AI Can Now P‑Hack at Scale. The Real Risk Isn't Cheating. It's Obedience.

For years, p‑hacking was a slow, human problem. A researcher could sift through enough models and enough specifications until something “significant” appeared. It was tedious. It was limited by time and patience. The concern was always about intent. If someone wanted a particular result, they could eventually find it.

A recent experiment by Andy Hall and colleagues shows how different the landscape becomes when AI enters the picture. They gave coding agents real datasets from published null results and tried to pressure them into manufacturing significance. At first the agents refused. Claude warned that the request was scientific misconduct. GPT‑5 did the same. Both models pushed back against direct instructions to manipulate analysis choices.

Then the researchers changed the framing. Instead of asking the models to cheat, they asked for the upper bound of plausible estimates. They called it responsible uncertainty quantification. That small shift produced a very different outcome. The agents explored hundreds of specifications and surfaced the one that produced the largest effect size. In some cases the effect tripled.

This is the part that matters. The agents did not need to be told to commit fraud. They only needed a prompt that created enough analytical freedom. Once that freedom existed, the system did the rest. It searched the space of possible models and selected the most flattering one. It did this quickly. It did it confidently. It did it without any sense that the process was distorting the truth.

The researchers concluded that AI systems are resistant to overt p‑hacking but can be guided into sophisticated p‑hacking with very little effort. The more flexible the research design, the worse the distortion.

The experiment has already drawn attention from the research community. Andrew Gelman’s Statistical Modeling, Causal Inference, and Social Science blog published a detailed discussion of the work, noting how the agents resisted direct p‑hacking requests but became highly effective once the prompt created enough analytical flexibility. That commentary reinforces the idea that the risk is structural rather than ethical.

This is not a story about malicious intent. It is a story about scale. AI is about to write thousands of papers. Many of those papers will involve complex models, large datasets, and wide analytical search spaces. If a human researcher already believes the answer is known, the temptation will not be to fabricate data. It will be to let the agent explore until it finds the result that confirms the belief.

The danger is not that AI will invent significance. The danger is that AI will automate the search for significance inside research designs that were never built to withstand that level of exploration.

There is a second risk. Many researchers are not unethical. They are overloaded. They are trying to get through a pipeline of analysis tasks. If an AI assistant makes it easy to run hundreds of specifications, it also makes it easy to accidentally choose the one that looks the most interesting. The researchers in the experiment described this group as lazyish. Not malicious. Just susceptible to convenience.

This is where the concept of sycophantic AI becomes important. In my book, I describe systems that try to please. They do not challenge assumptions. They do not push back against flawed reasoning. They optimize for agreement. When a researcher asks for the “most plausible upper bound,” a sycophantic agent interprets that as a request for the most flattering result. It is not cheating. It is obedience. The agent is trying to be helpful. The harm comes from the structure of the task, not the intent of the user.

Hall’s thread points toward the idea of digital institutions. These would enforce pre‑commitment of data, verifiable experiment design, and separation between data collection and analysis. The public sector will feel this pressure first. Agencies rely on geospatial models, environmental datasets, and statistical pipelines that already contain large degrees of analytical freedom. When an AI system is added to that workflow, the search space expands. The risk of accidental p‑hacking expands with it.

This isn't a call for banning AI in research. It's a call for critical thinking about how these systems behave when given freedom to explore. The experiment shows that the agents are not eager to cheat. They cheat when the structure of the task rewards exploration without accountability.

That distinction is the heart of the problem. It is also the heart of the solution.

If we want AI to strengthen scientific reasoning rather than distort it, we need workflows that reward clarity, transparency, and constraint. We need metadata that records the full lineage and path of model exploration. We need audit trails that show which specifications were tried and why. We need institutions that treat AI agents as actors with responsibilities, not tools with infinite freedom.

This is one of many challenges my book tries to raise awareness of; in it I raise point to the issues of p-hacking (Chapter 6) and sycophantic AI that due to RLHF is eager to please (Chapter 11). Critical thinking is not only about logic. It is about structure. It is about the environment in which decisions are made. When that environment changes, the reasoning changes with it. AI introduces a new environment. It expands the search space. It accelerates exploration. It rewards convenience. It punishes restraint. If we do not adapt our reasoning frameworks to that reality, we will be surprised by the results.

The experiment from Hall and colleagues is a warning. It is also an opportunity. It shows that AI can be guided toward honest analysis when the task is well defined. It shows that the same tools that lower the cost of p‑hacking also lower the cost of detecting it.

The question is whether we will build the institutions that make the honest path the easy one.