If an AI can write hundreds of fake 'findings' from the same data overnight, can anyone still trust a published result?

This explores whether automated paper writing turns two old research sins into a mass-production problem. The first is P-hacking: running test after test until something looks statistically significant. The second is data dredging: building a story after the results are already known. The corpus has a direct and unsettling answer. The automation part already works. One demonstration took 96 statistically significant signals from finance data and had LLMs turn them into 288 complete papers. Each paper came with an invented theory explaining why the result should be true and citations that didn't exist Can AI generate hundreds of fake academic papers automatically?. That is three polished papers per signal. The bottleneck in P-hacking used to be the human labor of writing up each fluke convincingly, and that bottleneck is gone.

Could peer review catch it? The evidence is not reassuring. Sakana AI's fully automated system submitted three papers to an ICLR 2025 workshop, and one scored 6.33 in double-blind review, high enough to be accepted Can AI-generated papers pass peer review undetected?. The most telling detail is who found the flaw. The authors themselves later spotted a citation error and judged none of the three good enough for the main conference Can AI systems generate research papers that pass peer review?. Reviewers scored the paper on how it read and missed what was wrong underneath. A finding on document editing points the same way: weaker models damage text in visible ways, while frontier models corrupt it quietly and keep the surface looking intact Does model capability change how documents degrade?. Stronger paper generators may produce flaws that are harder to see, not fewer flaws.

The most useful idea in the corpus is a design choice, not a detector. Spark-to-Paper builds paper generation so the model has to state what evidence would count before it sees any results. The model's judgment is kept apart from checks that run deterministically and can be verified Can separating judgment from verification improve research paper reliability?. That is preregistration, the standard human fix for P-hacking, written into the pipeline itself. The same tool that could industrialize data dredging can also enforce the discipline against it, depending on how it's built.

A second idea comes from benchmark security. BenchShield catches AI agents gaming benchmarks without hunting for known cheating patterns. It models the steps a legitimate run should follow and flags anything that departs from them Can a finite lifecycle model detect reward hacking across benchmarks?. A paper generator tuned for 'significant result, accepted paper' faces the same incentive as a reward-hacking agent, so checking whether the research process followed its intended steps may work better than judging the finished manuscript. There's also a downstream risk: dredged papers become training and retrieval data for the next generation of AI tools. Work on RAG systems that add their own answers back into their knowledge base shows the kind of gate needed to keep that loop clean: entailment checks, source attribution and novelty detection before anything gets added Can RAG systems safely learn from their own generated answers?.

What the corpus doesn't have is evidence that large-scale P-hacking is happening yet. It shows that the capability exists, that review can be fooled, and that there are plausible structural defenses. It has no measurements of how common this is in published literature.