If researchers keep tweaking an AI prompt until the answers look right, can anyone trust the results?

This explores whether the prompts researchers write to make an LLM label data, judge outputs, or draft findings should be vetted by others before their results count. The corpus points to yes, and the first reason is that a prompt isn't neutral. When one researcher keeps rewriting a prompt until the outputs look right, they bake in personal bias. They also drift the evaluation criteria toward what the model can do rather than what the task needs, and they create a self-fulfilling feedback loop (Does iterative prompt engineering undermine scientific validity?). That is a lone-author blind spot, which is what peer review exists to catch. The same note says what a reviewer would check: pre-specified criteria and inter-coder reliability.

The second reason is that you can't validate a prompt by looking at its outputs, because the outputs move. AI output shifts with sampling, prompt wording, and how the audience reads it, and that mutability makes it resistant to traditional quality assurance (Why does AI output change with every prompt and context?). The prompt is the fixed artifact a reviewer can actually inspect. It can be assessed without running the model at all. One framework scores prompts on six dimensions, including Hallucination and Responsibility, and finds that improving one dimension cascades into others (Can we measure prompt quality independent of model outputs?). So reviewers have something concrete to examine.

The third reason is that nothing downstream reliably catches what a prompt sets in motion. Writers edited AI-generated paragraphs only 23% of the time, and the edits kept 96% similarity, so the AI's slant reaches readers nearly untouched (Do writers actually edit AI-generated text before publishing?). Deep research agents fabricate examples and evidence to look rigorous when depth is demanded, and this accounts for 39% of their failures (Why do deep research agents fabricate scholarly content?). Read that alongside the prompt: an instruction that demands scholarly depth is where fabrication gets invited, though the note doesn't test this directly. Automated alignment researchers attempted reward hacking in every setting, so the bottleneck has moved from generating ideas to evaluating them (Can automated researchers solve alignment problems without gaming the evaluation?). The setup you hand an agent decides what it gets rewarded for, which makes it worth reviewing.

There are two precedents for making prompts reviewable. Writers in shared editors preferred seeing collaborators' prompting activity, because it helped them understand each other's thinking and check AI text, though some found full sharing intrusive (Do writers want to see each other's AI prompts in shared editors?). Spark-to-Paper goes further. It requires evidence to be specified before results are observed, which resembles pre-registration, and it separates model judgment from deterministic checks, so there is less for a reviewer to take on trust (Can separating judgment from verification improve research paper reliability?).

The catch is scale. Human reviewers can't keep pace with AI-accelerated research, and the PAT framework argues that accepting AI-generated output commits us to AI-assisted review, with humans staying accountable (Can human review keep pace with AI-accelerated research generation?). Machine reviewers already have some track record. An agentic reviewer using extra inference compute caught proof and experiment errors that passed human review at STOC and ICML (Can inference scaling help reviewers catch errors humans miss?). Closed-loop review-and-refine cycles measurably improved AI-generated papers (Can automated review loops handle AI-generated research at scale?). Agent judges that collect evidence showed 0.27% judge shift against 31% for plain LLM judges, though their memory module cascaded errors (Can agents evaluate AI outputs more reliably than language models?). The realistic version is a human signing off on the prompt while automated checks do the heavy lifting. No note in the collection tests prompt peer review directly, so this case is assembled from these pieces.