SQ8. What should the roles of academia and industry be, respectively, in the development and deployment of AI technologies and the study of the impacts of AI?
In most research areas, and historically in AI, there has been a relatively clear differentiation between the roles of academia and industry. Academics focus more on basic research, education, and training, while industry focuses more on applied research and development in commercially viable application domains. In the field of AI in recent years, however, this distinction has eroded.
Although academia and industry have each played central roles in shaping AI technologies and their uses, their efforts have been loosely coordinated at best. Now, as AI takes on added importance across most of society, there is potential for conflict between the private and public sectors regarding the development, deployment, and oversight of AI technologies.
The last five years have seen considerable debate regarding the appropriate roles and relationship of academia and industry in the development and deployment of AI applications.1 This debate arises from two facts. First, the commercial sector continues to lead in AI investment, research and applications, outpacing academia and government spending combined. In the US, private enterprises have spent over $80 billion on AI, while non-defense investment by the federal government in research and development is estimated at only $1.5 billion in 2020. Second, many researchers are opting out of academia for full-time roles in industry, and the long-term consequences of this shift are potentially worrying.2 To understand the extent to which these concerns might affect how AI develops and shapes society, we must consider the range of ideal roles academia and industry might play.3
Research and Innovation
It is now easier than ever to translate basic AI research into commercially viable products, thanks to the availability of relatively inexpensive, large-scale cloud computing, powerful open-source libraries, and pre-trained models for language, vision, and more. Access to such technology has created new incentives for university researchers, including faculty, postdocs, and graduate students, to create startups or seek other mechanisms to commercialize their intellectual property.
Meanwhile, the presence and influence of industry-led research at AI conferences has increased dramatically. For example, at the 2020 Neural Information Processing Systems Conference (NeurIPS), one of the premier, most widely attended and highly visible conferences in the area of machine learning, 21 percent of the papers were contributed by industrial researchers.4 This figure compares to 9.5 percent in 2005 (across all papers), the time at which this conference began to see a significant increase in submissions.5 This shift raises concerns that published research is becoming more applied (and perhaps less free to confront issues at odds with corporate interests), at the risk of stifling long-term innovation and value. On the other hand, industry’s increased presence might be helping catalyze the search for innovative solutions to real-world problems.
This increased mixing of academic and industrial research has raised concerns about the impact of "keeping up with the Joneses." A study of the amount of computing resources needed to train large natural-language processing models, such as the models known as transformers,6 noted that researchers trained nearly 4,800 models using the equivalent of 27 years of GPU compute time at an estimated cost of $103K-$350K (at cloud-compute market prices at the time). Such AI investments are impossible for most academic researchers.
Creating ways to share such models and evaluation environments would provide steps toward alleviating this imbalance. An interesting example comes from the decision by OpenAI to incrementally release the model parameters of their transformer-based GPT-2 network in 2019 and to provide access to its successor, GPT-3, in 2020.7 In 2021, the US National Security Commission on Artificial Intelligence recommended the federal creation of data repositories and access to large-scale computational resources.8 How to ideally allocate resources is an open problem that requires ongoing attention.
Research into Societal and Ethical Issues
As the line between academic and industry research in AI blurs, additional social and ethical issues come to the fore. Academic and industry researchers might have different perspectives on—and hence take different approaches to—many sociotechnical challenges that can be at least partially addressed by technical solutions, such as bias in machine-learned models, fairness in AI decision-making algorithms, privacy in data collection, and the emergence of polarization or filter-bubbles in social-media consumption.9
In addition, tighter coupling of academic and industrial research may reduce the focus on both longer-term problems and on those that might run counter to commercial interests. There is also a separate ethical consideration around IP ownership when students work directly with faculty whose intellectual products are partly owned by a company.
Companies that want to keep their customers satisfied have strong incentives to act when it comes to issues of privacy and fairness in the use of AI. One example of an action taken is corporate investment in The Partnership on AI, a nonprofit coalition of industry and university stakeholders committed to the responsible use of artificial intelligence.10 But an incentive to act is not necessarily aligned with the desire to get it right.
Development and Deployment
Application of advanced research and technology in real-world settings has traditionally occured outside of academia largely because of the high costs associated with development and deployment at scale. These include the costs of infrastructure, engineering, and testing; verification for robustness; and safety, logistics, and delivery—all of which are often most easily absorbed by companies with a commercial interest in deployment and the specific skills needed to manage these activities. While this dynamic remains largely intact in AI, the last few years have seen academic researchers increasingly able to take their technological innovations out of the lab and deploy them in the wild. A notable example is Duolingo, a language learning system built by academics at Carnegie Mellon, which went public in 2021 with a $5 billion valuation.11
Of course, not all real-world deployment is profit-oriented, and there’s no reason that nonprofit applications that benefit the public can’t be quickly created and put to use. For example, Oxford and Google collaborated on tracking COVID-19 variants,12 and, in the US, several universities are cooperating with companies c3.ai and Microsoft to promote urgent applications of AI for future pandemics.13 These developments have also played a major role in fostering non-commercial collaborations between industry and academia.
Education and Training
Many people from across the academic research spectrum have decried a perceived brain drain as a large number of AI researchers have left university and research institute posts to join the industrial ranks. Research suggests that this trend has intensified in recent years.14 According to one study, 131 AI faculty departed for industry (including startups) between 2004 and 2018, while another 90 took reduced academic roles for such pursuits.15 The study concludes that these departures had a negative consequence on Ph.D. training within the discipline. While there has not yet been a sustained dip in computer science Ph.D. graduates—nor, presumably, AI students—due to faculty departures, there is fear that one may develop.16
As student interest in computer science and AI continues to grow, more universities are developing standalone AI/machine-learning programs, departments, and related degree programs.17 Such programs include both traditional delivery and those that are partly, if not entirely, online.18 The trends outlined above raise questions as to who will staff these programs, and how they will feed into the pipelines needed to produce AI talent, from software and application developers to Ph.D. students and the next generation of academic leaders.19
A partial answer is to encourage industry to play a broader role in training. Internships, for example, where students spend a few months working in a company, offer current and recent students the ability to obtain valuable hands-on experience while addressing applied research questions or strengthening their skills in AI development and deployment. Such opportunities amplify university-based education and can often jumpstart students’ careers. Moreover, company-led courses are becoming increasingly common and can fill curricular gaps, especially if more students want access to basic AI education than universities can handle, or if students seek specialized skills that are best learned in the context of real-world applications.20
Societal Impact: Monitoring and Oversight
A controversy involving AI ethics research and researchers at Google in early 202121 spurred community-wide concerns about reliance on companies to monitor and govern their own ethics practices. For instance, a company can easily withdraw support from any ethics group or initiative whose findings conflict with its near-term business interests.
When it comes to the societal impacts of AI, stakes are high in the academia-industry relationship. Beyond questions of privacy and fairness lie concerns about the potential for AI and machine-learning algorithms to create filter bubbles or shape social tendencies toward radicalization, polarization, and homogenization by influencing content consumption and user interactions. However, studying and assessing these issues is easiest when academic-industry collaborations facilitate access to data and platforms.
Reducing some of the negative consequences of this more enmeshed relationship may require government regulation and oversight, particularly to guide how societal impacts are monitored, promoted, and mitigated. Any changes in regulation, however, should be made in consultation with the researchers who have the clearest idea of what the key issues are and how they should be addressed. Serious research is needed to guide effective policy, and that's where academic/industry collaboration can have the greatest impact.
 The National Security Commission on Artificial Intelligence Final Report (USA), 2021 https://www.nscai.gov/2021-final-report/; Kate Crawford, Atlas of AI, Yale University Press, 2021. https://yalebooks.yale.edu/book/9780300209570/atlas-ai
 Michael Gofman and Zhao Jin, “Artificial Intelligence, Education, and Entrepreneurship,” October 26, 2020. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3449440
 For an example of some existing efforts, see Yolanda Gil and Bart Selman, A 20-Year Community Roadmap for Artificial Intelligence Research in the US. Computing Community Consortium (CCC) and Association for the Advancement of Artificial Intelligence (AAAI), released August 6, 2019. https://cra.org/ccc/visioning/visioning-activities/2018-activities/artificial-intelligence-roadmap/
 See https://chuvpilo.medium.com/whos-ahead-in-ai-research-at-neurips-2020-bf2a40a54325. On the “counting methodology,” roughly speaking, credit for 1/n-th of a paper is given to each author of a paper with n authors, which is used to compute the fraction of all papers contributed by authors with either academic or industrial affiliation. For authors with multiple affiliations, their 1/n fraction is divided equally across all listed affiliations.
 Data drawn from comparing NeurIPS in 2020 versus 2015 here: https://papers.nips.cc/
 Emma Strubell, Ananya Ganesh, and Andrew McCallum, “Energy and policy considerations for modern deep learning research,” Proceedings of the AAAI Conference on Artificial Intelligence, 34(09), 13693-13696, 2020, https://ojs.aaai.org//index.php/AAAI/article/view/7123. We note that state-of-the-art large language models are even more massive than those reported in this paper (for example, Open AI’s GPT3 model has approximately 175B parameters).
 Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei, "Language models are few-shot learners." https://arxiv.org/abs/2005.14165v4; https://towardsdatascience.com/openais-gpt-2-the-model-the-hype-and-the-controversy-1109f4bfd5e8
 The National Security Commission on Artificial Intelligence Final Report (USA), 2021. https://assets.foleon.com/eu-west-2/uploads-7e3kk3/48187/nscai_full_report_digital.04d6b124173c.pdf
 We discuss separately below the question of meta-research, policy and oversight with respect to deployment, adoption and access. On the topics mentioned above, see Efrat Nechushtai and Seth C. Lewis, "What kind of news gatekeepers do we want machines to be? Filter bubbles, fragmentation, and the normative dimensions of algorithmic recommendations," Computers in Human Behavior, Volume 90, Pages 298–307, January 2019, https://doi.org/10.1016/j.chb.2018.07.043, https://arxiv.org/abs/1908.09635; Sorelle A. Friedler, Carlos Scheidegger, Suresh Venkatasubramanian, Sonam Choudhary, Evan P. Hamilton, and Derek Roth, "A comparative study of fairness-enhancing interventions in machine learning," Proceedings of the Conference on Fairness, Accountability, and Transparency, January 2019. https://dl.acm.org/doi/10.1145/3287560.3287589
 We discuss the evidence in detail below; see Michael Gofman and Zhao Jin, "Artificial Intelligence, Education, and Entrepreneurship," October 26, 2020. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3449440.
Cite This Report
Michael L. Littman, Ifeoma Ajunwa, Guy Berger, Craig Boutilier, Morgan Currie, Finale Doshi-Velez, Gillian Hadfield, Michael C. Horowitz, Charles Isbell, Hiroaki Kitano, Karen Levy, Terah Lyons, Melanie Mitchell, Julie Shah, Steven Sloman, Shannon Vallor, and Toby Walsh. "Gathering Strength, Gathering Storms: The One Hundred Year Study on Artificial Intelligence (AI100) 2021 Study Panel Report." Stanford University, Stanford, CA, September 2021. Doc: http://ai100.stanford.edu/2021-report. Accessed: September 16, 2021.
AI100 Standing Committee and Study Panel
© 2021 by Stanford University. Gathering Strength, Gathering Storms: The One Hundred Year Study on Artificial Intelligence (AI100) 2021 Study Panel Report is made available under a Creative Commons Attribution-NoDerivatives 4.0 License (International): https://creativecommons.org/licenses/by-nd/4.0/.