On November 19, 2026, I participated in a panel discussion at the DSTS / DSBS Joint Two-day Meeting on Statistics in an AI era. I gave the talk p < 0.05, Trust Me — Says the Machine framing the panel debate. The slides from the talk are linked above. This post summarizes my main points.
The leading question
Did the statistics profession just die due to the capabilities of frontier AI models? I don’t think so, but the answer may be uncomfortable.
What happened
In two years, coding agents went from handling tasks of a few minutes to tasks of many hours. In 2026, AI arrived in research mathematics. In September, OpenAI claimed a Lean-verified blow-up result for Navier–Stokes, and in October it released 722 AI-written mathematical manuscripts.
Mathematics was hit first for good reasons: it exists fully in digital form, correctness is objectively verifiable, and there is a large body of high-quality training material.
Statistics is affected too, substantially on derivations and proofs, on coding for analysis and reporting, and on study design and method selection. At the moment less so on which questions are worth asking, who is accountable, and oversight.
The fragile position
The job market for young people in AI-exposed jobs is already shrinking, in the US and in Denmark. Programmers, mathematicians and statisticians whose identity is tied to producing code, proofs or data analyses are facing an existential crisis.
A common position is that humans are needed because AI cannot do something we can. For example, write good papers, think “truly” creatively, or have human judgment. Until recently, few mathematicians believed that AI could solve open research problems. I find that any position based on AI cannot do X is fragile.
The defendable position
Assume instead that anything humans can do, AI can do, possibly better. Do we then roll over? The central question is this: Do we want a world where no human, not even in principle, can follow and verify a claim or decision made by AI?
My answer is no!
Humans should claim two roles: regulation (what can AI do?) and oversight (does AI comply?). I find it essential that we claim these two roles to be able to participate in decision processes in the future.
Two more roles are ours if we want them: framing the problem we want to solve, and valuation, making tradeoffs between values with no fixed exchange rate. It is irrelevant if we are “better” than AI in these roles. What matters is again that we take these roles to participate in the future development and not be mere spectators.
These roles are not defined by capacities that AI lacks. They are positions we can insist on. Whether oversight of a superhuman AI is possible at all is an open question, and it requires institutional support and the overseer’s real possibility to reject. But I see it as imperative to work on oversight, framing and valuation as the key human roles.
In concrete terms, a p < 0.05 claim must be justified through a transparent analysis, all the way. It must answer a question we actually want to ask, and we must have the right to contest it. That “all the way” matters is not a new insight, but a recent study demonstrates its contemporary value: coding agents report p < 0.05 in far more than 5% of simulated null examples ( arXiv:2611.23174). This is a well-known post-selection inference fallacy, which humans also commit, but it highlights why we should not accept claims from AI agents uncontested. We should, indeed, never accept claims on authority alone, but only when they can be transparently supported.
Our value should not be attached to our replaceable productivity skills but to our abilities to reason about and question claims or decisions. Human value is then attached to maintaining our agency.
Education
If humans are to fill the roles above, we need education to continue as an institution, and humans to keep the framing role in it and the oversight of it. We cannot educate narrowly toward a rapidly changing job market or focus on productivity skills. We cannot even know if there is a job market, but we can insist on human agency being of value in itself and that the education system builds on that.
For statistics, we need less tool training (PROC TTEST, t.test()) and more understanding of what H₀ and p < 0.05 actually mean. It means a sharper focus on limitations and breakdown points, on adversarial behavior, both by AI systems and in our own analyses, and on critical assessment over productivity tooling. Statistics is to me the fundamental formal science of empirical epistemology, where we investigate what we can learn from data and how to do that. Humans need to continue thinking about this and working with these problems, even if we outsource a great deal of the hard work.
For years, mastering SAS or R was a prerequisite for functioning as a statistician. That is no longer the case: reviewing code matters more than writing it from scratch. And even the reviewing may be partly automated, as long as humans maintain the ability to validate and be ultimately accountable if need be. The open question, as so often when a tool replaces manual labor, is what competences we lose along the way.
The risks are real. The connection to the job market is uncertain, individual incentives point to credentials rather than agency, and what we actually want from education is harder to measure. But the case for mathematical and statistical training does not rest on beating the machines at production. It rests on being able to follow, question and contest what machines as well as other humans tell us. It rests on humans maintaining their agency.