There are foreseeable ways AI-assisted research—and people who misuse AI—can harm human subjects and corrupt scientific inference. We can now avoid them in a context, ironically, powered by AI.

by James Lyons-Weiler, PhD, Popular Rationalism, ©2026

(Sep. 14, 2026) — AI can make a bad protocol look finished. It can produce polished prose, tidy tables, and an apparently comprehensive reference list before the investigator has done the hard work of deciding what would disconfirm the hypothesis, which variables may be colliders, whether the endpoints were defined in advance, or whether a data field can identify a participant. That is precisely why fluency is dangerous here: it can create the appearance of control where no scientific control exists.

In clinical research, the consequence of cheating reaches people. They may be recruited, exposed, reassured, classified, or treated on the strength of a method that cannot support the inference being made. Clinical practice based on nonsense can obviously harm. The Common Rule makes the connections explicit: an institutional review board must determine that risks are minimized through procedures consistent with sound research design and that remaining risks are reasonable in relation to anticipated benefits and the importance of the knowledge expected. When a study cannot produce reliable knowledge, the risk-benefit bargain degrades. Methodological failure becomes an ethical failure. In clinical research, methodology is king.

The modern standards point in the same direction. ICH E6(R3) places quality by design, participant protection, reliable results, and data integrity at the center of good clinical practice. The FDA’s January 2025 draft guidance on AI used to support regulatory decision-making (still draft and nonbinding) uses a risk-based framework tied to context of use, model risk, credibility assessment, documentation, and lifecycle maintenance. Speed does not displace responsibility. But tools built with compliance and enforcement in mind can increase adherence to responsibilities.

Abuse and neglect are different routes to the same failure. AI can prevent both.

Some AI-related failures are acts of abuse: inventing observations, manufacturing authority, generating analysis paths until the preferred answer appears, or disguising retrospective choices as prospective commitments (or vice-versa). Others are acts of neglect: failing to open a citation, failing to inspect an adjustment set, sending identifiable information to a system before de-identification, or failing to preserve the version of a model that influenced a decision. The intent differs. The participant and the scientific record can still absorb the damage.

AI lowers the time and technical friction required for both. It does not invent misconduct, bias, or carelessness. It can industrialize them. A responsible research platform must generate useful material while restoring friction at the exact moments when a person must verify, justify, protect, or sign off.

1. Protocol laundering

Researchers have long known the danger of HARKing (Hypothesizing After the Results are Known): presenting a hypothesis formed after results were known as though it had been specified in advance. AI adds a more scalable variant. A user can ask a model to reverse-engineer a hypothesis, endpoint, eligibility rule, or mechanistic rationale that makes an observed result look inevitable. The prose may be coherent. The chronology is false. I call this protocol laundering: exploratory decisions are washed through fluent language until they resemble prospective science.

The Study Design Architect is designed to force that chronology back into daylight. Research questions, hypotheses, endpoints, inclusion and exclusion criteria, and analysis commitments are placed into an up-front structured record. Revisions remain revisions, but they do not erase what came before. Version histories, visible differences, rollback, and human approval preserve the distinction between what was planned and what was learned. Exploratory work remains legitimate, but it must be labeled as exploratory.

Thus is was, thus it should always be.

2. Phantom evidence

A fabricated citation is not a cosmetic defect. It creates false authority. A real citation that does not support the sentence attached to it can be just as misleading. In a 2023 evaluation of 636 references generated by then-current ChatGPT models, 55 percent of the GPT-3.5 references and 18 percent of the GPT-4 references were fabricated. Those figures are a historical snapshot, not a benchmark for every model now in use. The underlying failure mode remains: plausible bibliographic syntax can be mistaken for evidence.

The Suite’s Citation Checker does not accept a reference because it looks scholarly. The record must resolve through a DOI, PMID, or other verifiable identifier, and the cited source must be checked against the proposition it is being asked to support. An unresolved citation or a content mismatch is surfaced rather than silently passed forward. The Power Curve applies the same principle to sample-size planning: effect-size estimates are drawn from PubMed material only when the numerical estimate and its confidence interval can be verified; when an abstract contains several estimates, the system presents them rather than choosing the convenient one for the user. And it provides reminders that attrition may ablate power, that subgroup analysis power is relevant, and that power-to-detect rare adverse events requires an a-priori definition of ‘rare’.

---

Read the rest here.