Does a private AI company have to follow the Common Rule at all?

The regulation binds federally funded research and institutions that have voluntarily extended their assurance to cover all human-subjects work regardless of funding. A self-funded company running its own product experiments is often outside its formal scope. The exposure arrives through three doors instead: collaborating with a university or academic medical centre, taking federal grant money directly or as a subaward, or working with data or participants from an institution whose assurance covers the whole project. Any one of those can make an otherwise unregulated experiment a regulated one.

What is the two-part test for human-subjects research?

First, is the activity a systematic investigation designed to develop or contribute to generalisable knowledge? Second, does it involve a living individual about whom an investigator obtains information through intervention or interaction, or obtains identifiable private information? Both parts must be satisfied. Product optimisation intended only to improve your own service typically fails the first part; the same experiment written up as a paper about how people respond to AI systems generally satisfies it, which is why intent to publish is such a reliable trigger.

Are prompt logs and conversation transcripts identifiable private information?

Often yes, and more robustly than teams expect. Private information covers information about behaviour in a context where an individual can reasonably expect no observation or recording, and information provided for a specific purpose that the individual can reasonably expect will not be made public. It is identifiable where the identity may readily be ascertained or is associated with the information. Free-text conversation is unusually rich in re-identifying detail, and stripping account IDs from a transcript that describes someone's employer, health situation and city is not de-identification in any meaningful sense.

Does the terms of service count as informed consent?

No. Informed consent under the regulation has required elements — a statement that the activity is research, its purposes and expected duration, procedures, foreseeable risks, benefits, alternatives, confidentiality provisions, whom to contact and an explicit statement that participation is voluntary and refusal carries no penalty. A clause buried in the terms accepted at signup satisfies almost none of them, and the voluntariness element in particular is hard to reconcile with participation being a condition of using the product at all. Waivers of consent exist and are sometimes appropriate, but they are granted by a review body, not assumed by the team.

Can we get approval after the experiment has already run?

Review bodies cannot retroactively approve completed research. Where an unreviewed study has already run, the realistic paths are a determination that the activity was not human-subjects research, a review of a secondary-use protocol over the existing dataset, or accepting that the work cannot be published. Journals and conferences increasingly require an approval statement or a formal exemption determination at submission, so this decision usually surfaces months after the experiment, when the cost of having skipped it is highest.

What about research on our own employees or contract annotators?

It raises the hardest version of the voluntariness question. Employees and contracted workers whose pay or standing depends on the requester are a population where genuine refusal is difficult, which is exactly the concern additional protections exist to address. Studies of annotator behaviour, model-assisted productivity or wellbeing effects on a workforce you control should be treated as sensitive rather than convenient, however easy the recruitment is.

You Called It an A/B Test. The Common Rule Calls It Research.

Both Gates, or Neither Applies

Gate 1 — Research?

A systematic investigation, including development, testing and evaluation, designed to develop or contribute to generalisable knowledge.

Satisfied by

+Intent to publish, present or post a preprint
+Hypothesis stated in advance and tested against a control
+Conclusions framed as being about people in general, not about your users
+A named collaborator whose output is an academic paper

Generally not

−Optimising a flow to raise your own conversion rate
−Internal quality evaluation with no external claim
−Operational monitoring and incident analysis

Gate 2 — Human subjects?

A living individual about whom the investigator obtains information through intervention or interaction, or obtains identifiable private information.

Satisfied by

+Manipulating what a user sees in order to measure their response
+Surveys, interviews and moderated sessions
+Analysing conversation transcripts tied to accounts
+Re-identifiable behavioural traces, however indirect

Generally not

−Benchmarks run against models with no human data
−Genuinely de-identified aggregate statistics
−Public documents about organisations rather than people

The useful consequence of a conjunctive test is that one honest answer can resolve it. An experiment whose findings genuinely stay inside the company clears gate one; a study that truly touches no identifiable human data clears gate two. Most disputes are about teams asserting both while planning to publish.

What the Team Says, What the Definition Asks

"It's just an A/B test, we run hundreds."

→

Is anyone going to write about what it showed? Volume is irrelevant to the definition; the intended contribution to generalisable knowledge is the whole of gate one.

"The data is anonymous — we removed the user IDs."

→

Could identity readily be ascertained from what remains? Free-text prompts routinely carry employer, location, health and family detail that no ID-stripping touches.

"Users agreed to research in the terms of service."

→

Were the required elements of informed consent present, and was refusal genuinely without penalty? Acceptance of terms as a condition of access answers neither question.

"Our university partner is handling the ethics side."

→

Is your company engaged in the research — obtaining data, interacting with participants, or receiving identifiable information? If so you have your own obligations, not a borrowed exemption.

"It's quality improvement, not research."

→

Where does the knowledge land? The distinction is defensible when findings stay internal and operational, and collapses the moment the same analysis becomes a public claim about human behaviour.

"It's minimal risk, so review isn't needed."

→

Who determined that? Minimal risk changes the level of review and may support an exemption, but that determination is made by a review body, not by the investigator who wants the answer.

Engagement Is How the Obligation Reaches You

Companies routinely assume a university collaborator absorbs the compliance work. Sometimes that is right and sometimes it is exactly backwards. The concept that controls the answer is engagement: an entity is generally engaged in research when its own employees or agents intervene or interact with participants for research purposes, or obtain identifiable private information for those purposes. A company that runs the experiment inside its own product, holds the raw transcripts and hands a partner an analysis is doing far more of the regulated activity than the partner is.

The practical version of this is a written determination made before data collection — who is engaged, whose review covers what, and whether a reliance arrangement is needed — rather than an assumption resolved by whoever wrote the acknowledgements section.

Six Designs That Deserve Review Regardless

Even where the formal regulation does not bind you, a serious deployment review is worth running on the categories below — because the reputational and litigation risk does not track the funding source, and because the first external question after an incident is invariably who reviewed this before it ran.

01Studies involving users who are minors, or products where age is unverified and minors are foreseeably present.

02Anything touching mental health, crisis behaviour, self-harm signals or emotional manipulation — the highest-risk category and the one most attractive to an AI research agenda.

03Deception or incomplete disclosure designs, which require a specific justification and usually a debriefing plan.

04Experiments on employees, contractors or annotators, where voluntary refusal is structurally difficult.

05Health-adjacent inference from behavioural data, which can pull in a separate regulatory regime alongside this one.

06Any design where the control arm is meaningfully worse for a user with a real problem in front of them.

The Cheapest Version of Getting This Right

A one-page determination form attached to any experiment where publication is plausible, answering both gates and naming who decided, costs a team almost nothing and resolves the entire question at the only moment it is cheap to resolve. Independent review boards will review commercial protocols for a fee where a formal record is needed. Both options are dramatically less expensive than the alternative, which is discovering at submission that a year of work is unpublishable and that the dataset behind it was assembled under a consent theory nobody would defend in writing.

Check What Your Site Says About User Data and Research

"We never use your conversations", "opt out anytime" and research-programme pages accumulate across docs, trust centres and old launch posts — and they are the record against which a consent theory gets judged.

See every claim your site makes in one pass. Run a free scan and check each against what your experiments actually do.

This article is general information and not legal or regulatory advice. The scope of federal human-subjects regulation depends on funding, institutional assurances and the specific facts of a study, and institutional policies frequently impose more than the regulation requires. Consult qualified counsel or a review board before relying on any conclusion here.