AI in Research: Integrity, Reproducibility, and Limits

Science works because of integrity. Its findings can be trusted only because they are honestly reported, properly attributed, reproducible, and subjected to scrutiny. AI, used carelessly, can undermine every one of these, which is why responsible AI use in research is fundamentally about protecting scientific integrity.

The previous lessons showed how AI accelerates research and where verification is required. This lesson addresses the standards that make AI-accelerated science trustworthy rather than corrosive: how to disclose AI use honestly, how to keep work reproducible, where the line between legitimate assistance and misconduct lies, and what AI simply cannot do in the scientific method.

The stakes are high because science is a cumulative, trust-based system. Each result builds on prior results that researchers take on faith were honestly and correctly produced. A single fabricated finding or corrupted analysis does not just harm one paper; it can mislead an entire line of research and erode the trust the whole system depends on. When a powerful, fluent tool can generate plausible falsehoods at speed, the safeguards that protect that trust become more important, not less.

So this final lesson is less about capabilities and more about responsibility: the practices that let a scientist use AI aggressively for acceleration while ensuring their work remains honest, reproducible, and genuinely scientific. The goal is to capture AI's benefits without compromising the integrity that gives science its value, because in research, a fast wrong answer is worse than a slow right one.

As AI became common in research, the scientific community established clear norms about disclosing its use and about who, or what, can be an author. These norms are now widespread across journals and institutions, and researchers must know them.

Two principles have broad consensus:

• AI cannot be an author. Major journals and publishing bodies have stated that AI tools cannot be listed as authors of a paper, because authorship carries accountability and responsibility that only a human can bear. An author vouches for the work and answers for it; a tool cannot. The humans remain fully responsible for everything in the paper, including any part AI helped produce.
• AI use should be disclosed. Many journals now require authors to disclose how AI was used, for example in writing, analysis, or figure generation, typically in the methods or acknowledgments. Transparency about AI's role is increasingly expected as a matter of honest reporting.

The reasoning behind both is the same: science runs on accountability and transparency. Readers, reviewers, and other researchers need to know how a result was produced to judge and build on it, and someone must be answerable for its correctness. AI disrupts neither principle as long as its use is disclosed and a human remains responsible.

The practical guidance for researchers: know and follow the AI-disclosure policies of your target journals and your institution, which are evolving, and never treat AI as a way to obscure how work was done. Being transparent about AI use is not an admission of weakness; it is part of the honest reporting that science requires. Concealing material AI use, by contrast, is a integrity failure, because it hides from the community something they need to properly evaluate the work.

The most important integrity boundary is between legitimate assistance and fabrication. AI can cross this line invisibly, so scientists must hold it clearly.

On the acceptable side, AI assists with real work you did:

• polishing the writing of findings you actually obtained,
• writing code to analyze data you actually collected,
• summarizing literature you then verify and read,
• helping structure arguments grounded in your real results.

On the unacceptable side, AI fabricates, producing content presented as real that is not:

• generating fake data or fabricated results,
• inventing citations to papers that do not exist (from the earlier lesson),
• creating figures or images that misrepresent actual findings,
• writing claims or conclusions the data do not support, presented as if they do.

The distinguishing test is simple: does the AI's output represent something real that you verified, or something invented presented as real? Assistance helps you communicate and analyze genuine work; fabrication manufactures false science. The first is a productivity tool; the second is research misconduct, as serious with AI as without it.

A subtle danger is that AI makes fabrication easy and convincing. It can generate realistic-looking data, plausible citations, and confident claims effortlessly, which can tempt a rushed or pressured researcher, and can also cause fabrication unintentionally if AI-invented content slips into a paper unverified. This is why the verification discipline of the earlier lessons is an integrity safeguard, not just a quality check: verifying every AI output against reality is precisely what prevents fabrication from entering your work by accident. The scientist's responsibility is to ensure that everything in their paper corresponds to real, verified science, no matter how it was produced.

The first lesson introduced reproducibility as a constraint on AI use; here it becomes a core integrity practice. A result is only solid science if others can reproduce it, and that requires transparency about how it was produced, which AI's opacity can threaten.

The integrity practices for reproducible AI-assisted research:

• Document AI's role. Record where and how AI contributed to the analysis or the work, in enough detail that another researcher understands what was done, consistent with journal disclosure requirements.
• Make the analysis inspectable. Where AI writes code, the code is the reproducible artifact: share it, document it, and version-control it, so the analysis can be rerun and checked. The reproducible record is the code and data, not an ephemeral AI conversation.
• Avoid hidden AI judgments in the pipeline. An analysis whose crucial step is an unexaminable "the AI decided" cannot be reproduced or scrutinized. Make the reasoning explicit and the steps repeatable.
• Be wary of non-determinism. Because general AI can give different outputs to the same prompt, do not build a result on a one-off AI output you cannot regenerate or justify.

The unifying principle is that AI should help produce transparent, reproducible science, not opaque, unrepeatable claims. Used well, AI writes the shareable code and drafts the documented methods that support reproducibility. Used poorly, it becomes a black box that undermines it. The difference is entirely in how the scientist integrates the tool.

This connects integrity to the everyday workflow: the reproducibility discipline is not a bureaucratic add-on but part of what makes AI-accelerated work count as science. A finding you cannot show others how to reproduce, because it rests on an untraceable AI step, is not yet a scientific result, however plausible it looks.

For all its power, AI has fundamental limits in science that no amount of capability removes, and understanding them keeps its use in proper perspective.

• AI does not verify truth. It generates plausible output, but in science, truth is established by experiment, observation, and replication, not by a model's confidence. A hypothesis is validated by testing it against reality, something AI cannot do for you.
• AI does not understand causation the way science requires. It can surface correlations and patterns, but establishing genuine causal mechanisms, the heart of much science, requires careful experimental design and reasoning that remains the scientist's work.
• AI cannot take responsibility. As the authorship norm reflects, accountability for a scientific claim must rest with a human who vouches for it.
• AI does not replace the scientific method. The cycle of question, hypothesis, controlled experiment, analysis, and peer scrutiny is what turns ideas into knowledge, and AI accelerates parts of it without replacing the method itself.

The deepest point is that science is not just producing plausible answers; it is establishing which answers are actually true, through a rigorous, communal, evidence-based process. AI is superb at producing plausible answers, which is exactly why it must be embedded within science's verification machinery rather than trusted in place of it. A fluent AI explanation is a hypothesis at best, not a validated finding.

Specialized scientific models, like structure predictors, are more than plausible-answer generators, they make validated predictions, but even these are typically confirmed experimentally for important cases, and they operate within, not outside, the scientific method. The final safeguard, for AI-assisted work as for all science, is peer review and replication: the community's checking of claims against evidence, which no AI removes the need for and which remains the ultimate guarantor that AI-accelerated science is real science.

Bring the whole cursus together into a coherent practice for using AI in research with integrity.

The unifying stance across the cursus is a rigorous partnership. AI is a powerful accelerator of research, speeding literature review, removing the coding bottleneck, clarifying writing, and, in specialized forms, making genuinely new predictions. But it is safe and valuable only within the discipline of science: verify everything, keep it reproducible, disclose its use, never let it fabricate, and remember that AI produces plausible answers while science establishes true ones.

The scientists who benefit most are those who use AI aggressively for speed and reach while holding uncompromising rigor over its output. They let AI handle the volume and the tedium, and they apply undiminished scientific judgment, verification, and integrity to everything it produces. That combination, acceleration plus rigor, lets AI advance science rather than pollute it.

The closing thought: AI does not change what science is, a disciplined, honest, communal pursuit of verified truth. It changes how fast scientists can do the supporting work and, through specialized models, what predictions are possible. Used with integrity, it is one of the most powerful research tools ever created. Used without it, it is a fast way to produce convincing errors. The difference is the scientist, and the standards they refuse to compromise.

AI accelerates research within science's integrity guardrails, human authorship and accountability, disclosure, no fabrication, reproducibility, and verification, while experiment, replication, and peer review remain the ultimate test of truth.