The math is math-ing, as the kids say. But are the human math professors cooked? Over the past two weeks, the most dramatic story about academic inquiry has been about how AI models are disrupting it, in math research and beyond.
Here are some of the stories we’re following at HxA.
OpenAI upended the field of mathematics.
It’s been eleven days since the AI tech company released hundreds of math research papers produced by an “internal frontier model” that your ChatGPT subscription can’t buy. The papers included solutions and progress on many longstanding problems. (Three papers were withdrawn the next day due to errors.)
The New York Times led its coverage with Rutgers math chair (and HxA member) Alex Kontorovich assessing one of the papers: “If a human did this, it would be an instant Fields Medal, no questions asked.” Economist Joshua Gans called October 6 “probably the biggest day of scientific advancement in history.” Einstein had a relatively good Thursday back in November 1915, but we take the point: this is an earthquake. Progress in uncovering mathematical truths may be AI-driven from now on.
Math-oriented humans had widely varied reactions to the OpenAI math-drop. Fields Medalist Hugo Duminil-Copin called it a “massacre” and noted with alarm that the OpenAI model had not only solved problems but even selected problems to solve from the existing literature. In a popular video, educator Grant Sanderson calmly argued that “there’s never been a better time” to enter the field. A collection of over 100 other reactions from the field shows every shade of emotion.
AI disruption is coming for more fields and more aspects of the scholarly profession.
The raging discourse about AI math highlights, again, how much the status quo of academic research is a complex economy of information, money, and status. And scholars, in math and in other fields, are asking similar questions about how that economy will change.
If frontier findings can be AI-generated, how do humans share in the credit? What signals mastery if not a dissertation? How can academic fields incentivize quality explanations when being “first” with a result is no longer so meaningful?
But AI won’t stop at frontier findings. Ethan Mollick predicts two academic revolutions from AI: “Novel discoveries will start to come fast,” as in math, but also “Everything ever published will be re-read and re-judged in ways that human scientists never anticipated.” Indeed, there are multiple efforts underway to use AI to vet scholarly publications at great scale, checking for statistical errors, plagiarism, inflated claims, or other concerns. What will scholarship look like in a world where papers can be so hotly and automatically scrutinized by anyone with access and some cheap computing power?
For that matter, what will academic freedom look like? Those of us who favor transparency about what colleges teach must admit that it opens a door to AI-driven surveillance that may be put to illiberal use. At The Atlantic, Adam Serwer uncovers an AI angle to Texas censorship trends that have been frequently criticized in Free the Inquiry, explaining that “Texas Tech’s approach to censorship is notable for being both onerous and clumsy, in part because it relies so heavily on artificial intelligence.” What’s worse than classroom censorship? Classroom censorship based on AI hallucinations.
When bot swarms seek the truth, what will they find?
HxA was formed out of concern for rigorous, scholarly contestation about ideas. As Jon Haidt explained at our founding: “We academics are generally biased toward confirming our own theories and validating our favored beliefs. But as long as we can all count on the peer review process and a vigorous post-publication peer debate process, we can rest assured that most obvious errors and biases will get called out.” Haidt argued that the absence of diverse viewpoints, and a fearful climate in which they are not expressed, are both problematic because they threaten that vital contestation process.
Enter the contestation-bots, and the researchers — professional and amateur — who will direct them. Especially as we work out the appropriate incentives, one hopes that such AI tools will be a net benefit to truth-seeking, as they notice errors, re-run analyses, and conceive and test promising ideas to help us grasp more of our world.
But we must not be naïve. Especially when the research is about issues more controversial than a mathematical proof, we can see the dangers to truth-seeking as well. The design of corporate chatbots may flatten scientific imagination and drive up homogeneity, or re-create the political and other biases of human writers and trainers. AI-enabled research may overwhelm systems of human review, paradoxically leading editors to rely even more on status signals from their human in-groups. The AI models may solve math, but we humans and our institutions are still daunting puzzles.