augmented reality, sleep therapy, supply chain planning, and
more.
One of the contributions of the paper is to compare AI startups and
AI entrepreneurs and to non-AI startups and non-AI entrepreneurs in
other fields of the information technology (IT) sector. In Table 2,
we compare characteristics of 363 AI entrepreneurs from 177 AI
startups with charac- teristics of 985 entrepreneurs from 591 IT
startups that are not AI.15 We find that the distribution of the
educational levels for AI founders is skewed towards higher
degrees. Specifically, we find that AI entrepreneurs, on average,
hold higher academic degrees than non-AI entrepreneurs, with as
many as 26% holding a PhD degree, as opposed to 10% for non-AI
entrepreneurs. And 24% of AI startups have at least one founder
with a PhD degree, while only 7% of other startups in the IT field
have at least one PhD-holding founder. In terms of early-stage
funding, compared to other IT startups, AI startups raised $1.01
million more in the first round and $2.9 million more in the series
A round. We also find that AI entrepreneurs are less likely to be
the sole founder relative to non-AI entrepreneurs. They have a
larger time gap between the graduation year for their highest
degree and the year when they established their startup. All these
facts suggest that AI startups require a higher level of expertise
and knowledge than startups in non-AI fields.
We document that top North American universities produced the most
alumni who established AI startups after receiving their highest
degree (see Figure OA1 in the Online Appendix). In our sample, 78
MIT graduates and 72 Stanford graduates established AI startups.
Carnegie Mellon University, ranked third according to this metric,
produced 39 AI entrepreneurs. The Canadian university with the most
AI entrepreneur alumni is the University of Waterloo with 22 AI en-
trepreneurs.
2.2 AI Professors
For AI professors leaving for an industry job, we exploit a
hand-collected sample from LinkedIn. To obtain AI professors from
LinkedIn, we employ two searching processes. The first
involves
15Our sample includes only founders who graduated on or after 2010
because we need a lag between professors’ departure and students’
graduation.
8
directly searching in Google with inputs such as:
site:linkedin.com/in/ “Professor” and “Artificial Intelligence”16,
which allows us to retrieve all the LinkedIn profiles returned by
Google.
Our second method is to search in LinkedIn using reviewers’ and
program committee mem- bers’ names of AI related conferences. We
picked three of the most prestigious conferences in the field of
machine learning and artificial intelligence, which are
International Conference on Ma- chine Learning (ICML), Neural
Information Processing Systems (NeurIPS), and the Association for
the Advancement of Artificial Intelligence (AAAI), and extracted
the names of reviewers and committee members from 2008 to 2018.
Because there are also reviewers and committee members from the
industry, this procedure includes faculty who left academia for the
industry. We manually check each person’s LinkedIn page and
sometimes their academic webpage to verify the results.
As we only consider tenured or tenure-track faculty, we exclude the
sample people with titles like “Adjunct Professor”, “Clinical
Professor ”, and “Research Professor”. We classify assistant
professors as tenure-track faculty and associate or full professors
as tenured faculty.
The final sample includes 221 tenure-track and tenured AI faculty
in North American univer- sities who started to work for a private
company or established their own startup between 2004 and 2018.17
This sample includes 131 faculty who left their academic positions
completely and 90 faculty who retained their university
affiliations. Figure 1 shows the aggregate time trend for the
combined sample of complete and partial departures (i.e., still
retain the university affiliations). We also identify 160
tenure-track or tenured faculty who moved within North American
academia. In our empirical analysis, we use both the faculty who
leave for industry and for another university.
We also download data of faculty size at the top 100 universities’
computer science depart- ments from CSRankings.org, which provides
the number of full-time, tenure-track and tenured computer science
(CS) faculty for each year based on data from DBLP.18 DBLP provides
open bibliographic information on major computer science journals
and proceedings. Its website states that “DBLP indexes over 4.4
million publications, published by more than 2.2 million authors.
To this end, DBLP indexes about 40,000 journal volumes, more than
39,000 conference and workshop proceedings, and more than 80,000
monographs.”19
16Another example of a search input is: site:linkedin.com/in/
“Professor” and “Deep Learning” . In general, we search for
professors who mention one of the subfields of AI, such as natural
language processing or autonomous driving, on their LinkedIn
profiles.
17Our sample can underreport the actual number of AI faculty
departures to industry because some professors might not have
LinkedIn accounts or leave their LinkedIn profiles
out-of-date.
18See http://csrankings.org/faq for detailed information about
CSRankings.org. 19Source: https://dblp.org/faq.
9
Last, we downloaded citation data of AI faculty from Google
Scholar, which are used as a proxy for the quality of their
research.
2.3 AI Brain Drain
Although we use AI brain drain to generate the variation in
student-specific knowledge, AI brain drain is an interesting
phenomena on its own.20 One of the contributions of the paper is to
use a hand-collected dataset of AI faculty departures to document
the brain drain of AI professors from universities into the
industry. Despite lots of public interest in the brain drain of AI
professors to the industry (Lowensohn 2015; Economist 2016; Sample
2017; Spice 2018; Procaccia 2019), there is no study that
systematically investigates this issue.
We identify 221 tenured and tenure-track AI faculty who accepted an
industry position or started their own startups between 2004 and
2018. Figure 1 shows an exponential trend in the number of faculty
departures. In our sample, there were no AI professors leaving
academia in 2004 whereas 41 AI professors left completely or
partially in 2018. Figure 1 also shows the citation ratio of AI
faculty who left for an industry job to all AI faculty. For each
year and each university, the citation ratio is calculated using
the sum of Google Scholar citations from faculty who left for an
industry job (stock at the time of departure) divided by the sum of
citations from all AI faculty in this university; we then average
across universities to get a proportion of citations for a given
year. Only 4% of total AI citations from a university are from
professors who left for the industry in 2010, on average, whereas
AI professors who left academia for an industry job in 2018 account
for about 20% of a university’s total AI citations.
Panel A and B of Figure 2 show the top 18 North American
universities in terms of the number of faculty who took industry
positions or established own startups from 2004 to 2018. The three
universities that lost the most AI faculty are Carnegie Mellon
University (CMU), the University of Washington, and UC Berkeley.
CMU lost 16 tenured faculty members and no untenured faculty, and
the University of Washington lost 7 tenured and 4 assistant
professors. For Canadian univer- sities in our sample, the
University of Toronto lost the most AI professors, 6 tenured
faculty and 3 assistant professors.
20Previous studies of brain drain from universities focused on
outcomes within academia. Waldinger (2010) inves- tigates the
effect of the expulsion of mathematics professors in Nazi Germany
on PhD student outcomes. Borjas and Doran (2012) study the effect
of immigration of Soviet Union mathematicians to U.S. on the
productivity of American mathematicians. Our paper focuses on the
link between AI brain drain and economic and financial outcomes,
such as firm formation and funding.
10
Panel A of Figure 2 further considers whether there is any
difference between the departures of tenured faculty and untenured
faculty for an industry job. We see that the two groups of faculty
members exhibit similar trends. Panel B further looks into whether
AI professors still keep their university affiliation after they
accept an industry position or start their own startups. We
identify departed AI professors with university affiliation in two
steps. First, we select professors who accept an industry position
with an open-end date (i.e., present) for their university
positions on their Linkedin profiles. An open-end date indicates
that the professors are still employed by the university. However,
to ensure that an open-end date on Linkedin profile is up-to-date,
we manually check the professors with an open-end date on their
universities’ websites. If they still appear on the universities’
website with active teaching or supervising record, we consider
them departures with university affiliation. Panel B of Figure 2
depicts the AI brain drain for the group with and without
university affiliation. Both groups exhibit similar growing trend
from 2004 to 2018.
Figure 3 presents firms that hired most of the AI professors. In
our sample, Google, Amazon, and Microsoft poached the most AI
professors from North American universities. From 2004 to 2018,
Google and its subsidiary, DeepMind, together hired 23 tenure-track
and tenured AI professors from North American universities. Amazon
and Microsoft respectively hired 17 and 12 AI professors. Apart
from technology firms, we also see that large firms from the
finance industry poach AI professors, such as Morgan Stanley,
American Express, and JP Morgan. It is worth noting that that these
publicly traded firms hired about 45% of the 221 professors in our
sample.
To induce the best AI professors to leave their tenure-track or
tenured positions, the industry offers them millions of dollars in
compensation (e.g., Metz 2017, 2018a). The corporate poaching of AI
faculty has raised public concerns about its negative impact on
universities (e.g., Economist 2016; Sample 2017). “That raises
significant issues for universities and governments. They also need
A.I. expertise, both to teach the next generation of researchers
and to put these technologies into practice in everything from the
military to drug discovery. But they could never match the salaries
being paid in the private sector.” said Yoshua Bengio, one of the
2018 Turing Award winners and a professor at the University of
Montreal (Metz 2017).
Although the compensation data that AI professors receive from
industry positions are not publicly available, anecdotal evidence
may provide a sense of how large the pay gap is. Ilya Sutskever,
Geoffrey Hinton’s student, went to work for OpenAI in 2016, a
non-profit organization that aims to democratize AI, where he was
paid more than $1.9 million dollars that year. “I turned down
offers for multiple times the dollar amount I accepted at OpenAI.
Others did the same.” Mr.
11
Sutskever said (Metz 2018a).21 As a comparison, the highest-paid AI
professor at UC Berkeley received $448,742 in 2018.22 The large pay
gap and better computational resources and access to the commercial
datasets offered by industry positions make it nearly impossible
for universities to retain their best AI professors. In this paper,
we use AI brain drain as a source of variation in students’
AI-specific knowledge to provide a causal evidence about the effect
of specialized education on entrepreneurship.
3 Empirical Analysis
This section presents empirical evidence of how domain-specific
knowledge affects the creation and funding of AI startups by
exploiting the variation of AI faculty’s departures to industry. We
have two concerns about using AI brain drain as a shock to
students’ domain-specific knowledge. First one is that AI brain
drain may generate variation in other factors, such as connection
to VC, that could also affect formation and financing of startup by
university graduates. We address this concern in Section 4. The
second concern is that the variation generated by AI brain drain
may not be completely exogenous. We address this concern in Section
5.
3.1 Empirical Design
For the extensive margin analyses, we conduct university-level
analyses. Specifically, we test whether AI professors’ departures
from a university for the industry affect the number of AI star-
tups established by students who graduated from this university.
Given that AI faculty departures could repeatedly occur at a given
university, instead of a difference-in-differences specification,
we have a simple panel OLS specification as follows23
ln(1+AI Entrepreneur j,t) = αt +θ j +βAI Brain Drain
j,[t−6,t−1]+φXj,t + ε j,t (1)
, where αt are the graduation year fixed effects, θ j are
university fixed effects, AI Entrepreneurr jt
is defined as the number of graduates who graduate from university
j in year t and then start AI
21“Well-known names in the A.I. field have received compensation in
salary and shares in a company’s stock that total single- or
double-digit millions over a four- or five-year period.” (Metz
2017). See also Metz (2018b).
22The wage data is from https://ucannualwage.ucop.edu/wage/
23Because the dependent variable, AI Entrepreneur jt , is count
data, in addition to OLS, we also use both Poisson
and Negative Binomial model for the same regressors as a robustness
in Section 6.
12
startups, β captures the effect of AI Brain Drain on the students’
ability to start AI firms after they graduate, and other regressors
are defined in Table 1. Because of the lag between graduation and
startup inception (the average is about two years), graduates in
2018 mechanically have fewer AI entrepreneurs than graduates in
2010. We control for this by adding time fixed effects. In
addition, we add university fixed effects to control for the fact
that some universities produce more AI entrepreneurs because of
their location near areas like Silicon Valley or other
factors.
To measure the AI faculty departures, we first look at total AI
faculty departure during the time window, [t−1, t−6], where t
indicates the student’s graduation year.24 The corresponding
variable, AI Brain Drain[t−6,t−1], is calculated as the sum of each
year’s AI faculty departures scaled by each year’s AI faculty size
for past six years at each university.
The time window [t−1, t−6] does not allow us to disentangle the
effects between those pro- fessors who have more opportunity to
interact with students who graduate in year t and those with less
opportunity or no opportunity to interact with students who
graduate in year t. To address this issue, we measure AI faculty
departure separately at two time windows [t− 3, t− 1] and [t− 6,
t−4] for the departures, where t is the graduation year.
Effectively, we break down the variable AI Brain Drain[t−6,t−1],
into two variables, AI Brain Drain[t−6,t−4] and AI Brain
Drain[t−3,t−1]. If a professor departs four to six years prior to
student’s graduation year, this professor has probably no
interaction with the student because undergraduate programs last
four years, master programs are usually two years long and PhD
programs are five years long, with most of the classes taken in the
first couple of years.
t−6 t−5 Departures Period 1
t−4 t−3 t−2 Departures Period 2
t−1 t
Startups by Graduates
This figure plots the timeline for our empirical design. For
departures, we focus on tenure-track and tenured AI professors who
left academia for an industry job in two time windows, [t − 6, t −
4] and [t − 3, t − 1]. For student entrepreneurs, we look at people
who graduate in year t and subsequently establish an AI
company.
24We use a window instead of putting a specific year because it is
difficult to determine precisely the year when a professor actually
stopped perform its teaching and supervisory duties at the
university. Professor can take a leave of absence before they
officially resign.
13
Moreover, we look at the effects of tenured professors’ departures
and untenured departures separately because of the potential
differences in their impact on students. In addition, untenured
professors could be leaving academia because they were denied
tenure. Therefore, departures by tenured professors are likely to
have a larger effect on students. So we further break down the
vari- able AI Brain Drain[t−6,t−4] into Tenured Brain
Drain[t−6,t−4] and Untenured Brain Drain[t−6,t−4]. For the period
[t-3,t-1], we break down the variable AI Brain Drain[t−3,t−1] into
Tenured Brain Drain[t−3,t−1]
and Untenured Brain Drain[t−3,t−1].
3.2 The Creation of AI Startups (Extensive Margin)
The regression sample includes 69 universities North American
universities that produced at least one AI entrepreneur who
received his/her highest degree before 2010. 59 of these
universities are in the US and 10 universities are in Canada. Table
3 presents summary statistics for these 69 universities from 2010
to 2018. From this table we learn that an average department has
around 7.8 AI professors during our sample period. For our
comparative statics we will use departure of one AI faculty member,
which would be 12.8% of the average AI group.
The OLS results of the estimation of equation (1) appear in Table
4. In column (1), the coeffi- cient on the variable AI Brain
Drain[t−6,t−1] is negative and statistically significant at 5%
level.25
Once we break down AI Brain Drain[t−6,t−1] into AI Brain
Drain[t−6,t−4] and AI Brain Drain[t−3,t−1]
in column (2), only the coefficient on AI Brain Drain[t−6,t−4] is
negative and statistically significant at 5% level.
In column (3), when we examine the impacts of tenured and untenured
faculty separately, Untenured faculty leaving for an industry job
has no effect, but departures by tenured profes- sors have a
statistically significant effect. In terms of economic
significance, the coefficient on Tenured Brain Drain[t−6,t−4] is
-0.863 in column (3), which implies that the departure of one
tenured AI professor during the time window [t−6, t−4] on average
reduces the number of future AI entrepreneurs who graduate in year
t by about 11%.26
In addition, in column (4), we show that AI faculty departures have
no effect on Non-AI en- trepreneurship (i.e. students who start
firms after their graduation in the area of information tech-
25Robust standard errors are clustered at university city level
because we cannot assume that two observations in the same city are
independent.
26One departure represents 12.8% of the sample average AI faculty
size. 11% is calculated as the coefficient, -0.863, on Tenured
Brain Drain[t−6,t−4] multiplied by 12.8%.
14
nology but not in AI). This result implies that the negative effect
is unlikely to be driven by would- be AI entrepreneurs switching to
other IT fields. It also suggests that that the effect is driven by
an unobservable university-level shock because such shock would
affect non-AI entrepreneurs as well. Overall, the results in Table
4 show that faculty departures to the industry have a long-lasting
negative effect on startup formation.
3.3 Examining Heterogeneous Effects
In this subsection, we investigate various heterogeneous effects
for the results of the extensive mar- gin. Table 5 examines whether
the negative effects shown above are heterogeneous with respect to
the CS department’s ranking. We also study whether Canadian
universities are affected differently from US universities. As
shown in columns (1) through (3), the effects are much stronger for
the Top 10 CS departments, as ranked by CSRanking.org based on the
number of publications and ad- justed for faculty size. In
addition, in column (4), we do not see the impacts of faculty
departures on Canadian universities being significantly different
from those on American universities.
Table 6 investigates heterogeneous effects with respect to whether
the AI entrepreneurs’ highest degrees are a bachelor or PhD,
whether the startup utilizes or develops the deep learning technol-
ogy, and whether the gap between graduation and inception is less
than two years. In column (1), the dependent variable only includes
AI entrepreneurs whose highest degrees are bachelor and is
calculated as the natural log of one plus the number of AI
entrepreneurs with only bachelor degrees. The coefficient on our
main faculty departure measure, Tenured Brain Drain[t−6,t−4], is
statistically significant at 5% level, and the magnitude is similar
to that in the main results shown in Table 4. In column (2), we
only focus on AI entrepreneurs who hold PhD degrees. The
coefficient on the variable, Tenured Brain Drain[t−6,t−4], is
statistically significant at 5% level. Compared to the
undergraduate results in column (1), the magnitude for PhD
entrepreneurs is about 30% higher. The results from columns (1) and
(2) indicate that the negative effects apply to both undergraduate
and PhD students.
Recent advancements in deep learning algorithms is a major factor
in the renewed interest in AI. Figure OA4 in the Online Appendix
shows the exponential growth in Google search index for the term
“deep learning” and in the total number of Google citations for the
three most influential deep-learning papers published before 2010
by the three 2018 Turing Prize winners.
Deep learning could be the Invention of the Method of Invention
(IMI), a term coined by
15
Griliches (1957) in his seminal work on hybrid corn. Cockburn et
al. (2018) write “If it [deep learning] is also a general-purpose
IMI, we would expect it to have an even larger impact on
economy-wide innovation, growth, and productivity as dynamics play
out—and to trigger even more severe short-run disruptions of labor
markets and the internal structure of organizations.” They
continue, “It is possible that deep learning will change the nature
of scientific and technical advance itself.”27
Given the importance of GPT technologies for the economy (David
1989; Autor et al. 1998; Bresnahan and Trajtenberg 1995; Bresnahan
et al. 2002; Rosenberg and Trajtenberg 2004), in columns (3) and
(4) of Table 6 we look specifically at AI startups in the area of
deep learning. There are 106 startups in our sample that have a
keyword “deep learning” in their description. Indeed, these
startups operate in very different fields (cybersecurity, retail,
breast cancer early detection, solutions for the oil and gas
industry, augmented reality, sleep therapy, supply chain planning,
and more), consistent with the possibility that deep learning is a
general-purpose technology that can be applied across many
industries.
In column (3) of Table 6, the dependent variable only includes AI
entrepreneurs who establish startups that use deep learning
technology. The coefficient on Tenured Brain Drain[t−6,t−4] is
statistically significant at 5% level, and the magnitude is very
similar to that in the main results shown in Table 4. In column
(4), we focus on AI entrepreneurs who establish startups in other
sub- fields in AI. The coefficient on Tenured Brain Drain[t−6,t−4]
is statistically insignificant. These results (i.e., the negative
effects are the strongest on deep-learning startups) suggest that
founders’ ability to acquire deep learning knowledge from
professors is particularly important for AI startup formation.
Among all the subfields in AI, establishing a deep-learning startup
probably relies mostly upon domain-specific knowledge. This result
is consistent with our knowledge transfer channel, for which we
have an extensive discussion in Section 4.
In columns (5) and (6), we compare the impacts of faculty
departures on graduates who start new firms within two years after
their graduation and those who start new ventures two years after
they graduate. This comparison is important because if students
have been out of school for several years then there are many other
factors that may drive entrepreneurial activity such as exposure to
ideas through employment, personal wealth, opportunity costs, etc.
Therefore, we expect the
27See also Gil et al. (2014) who write “The challenge presented by
advances in AI is that they appear to be research tools that not
only have the potential to change the method of innovation itself
but also have implications across an extraordinarily wide range of
fields.”
16
impacts of AI faculty departure to be greatest on students who are
right out of school. This is indeed what we observe from columns
(4) and (5). In column (4), we only focus on graduates who start
new firms within two years, and the coefficient on Tenured Brain
Drain[t−6,t−4] is statistically significant at 5% level, while in
column (5), there are no effects on graduates who established
startups two years after they graduate.
3.4 Early-Stage Entrepreneurial Financing (Intensive Margin)
Early-stage startups are particularly hard to finance, given the
lack of tangible assets and the severe asymmetric information (Hall
and Lerner 2010). As a result, the information about the human
capital of a founding team becomes important in terms of attracting
early-stage funding (Bernstein et al. 2017). However, through which
mechanism (e.g., knowledge, networking, and signaling) the human
capital affects early-stage financing still remains a question.
Moreover, as we discussed earlier, AI startups are much more likely
to have PhD founders than other non-AI startups in the IT sector.
Thus, in the field of AI, knowledge transfer from university
professors to AI entrepreneurs may have a significant impact on
their human capital. In this section and Section 5, we provide
causal evidence that it is the knowledge, rather than signaling or
networking, through which human capital affects funding.
As shown in Panel B of Table 2, as of the year 2018, 85% of the AI
startups in our sample are still in the early stage (i.e., no later
than series A round). Therefore, our focus is on the first and
series A rounds. Table 2 also shows that compared to other IT
startups, AI startups raised $1.01 million more in the first round
and $2.9 million more in the series-A round.
Our extensive-margin results show that the negative impacts are
mainly due to tenured faculty departures in time window [t − 6, t −
4]. Based on this result, for intensive margin, we again examine
the impacts of faculty departures for tenured vs. untenured and
time window [t−3, t−1] vs. [t−6, t−4]. As a result, we use the
following OLS model to test the effects of faculty departures on
students’ ability to attract startup funding.
Funding Amounti, j,k,t,τ = ατ +θ j +δk + γ1Untenured Brain Drain
j,[t−1,t−3]+ γ2Untenured Brain Drain j,[t−4,t−6]
+ γ3Tenured Brain Drain j,[t−1,t−3]+ γ4Tenured Brain Drain
j,[t−4,t−6]+φXi,j,k,t,τ + εi, j,k,t,τ
(2)
, where the dependent variable, Funding Amounti, j,k,t,τ , is the
total funding amount from a financ- ing round (either first-round
or series-A in our case) in year τ for a startup founded
(co-founded)
17
in state k by entrepreneur i who graduated from university j in
year t, ατ is the financing year fixed effect, and θ j is the
university fixed effect, δk is the state fixed effect.28 The
financing year fixed effects can control for the cyclicality of
venture financing, univeristy fixed effects can control for
alumni’s VC connection, and startup’s state fixed effects factor in
venture capital industry develop- ment in different states in the
US. Entrepreneur-level control variables include the university
rank, the number of years between graduation and startup inception,
and prior exit (acquisition or IPO) experience. Firm-level control
variables are the number of founders and the number of investors at
each financing round (either first-round or series-A round in our
case).
Table 7 shows the effects of AI faculty departures on AI startups’
early-stage funding. The OLS results in columns (1) the effect of
AI faculty departures on AI startups’ first-round funding. Only the
coefficient on Tenured Brain Drain[t−6,t−4] is negative and
statistically significant at the 1% level. The coefficients on
Tenured Brain Drain[t−3,t−1], Untenured Brain Drain[t−3,t−1], and
Untenured Brain Drain[t−6,t−4] are all insignificant. Together,
these results indicate that only the tenured AI faculty departures
during [t−6, t−4] have a negative impact on the first-round funding
of AI startups by entrepreneurs who graduate in year t. Untenured
AI departures or early tenured AI departures (i.e., during [t−3,
t−1]) have no impacts on entrepreneurs who graduate in year t. For
the economic magnitude, the departure of one tenured AI professor
in [t−6, t−4] results in about a $1.68 million decrease in
first-round funding.29 These results are consistent with the
knowledge transfer channel because tenured professors are, on
average, more knowledgeable. And students are likely to take
classes and do research under the supervision of the professors who
depart during students’ studies. Therefore, the negative effect
should be larger for tenured faculty departures and for the
departures before students’ arrival on campus.
In column (2) and (3), we show the effects of AI faculty departures
on AI startups’ funding growth from the first to the second round
and funding of Series-A round, respectively. The results are
consistent with those in column (1) in the sense that only tenured
AI faculty departures during [t− 6, t− 4] have a negative impact,
and untenured AI departures or early tenured AI departures (i.e.,
during [t − 3, t − 1]) have no impacts. As for economic magnitude,
the departure of one tenured AI professor in [t − 6, t − 4] could
decrease the funding growth rate of a representative AI startup in
our sample from 630% to 246%. And the departure of one tenured AI
professor in
28First-found is the first funding round we can observe for a given
firm in the CrunchBase database. 29One departure represents 12.8%
of the sample average AI faculty size. The $1.68M reduction in
funding is
calculated as the coefficient, -13.1, on Tenured Brain
Drain[t−6,t−4] in column (1) multiplied by 12.8%.
18
[t−6, t−4] results in about a $6.6 million decrease in series-A
round funding.30
3.5 Faculty Replacement and Partial Departures
In this subsection, we first test whether professors leaving for
other universities, rather than indus- try, could pose similar
negative effects on entrepreneurship by university graduates. Then
we show whether the negative effects are similar between professors
who took a full-time industry job and professors who also have
university affiliations.
3.5.1 Faculty Replacement
Besides the 221 departures from academia to the industry, there are
also 160 departures to a dif- ferent North American university.31
In columns (1) of Table 8, we study whether the effect of the AI
faculty departures to other universities on AI startups is the same
as the effect of faculty departures to the industry while
controlling for the industry departures. Unlike the findings for AI
faculty who accepted an industry job, the results in column (2) of
Table 8 show that tenured or tenure-track professors who move to
other North American universities do not have a negative effect on
AI startups. This could be because AI faculty who accept industry
jobs might be different from those who choose to stay in academia
in terms of the quality.
In columns (2) through (5) of Table 8, we examine whether new AI
tenured/untenured profes- sors, who come to the universities that
previously experienced AI faculty departures, can remedy the
negative effects due to the tenured faculty departures. In column
(2), the coefficient on the interaction term, Tenured Brain
Drain[t−6,t−4]×Untenured Move In[t−3,t−1], is negative and statis-
tically significant. This result implies that the negative effect,
due to tenured faculty departure in time window [t − 6, t − 4], is
amplified by untenured faculty who come in later in time win- dow
[t − 3, t − 1]. In column (3), the coefficient on the variable,
Tenured Brain Drain[t−6,t−4]× Tenured Move In[t−3,t−1], is
statistically insignificant, suggesting that new tenured professors
are unable, if not worsen, to offset the negative effects on
student entrepreneurship.
In columns (4) and (5) of Table 8, we examine the replacement
effect of substitute professors who come from the higher-ranked
universities and those who come from lower-ranked institutions
separately. We consider a university to be ranked lower (higher)
than another one if it is ranked at
30The calculation procedure for economic magnitude in column (2)
and (3) is similar to that of column (1) 31We extract this
information from the faculty’s LinkedIn profiles.
19
least five spots lower (higher), based on the CS department
ranking. In column (4), the coefficient on the interaction term
between Tenured Brain Drain[t−6,t−4] and Move In From Low[t−3,t−1]
(i.e., the number of professors who come from lower-ranked
universities during [t− 3, t− 1] scaled by the AI faculty size) is
negative (significant at the 1% level), suggesting that the
negative effects are worsened by faculty who come from lower-ranked
universities. Furthermore, even for professors who come from
higher-ranked universities, the negative effect is not offset, as
suggested by the results from column (5). The sample size is small
for this category because tenured professors rarely move to
lower-ranked departments.
The findings from columns (2) through (5) in Table 8 suggest that
AI professors who leave for industries are probably the top
researchers in the field of AI and are hard to replace them with
someone of similar quality, especially during the times when all
departments are experiencing some level of brain drain. As a
result, students continue to be negatively affected by AI faculty
who joined the industry four to six years prior to their graduation
even if new AI faculty join the department.
3.5.2 Partial Departure (University Affiliation)
In Table 9, we examine the effects separately for AI faculty who
work for companies while keeping university affiliations (referred
to affiliated brain drain) and AI professors who took a full-time
industry job (referred to non-affiliated brain drain). The details
of how we classify a departure as affiliated brain drain are
presented in Section 2.3. Column (1) of Table 9 examine the effects
of affiliated and non-affiliated AI brain drain separately in the
departure time window [t− 6, t− 1]. The coefficient on the
affiliated brain drain is negative and statistically significant
while the coefficient on non-affiliated brain drain is
insignificant. In Table OA11 of the Online Appendix we check
whether the effect of non-affiliated brain drain and affiliated
brain drain depends on the new faculty who join the department. For
the affiliated brain drain, the type of replacement does not
matter. This is not surprising because the new faculty are not
replacing the existing faculty who dedicate part of their time to
the institution and another part to the outside job. However, for
non-affiliated brain drain we find a striking result. When
professors who leave completely are replaced by professors from
lower ranked departments, there is a strong and significant
negative effect on students’ entrepreneurship. When these
professors are replace by professors from higher ranked departments
the effect is positive and statistically significant.
These results suggest that affiliated departures have a negative
effect unconditionally, while
20
non-affiliated departures have a negative effect only if the
departed professors are replaced by professors from lower ranked
schools. This finding is consistent with knowledge-transfer channel
because AI faculty who have dual positions (i.e. industry and
academia) do not allow the university to hire someone instead. As a
result, the net effect of the departure is only negative because
these professors start to teach less and allocate less time to
mentoring students.32 Professors who leave universities completely
free up a position. If this position can be replaced by the same
quality or higher quality professor, we should not see a negative
effect according to the knowledge transfer story. If anything, the
effect can be positive if the university can attract someone
better. However, if the replacement is of lower quality, then the
net effect of the departure on students’ knowledge is
negative.
In addition, although university fixed effects rule out the school
signalling effect (Spence 1973), the fact that the negative effect
is coming from affiliated brain drain allows us to further rule out
the signaling effect coming from departed professors. For example,
maybe venture capitalists know professors and think that professors
participate in admissions of students. When a star professor
leaves, the perceived quality of students is reduced. However, the
perceived quality should drop more for complete departures
(non-affiliated brain drain) and for professors who depart after
en- rollment of students to the university. We find the opposite,
suggesting that the results are not driven by dynamic
signaling.
We then turn our focus on the group of affiliated AI departures. In
column (2) of Table 9, we examine the effects separately for
departure time windows [t − 3, t − 1] and [t − 6, t − 4]. In column
(3), we show the effects separately for tenured professors who
partially left university and untenured partial departures. The
results in column (2) and (3) are consistent with our main results
in the sense that for affiliated AI departures, the negative
effects concentrate in the departure window [t−6, t−4] and in the
group of tenured AI professors.
32Professors who keep their positions are more likely to be
influential and highly cited researchers making their partial
departure even more important for students.
21
4.1 AI Brain Drain and Students’ AI knowledge
AI startups require a highly specialized human capital that
constitutes a high barrier to entry for many potential
entrepreneurs. Table 2 shows that in our sample, 26% of AI
entrepreneurs hold a PhD degree, whereas it is only 10% for non-AI
entrepreneurs. Moreover, 29% of AI entrepreneurs’ highest degrees
is a Master’s degree, while 21% of non-AI entrepreneurs’ highest
degree is a Master’s degree. Moreover, Medium (Stanford 2018)
writes, “In a report, Paysa found that 35 percent of AI positions
require a Ph.D. and 26 percent require a master’s degree. Why?
Because AI is a rapidly growing field and when you study at the
Ph.D. level and participate in academic projects, they tend to be
innovative if not bleeding edge, and that gives the student the
experience they need for the work environment.”
The knowledge transfer channel attributes the negative effect of AI
faculty departures on star- tups to the reduced quality and
quantity of AI knowledge acquired by students. If students do not
have access to the best AI professors then they are less likely to
become AI entrepreneurs after graduation. We find a strong
empirical support for this channel.
First, we show that the effect is the strongest for students in the
top universities, startups in the area of deep learning, departures
by tenured professors, and departures by professors four to six
years prior to students’ graduation. These results are consistent
with knowledge transfer channel because the quality of knowledge to
be transferred to students is probably higher for tenured profes-
sors in top universities, deep learning requires high training and
the scarcity of knowledge makes the knowledge transfer particularly
important, and departures that take place prior to students’ en-
rollment to the program reduce students’ ability to acquire
knowledge more than departures that happen after students enroll
and take some classes from the departing professors.
Second, the IV regression results, shown in the Section 5.2,
indicate that professors with the highest levels of knowledge,
proxied by the number of their citations, are more likely to leave
academia. The negative effect on entrepreneurship is driven by
departures that are predicted by citations, suggesting that the
results come from departures of the more knowledgable professors
and not the less knowledgable. In fact, we find that when the
departing professors are replaced by professors from a lower ranked
computer science departments, the negative effect of the departures
to the industry is even stronger.
However, it is possible that the brain drain of AI professors into
the industry can affect the
22
number as well as the financing of AI startups established by
students at the affected universities via other channels. In this
section, we discuss possible economic channels and argue that the
most likely channel is the disruption in the transfer of critical
knowledge from professors to students.
4.2 Alternative Explanations
Professors’ VC connections. It could be that an AI professor does
not contribute any knowledge to students, but rather helps them
find investors given the importance of connection in VC’s invest-
ment decision (Gompers et al. 2020). When faculty departs for the
industry, future students lose the industry connections that these
professors have.
According to the connection channel, the negative effect of faculty
departure should be stronger for local startups because professors
are more likely to know VCs in their city. Table OA3 in the Online
Appendix shows that the negative effect of AI faculty departures is
not driven by students who establish their startups in the same
city as their universities’ city. To further test the VC connection
channel, Table OA4 in the Online Appendix shows that, conditional
on being AI en- trepreneurs, AI founders who are affected by AI
faculty departures are not more likely to have co-founders who hold
an MBA degree. If the founders have difficulties to get VC
connections due to faculty departures, they should be more likely
to partner with MBAs who can find such connections via their
b-school’s alumni network. We do not find a support for this
explanation. This channel also does not explain why the effect of
AI faculty departures is stronger when these professors are
replaced by faculty from lower ranked schools.
VC’s educational background. Venture capitalists may be more likely
to invest in startups whose founders share the same educational
background with them. In our case, it is possible that venture
capitalists are more like to invest in entrepreneurs who took the
AI courses from the same AI professors as they did. If this was the
case, then it might explain why the departures during [t − 6, t −
4] exhibit a negative effect but the departures during [t − 3, t −
1] do not. Moreover, this alternative explanation cannot explain
why startups in deep learning are disproportionately affected. It
also cannot explain why the results are stronger for startups
established within two years from students’ graduation.
Professors’ general skills. It is possible that professors who
leave for industry or establish their own startups have better
general skills, such as leadership skills. Therefore, instead of
AI- specific knowledge, students may just learn the managerial
skills from the departing professors.
23
However, any non-specialized skill transfer from professors to
students should be unrelated to the type of technology used by
startups. The fact that faculty departures disproportionately
affect deep learning startups is inconsistent with this
explanation.
Following professors. A professor’s departure for an industry job
may increase the probability that the professor’s previously
supervised students will join his or her current firm. When profes-
sors help students to get lucrative positions at their companies,
it could increase the opportunity costs of establishing a
startup.
It is plausible that poached students by departed professors are
more knowledgeable in the area of AI. If that is the case, then
this explanation is consistent with our argument that specialized
knowledge is needed for AI startups. The concern is that those
professors hire students not based on their AI knowledge but based
on some other traits (e.g., good communication skills). In this
case, when we see a drop in AI startups, we cannot attribute it to
the lack of knowledge or conclude that knowledge is
important.
The fact that in Table 6, we find deep learning startups to be most
affected is critical here, because it is advanced technology that
requires specialized knowledge but does not require more
communication skills or soft skills than other technologies. Also,
soft skills are much harder to assess without direct interaction.
If those soft skills are the explanation for the negative effect of
faculty departures on the number of startups by students, we would
expect to see this negative effect especially pronounced for
faculty departures that take place one to three years prior to stu-
dents’ graduation, rather than for departures that take place four
to six years prior to graduation. Our empirical results show that
the negative effect exists for departures that take place four to
six years prior to graduation and there is no effect for the more
recent departures. Moreover, this ex- planation cannot explain why
the negative effect on the startups is larger when the departing
faculty is replaced by faculty from lower ranked schools. We
conclude that the reduction in the number of startups is unlikely
to be explained by students following professors to join their
companies and to forego the opportunity to establish their own
startup.
Substitution to other fields. Departures of AI faculty could cause
would-be AI entrepreneurs to switch to other fields. For example,
if a student is interested in AI but majors in a different field or
switches a thesis topic because of the lack of faculty to teach AI
classes or supervise AI research. In columns (5) of Table 4 we test
whether AI professor departures affect the number of graduates from
the same university who later establish startups in other fields of
information technology. We do not find an empirical support for the
substitution effect because AI faculty departures do not
24
positively affect the number of non-AI startups in other fields of
information technology. Moreover, in Table 12, we do find that the
AI faculty departures do not have any impact on the early-stage
funding of non-AI (IT) startups established by affected university
graduates.
School selection. We would expect fewer students who are interested
in AI enroll to a univer- sity that has experienced a lot of AI
faculty departures. As a result, faculty departures could affect AI
entrepreneurship by a decline in enrollment of future AI
entrepreneurs, who instead enroll at other universities that are
less affected by AI faculty departures. If we are worried that
students might select another university after professors join tech
firms, we should be looking on this prob- lem from the perspective
of the students. Specifically, what is the information set
available to them? Do they search each professor on the
university’s websites and do they search professors on LinkedIn? It
is a plausible assumption that undergraduate students and master
students who are not expected to work on research would not do
that. We also have four tests to rule out this hypothesis.
In Table 9, we show that AI faculty who left partially (i.e.,
working for companies while keep- ing university affiliations) show
quantitatively similar negative effects to our main results shown
in Table 4. This result is inconsistent with school selection
explanation. Because the degree of the outside activity is not
necessarily observable for students in general and for
undergraduate stu- dents in particular. According to the selection
story, the negative effect should be stronger for full departures
than for partial departures. We find the opposite.
To further investigate the selection channel, we study the number
of AI startups in other univer- sities that are similar to the
focal university in terms of ranking or geography. If students
switch to another similarly ranked university or university in the
same locations as a result of faculty depar- tures in a given
university, then we should see an increase in AI startups founded
by graduates of the other universities. In columns (1) and (2) of
Table 10, we examine whether prospective students within the
universities that have similar CS department rankings will switch
to universities that are less affected by AI faculty departure.
Specifically, we divide the 69 universities in our sample into five
groups, based on the variable, Rank, that is defined in Table 1. We
then test whether AI faculty departures during the time window
[t−6, t−4] at one university can predict the sum of AI
entrepreneurs who graduate in year t from other universities within
the same ranking group. We do not find any evidence supporting the
redistribution of students within the universities that have
similar CS department ranking.
In columns (3) and (4), we examine the redistribution of students
within a region. We select
25
36 universities in our sample based on 12 regions, the details of
which are shown in Table 10. The dependent variable is the sum of
AI entrepreneurs who graduated in year t from a university in the
same region but not the university in question. For example, we
would regress the number of AI startups by Harvard alumni on AI
faculty departures at MIT. Consistent with the results from testing
the redistribution of students within the same ranking group, the
results from columns (5) and (6) do not support the redistribution
of students within a region.
Besides the tests of student redistribution, in Table OA5 in the
Online Appendix, we test whether AI faculty departures affect
universities’ ability to attract talented students. To measure the
quality of incoming students, we use the number of recipients of
the NSF Graduate Research Fellowship Program (GRFP) in the AI field
as a proxy for student quality. If faculty departures neg- atively
affect enrollment by high quality students, who later become AI
entrepreneurs, we would expect to see a drop in the GRFP recipients
following faculty departures. We do not find such an effect of
tenure departures neither in the [t − 3, t − 1] nor in the [t − 6,
t − 4] window on the number of the fellowship recipients in year t
at a given university. Together, these findings suggest that the
school selection channel is unlikely to explain the drop in the
number of AI entrepreneurs following AI faculty departures to the
industry.
5 Unobservable Factors
There is a potential concern that our results are affected by some
unobservable factors that are not controlled by fixed effects and
controls in the benchmark specification. A deteriorating local
economy is unlikely to be such an unobservable factor because if
our results are driven by the local economy, we should also observe
the impacts of faculty departures during [t−3, t−1] on AI
entrepreneurs who graduate in year t. Nevertheless, it is possible
that some local shocks affect both the AI faculty departures during
[t−6, t−4] and the AI startups by entrepreneurs who graduate in
year t, while they have no impacts on the AI faculty departures
during [t−3, t−1]. We address this omitted-variable concern by
conducting a placebo test at city-level and employing an
instrumental variable approach.
26
5.1 Location-Specific Factors
We conduct a test to address the concern that universities’
unobservable location-specific factors may confound our results.
Our test only focuses on the entrepreneurs who found AI startups in
a city where they did not attend a university. If our results were
driven by some unobservable local shock, this shock would hinder
startup activities overall, regardless of whether or not
entrepreneurs are alumni of local universities.
Table 11 provides a city-level analysis with the dependent variable
being the number of en- trepreneurs who found AI startups in city j
during year t but are not alumni of city j’s universities. For
independent variables, we aggregate the university-level measures
in previous tests by taking the average for each city and year.
Instead of a negative correlation between faculty departures and
non-alumni entrepreneurs’ local startup activities because both are
driven by negative local shocks, Table 11 presents a positive
correlation. This positive correlation makes it implausible that
negative local shocks explain our result that AI faculty departures
reduce the number of AI startups founded by alumni.
5.2 Instrumental Variable Approach
t−8 t−7
First-Round Funding
Series-A Funding
This figure plots the timeline for our instrumental variable
approach. For citations, we use AI faculty’s average citations
received during [t−8, t−6] at a given university. For AI
entrepreneurs in our sample, they start new AI startups, on
average, in year t + 2.2. (i.e., over two years after graduation),
receive their first-round funding, on average, in year t+3 (i.e.,
three years after the AI entrepreneurs have graduated), and receive
their Series-A-round funding, on average, in year t +4.5 (i.e.,
four and a half years after the AI entrepreneurs have
graduated).
Previous analysis shows that local or city-level unobservable
factors are unlikely to be the driving force behind the negative
impacts we documented earlier. To address the residual concern of
omit- ted variables, such as university-level unobservable shocks,
we rely on the instrumental variable approach.
27
5.2.1 Instrumental Variable: Average Number of Citations
We use the average number of AI-paper citations received during
[t−8, t−6] at a given university as an instrumental variable to
address the school-level unobservable factors.33 A high number of
citations is likely to attract more industry attention and thus is
more likely to generate an offer for the professor to leave
academia (Zucker et al. 2002). Moreover, given the time period of
our instrumental variable, it is unlikely that the number of
citations received during [t− 8, t− 6] at a given university would
be correlated with some school-level shocks that affect AI faculty
departures during [t−6, t−4] and startup creations and financing by
entrepreneurs who graduated in year t.
Since we use the average number of citations, it is possible that
the professors who actually left are those with fewer citations. To
verify whether the departing faculty members are actually above the
average in terms of the citations, we conduct a Logit analysis to
see whether the cumulative citations can predict the probability of
a professor leaving for industry. In Table OA6, after con- trolling
for the year fixed effects, the university fixed effects, and the
number of years since first AI publication, we find that the
cumulative citations are positively correlated with the probability
of leaving for industry.
To establish the causal link between faculty departures and the
outcome variables, we use the following 2SLS model:
First Stage: High Tenured Brain Drain j,[t−6,t−4] = αt +θ j
+β1Average Citation j,[t−8,t−6]+ γ1X j,[t−6,t−4]+ ε jt (3)
Second Stage (Int. Margin): Funding Amounti, j,k,t,τ = αt +θ j +β2
High Tenured Brain Drain j,[t−6,t−4]+ γ2X j,[t−6,t−4]+ ε jt
(4)
Second Stage (Ext. Margin): ln(1+AI Entrepreneur jt) = αt +θ j +β2
High Tenured Brain Drain j,[t−6,t−4]+ γ2X j,[t−6,t−4]+ ε jt
(5)
To improve the predictive power of the first stage, instead of
using the continuous variable to measure the AI faculty departures,
we construct a dummy variable, High Tenured Brain Drain[t−4,t−6],
that is equal to one if a university experiences above-median
departures in a given year (i.e., Tenured Brain Drain[t−4,t−6] is
above the median level in year t). It is easier to predict whether
a university will experience more faculty departures for industry
jobs relative to other universities than to predict what percentage
of AI faculty members leave for the industry based on the
univer-
33The results are qualitatively similar if we use average citations
received during [t−6, t−4].
28
sity’s average number of citations. It is important to emphasize
that there is a significant time gap between the departures that we
instrument and the outcomes that we measure. First, the outcomes
that we measure are for students who might not have been even
enrolled at the university at the time of the departure. Second,
with an average gap of two years between student’s graduation and
establishing a startup and an average gap of one year between the
startup’s inception and receiving funding for the first time, the
startups are established and financed on average about ten years
after the citations were received. These two features of the
empirical design provide further validity to the instrument.
5.2.2 Exclusion Condition
For the exclusion condition to hold, our instrument, the average
number of citations received dur- ing [t−8, t−6], needs to affect
the creation and the funding of startups by affected graduates only
via AI professors’ departures during [t− 6, t− 4]. Obviously, the
Google citations and the future number of startups at a given
university are both affected by the university-level characters.
How- ever, this concern is largely mitigated by university fixed
effect and time-variant computer science department rankings. Any
story that claims that the exclusion restriction is violated should
not be captured by the fixed effect and the rankings.
Moreover, given the large time gap between the citation and the
outcome, an example of a violation of exclusion condition should
also be able to explain why Google citations received in time
window [t−8, t−6] directly affect: (i) the number of startups by
students more than 6 years later (on average at time t + 2.2 ),
(ii) the amount of funding that these startups attract in the first
round (on average t +3), (iii) the amount of funding in round A (on
average t +4.5), and (iv) the growth rate in the amount of funding
between the first and the second rounds of funding.34 As a result,
it is unlikely that the citations received during [t − 8, t − 6]
would affect the number of startups and the funding that these
startups received a decade later for reasons that are not related
to faculty departures.
One major concern about the exclusion condition is that some
university-level unobservable shock (e.g., a loss of major donors
or a budget cut) may cause AI faculty with high citations to be
more likely to leave because their outside options are better than
those with low citations. In the meantime, those types of
university-level shocks could also negatively impact students
either
34For more information about firm-level statistics, see Table
2
29
because the affected universities are unable to attract
high-quality students or because they cannot provide enough
resources to students.
To address this potential violation of the exclusion condition, we
conduct a placebo test with the formation and the early-stage
funding of non-AI startups in the IT sectors as the outcome
variables. Given that the impact of a university-level shock is
hardly contained only on AI faculty and students who study AI, this
placebo test can address the concern that a university-level
unobservable shock could violate the exclusion condition for our
instrumental variable. The results from Table OA1 are inconsistent
with this concern.
5.2.3 The Results of 2SLS
Table 13 presents the 2SLS results for the extensive margin
results. Consistent with the results from the OLS regression in
Table 4, the second-stage results from column (2) show that the
coefficient on the fitted value of High Tenured Brain
Drain[t−4,t−6] obtained from the first stage is negative and
statistically significant at 5%.
Column (3) in Panel A of Table 13 shows the first-stage results
with a R2 of 0.53 and a Kleibergen-Paap F Statistic of 6.2. The
results show that universities with AI faculty who have a high
number of citations tend to have an above-median fraction of AI
professors leaving for an industry job. The first-stage results are
consistent with Zucker et al. (2002) who show that faculty quality
is a major determinant of professors leaving for an industry job.
Given that Kleibergen- Paap F Statistic is less than the
conventional threshold, 10, for weak instrument (Stock and Yogo
2002), we employ Anderson-Rubin (AR) test to draw a weak-instrument
robust inference. In just-identified case with single endogenous
regressor, Moreira (2009) shows that AR test is op- timal because
it is uniformly most powerful unbiased estimator. With 95%
confidence level, the confidence set for the coefficient on High
Tenured Brain Drain[t−4,t−6] is [−∞,−0.207]. Weak- instrument
robust inference indicates that the coefficient is negative and
statistically significant at 5%. Also, the 2SLS results are robust
if we use citations received during [t-6, t-4]. Moreover, using two
alternative instrumental variables, the citations based on papers
coauthored with authors who work in industry and the number of
industry coauthors, we get similar results for extensive margin.
This complementary analysis appears in the section OA.2 of the
online appendix.
Table 14 shows the effect of AI faculty departures on AI startups’
first-round funding. The OLS results in columns (1) and (2) show
that the coefficients on both departure measures, Tenured Brain
Drain[t−6,t−4] and High Tenured Brain Drain[t−6,t−4], are negative
and statistically significant
30
at the 1% level. This result implies that AI startups with founders
who are negatively affected by AI faculty departures, attract less
funding in the first round.
Columns (3) and (4) show the 2SLS results with the average number
of citations for AI pro- fessors as the instrumental variable. The
coefficient on the fitted value of High Tenured Brain
Drain[t−6,t−4] is negative and statistically significant at the 1%
level. Because the coefficient on faculty departure during the time
window [t−3, t−1] (Tenured Brain Drain[t−3,t−1]) is not statis-
tically significant in the OLS regressions, we do not include it in
the 2SLS. Moreover, the economic magnitude from the 2SLS regression
is much larger than that from the OLS model in column (2). Since AI
faculty from top schools are more likely to have high citations and
get poached by firms from industry, the variation of fitted value
of High Tenured Brain Drain[t−4,t−6] obtained from the first stage
mainly comes from top schools. As a result, the larger magnitude
from the 2SLS regression is consistent with our results from Table
5 that the negative effects due to AI faculty departures are
stronger for top 10 CS departments. In addition, the
Kleibergen-Paap F statistic from the first stage is 190, which is
greater than the threshold, 10, for weak instruments (Stock and
Yogo 2002), and the R2 from the first stage is 0.75.
Table 15 shows the effect of AI faculty departures on the series-A
funding round of AI startups. The OLS regressions in columns (1)
and (2) show that the coefficients on both departure measures,
Tenured Brain Drain[t−6,t−4] and High Tenured Brain Drain[t−6,t−4],
are negative and statistically significant at the 5% level,
consistent with the findings from the first-round funding
regressions. Columns (3) and column (4) in Table 15 report the
first-stage and the second-stage results of 2SLS respectively. The
coefficient on the fitted valu