Note: `Decorrelation' does not mean `no correlation'
Over a period of perhaps 100 years, several academic pursuits/goals tended to reinforce one another more directly. For instance, a PhD degree trained a student, who solved an open problem, advanced the field, produced public knowledge, certified expertise, and generated ideas and talent that industry could use.
Of course, this correlation/alignment was never perfect, and of course, it was weakening before the current wave of AI. Research has become more specialized, collaborative, capital-intensive. Tractable problems are being replaced by problems with emergent complexity, and solving them may perhaps have utilitarian value (or not), but not necessarily advance understanding or the field. Teaching has long separated from frontier research, and publications (in the mean) have become noisy* and imperfect proxies for knowledge and mastery. In some fields, companies have been advancing the frontier faster than universities.
Let me make it clear that all of this was true even before the current wave of AI.
Over the past couple of years, AI has introduced a stronger decorrelation. The recent Fields medalists statement on AI and mathematics captures this sentiment well: It says that solving difficult problems was historically a useful proxy for producing conceptual understanding. That relationship is becoming less reliable and AI system may produce a proof, design, code, or prediction without training a person or providing a compact human explanation. It also says that, a person may develop considerable understanding without being the first to reach the answer.
This is a well-crafted/intended statement, but it needs some augmentation
1. Nothing is preventing humans from using AI to break down its solution into a form that advances human understanding. In fact, that is an easier task for AI than solving otherwise intractable problems!
2. In many cases, the nature of the problem is the main issue. Not the tool used to solve it! Science itself has been changing.. it used to progress by identifying pockets of reducibility (as Wolfram would say) and lead to tractable problems: phenomena that could be compressed into understandable mechanisms, and manageable chains of reasoning. In many fields, the remaining frontier is increasingly characterized by interacting scales, large search spaces, incomplete observations, nonlinear feedback, and strong context dependence. The supply of important problems that fit comfortably within a single human conceptual framework may be shrinking.
3. AI can help with this emergent complexity, but the resulting progress and incentives may look different. A verified theorem or a working drug remains valuable, but the stack of answers could grow faster than the stock of human understanding.
Universities will (and should**) continue to perform their usual roles: discovery, education, certification, knowledge stewardship, workforce development, and the creation of intellectual work. The difference is that success in one role may no longer imply success in the others.
Universities will (and should**) continue to perform their usual roles: discovery, education, certification, knowledge stewardship, workforce development, and the creation of intellectual work. The difference is that success in one role may no longer imply success in the others.
For a long time, these distinctions did not have to be made very sharply because the goals were highly correlated. AI does not make any of them unimportant. It simply makes their separation harder to ignore, and we can be more forthcoming about this, and redefine certain value propositions and incentives. Here's a mathematician wondering how to shift academic credit from "proofs" to "understanding". If the math community can have such discussions, it is a positive sign for the broader community.
Personally, I think that further decorrelation is inevitable and universities should try to sharpen and double down on what they are good at / more uniquely positioned to do well. I think curiosity-driven research should be #1 on the list, for instance.
=======
* Example: ICLR 2027 has apparently gotten 60k+ submissions. More than all the previous years combined.
** There is no other organization that is capable of producing curiosity-driven long term research and in training the next generation of scientists (AI jingoists might tell you things like 'you don't need formal education.. you can learn by yourself what you need to solve a problem' which have little merit because the world in which people are learning from AI is still a world that is underpinned by formal education. Of course, many things can be done better at universities... but...