Ten years ago, Mark Zuckerberg made a surprise appearance at the NeurIPS academic conference, announcing the launch of Facebook’s Fundamental AI Research (FAIR) unit, signaling that AI research had moved from university labs to the heart of Big Tech.
Fast forward to today: Meta announced drastic cuts at FAIR even as AI becomes a multi-billion dollar global industry. DeepMind no longer publishes technical details of its main AI models and introduced six-month embargoes and stricter internal document review to maintain a competitive advantage. Similarly, OpenAI is now called ClosedAI and, like other enterprise labs, increasingly favors technology blogs and internal product deployments over peer-reviewed publications or open source releases.
The wave of openness in the field of AI is receding – and with it, the foundations of scientific progress itself.
What we mean by “public good”
Open science is above all a public good: knowledge that benefits everyone rather than a privileged few. When research is shared openly, innovation accelerates, duplication is minimized, and ideas build on each other. In AI research, these shared open-source tools, datasets, libraries, and benchmarks have enabled advances from a lab and disseminated globally – from students to startups to large-scale industry deployments.
But when AI knowledge is privatized, we lose more than transparency: we lose the cross-pollination of ideas that drives true scientific progress. Universities and public institutions are uniquely positioned to take on this public good role because they are not structured primarily around shareholder returns or product deployment; they can prioritize openness, reproducibility, talent training and global participation.
How Openness Built Modern AI
The history of artificial intelligence is inseparable from the history of open science:
• The back-propagation algorithm, first shared openly in the 1980s, enabled the revival of deep learning.
• Effective deep learning techniques were subsequently developed in universities, notably for speech and image recognition in the laboratory of Geoff Hinton at the University of Toronto.
• Open datasets like TIMIT, TREC, MNIST, ImageNet, and Stanford Alpaca have provided reproducible benchmarks and common ground for AI advancements.
• Open source libraries/codes such as the Stanford CoreNLP toolkit and later TensorFlow, PyTorch and FlashAttention provided free access to cutting-edge techniques.
• Shared benchmarks and challenges (e.g., GLUE competitions, ImageNet) have trained generations of AI researchers and engineers.
This ecosystem created a flywheel of innovation: researchers published code and data, others used and improved them; students learned lessons from it; startups and industry have translated these advances into products. This was not accidental: it was the public good function of open science in action. In this context, the current decline of companies in relation to openness is worrying. This marks a shift from science as a shared effort to research as a proprietary product strategy.
Industry decline – and talent market failure
The retreat from open science is understandable: corporate AI labs face immense commercial pressures and fierce competition. Models are expensive, research is expensive, and first-mover advantage matters. Yet this change has broader implications for the public good and for education.
A striking indicator is the talent market: reports suggest that meta-platforms offered signing of packages of around 100 million dollars or more to top AI researchers in a desperate attempt to recruit elite talent.
This means a failure of the university market: the institutions that should train the next generation of talent simply don’t have enough computing, data, or the right balance of researchers and software engineers to meet the demand for experts in developing large AI models. Research students working within these larger teams are the right way to learn these important skills. If universities fail to train students in the ways required for future jobs, we lose not only individual opportunities, but also the broader workforce capacity needed for innovation and research for the public good.
The University’s Moment and the Public Good of Openness
Now is the time for universities to reaffirm their historic role in promoting AI as a public good. Academia and the nonprofit sector have the capacity to prioritize openness, ethics, shared infrastructure, and global access over short-term commercial gains.
This means investing in open data and open model initiatives that remain freely accessible for research and education; build global partnerships that share compute, data and expertise across borders and disciplines – so that knowledge is not siled in a few companies or countries; and foster interdisciplinary team science that integrates social sciences, ethics, design as well as technical research on AI to ensure that AI meets human needs and societal values.
Universities must not only publish, but also support the public good ecosystem that open science represents. In doing so, they preserve the foundations of talent development and discovery that fuel every advance in AI.
Carry the coat forward
At the Stanford Institute for Human-Centered Artificial Intelligence (HAI), we believe that the next chapter of AI must combine scientific openness with human-centered values such as dignity, fairness, and the common good. As industry prioritizes products and competitive advantage, our goal is to cultivate a network of global collaborations among like-minded universities, governments, nonprofits, and industry partners that uphold the public good mission. This is not about branding or competition, but about managing open science institutions and practices.
The most important problems facing the world require a new approach to scientific research: team science. Team science requires not only the broader collaborations of interdisciplinary academic researchers and software engineers that only industry has today, but also the calculations and data that go with it. We need new academic models to achieve the advances that team science will enable: distributed academic research centers, connected across continents, that share leadership, data, calculations, models and professional talent. Work in these centers will focus on human flourishing rather than commercial exclusivity.
The question is whether we will rebuild the open science institutions that made AI possible in the first place – or whether we will instead allow them to be eroded by concentrated commercial power. We have a fleeting opportunity to shape the trajectory of AI before it shapes us.
John Etchemendy, James LandayAnd Fei Fei Li are co-directors of the Stanford Institute for Human-Centered Artificial Intelligence (HAI), and Christopher Manning is associate director of Stanford HAI.