A medical writer is about to submit a Module 2.5 Clinical Overview, and by any honest accounting, sixty percent of the words on the page came out of an AI tool. Certara CoAuthor drafted the structure, populated the efficacy narrative from the loaded CSR, and produced the safety summary in a form the writer barely had to touch. The writer edited, verified, reconciled the citations, and tightened the prose, but the bulk of the text originated with the model. And on the cover page, under the line that reads "prepared by," is the writer's own name. Not Certara's. Not the model's. The writer's. That single fact, that the human name still goes on the document no matter how much of the draft the machine produced, is the load-bearing principle of responsible AI in life sciences, and it is the one this entire first level has been quietly building toward. This lesson is about why accountability does not, and cannot, transfer to the tool, why the FDA-EMA principles make this explicit, and why the named author owns the document and the named QPPV owns the PSUR even when most of the words were generated.

The Name on the Cover Page Is the Whole Point

Every regulated document carries a human signature, literal or functional, that says a named person stands behind its content. The Clinical Overview names its author. The Clinical Study Report names the responsible writer and the responsible signatory. The Periodic Safety Update Report, the PSUR, is owned in the European system by the Qualified Person for Pharmacovigilance, the QPPV, a named individual personally accountable for the safety system. These names are not formalities. They are the mechanism by which the regulatory system assigns responsibility to a person who can be questioned, who can be held to account, and who has attested that the document is true and complete to the best of their knowledge.

An AI tool cannot occupy that role, and the reason is structural rather than sentimental. A model cannot be questioned in an inspection, cannot attest to anything, cannot be held accountable, and cannot stand behind a conclusion, because it has no knowledge, no judgment, and no legal standing. It produces text. The accountability the cover page assigns has to land on a person, because only a person can do the things accountability requires: certify, answer, and bear the consequences. The proportion of the draft the model produced is irrelevant to this. Whether the human wrote ten percent or the model wrote ninety, the name on the cover page is a human's, and that human owns every word above their signature, including the words they did not personally write.

This is why the named-author principle has run through every lesson in this level like a spine. When we said the model completes patterns rather than retrieving truth, the consequence was that the author owns the gap between a plausible answer and a true one. When we said a fabricated citation is a Part 11 problem, the consequence was that the author who signed inherited it. When we said sensitive content must stay in a governed tool, the consequence was that the human, not the vendor, is responsible for the breach. Every thread terminates at the same place: a named person who is accountable for what the document says, regardless of how it was produced.

What Accountability Actually Means When Half the Words Are Generated

Accountability is often treated as a vague virtue, but in a regulated context it has a concrete operational meaning, and AI sharpens rather than softens it. To be accountable for a document is to be the person who has verified its claims against their sources, who can explain and defend every statement it makes, and who accepts the consequences if it is wrong. None of those obligations is reduced by the document having been drafted by a tool. If anything, AI raises the verification burden, because a fluent generated draft can carry confident errors that a blank page never could, and the author who signs is certifying that they caught them.

This reframes what the author's work actually is in an AI-assisted workflow. The writer who used Certara CoAuthor did not author the document by typing it; they authored it by owning it, which means they reconciled every efficacy number to the source table, confirmed every cross-reference points to a real target, checked that no analysis population was silently swapped, and satisfied themselves that the benefit-risk characterization is faithful to the data. The drafting was accelerated; the authorship was not delegated. When the author signs, they are not attesting "I typed this." They are attesting "I have verified this and I stand behind it," and that attestation is exactly as binding whether the first draft came from a keyboard or a model.

The trap to avoid is the quiet psychological shift in which a fluent AI draft feels finished, and the author's verification softens into a proofread. A model's even confidence makes its output feel authoritative, and a busy writer can slide from verifying claims against sources to merely reading the prose for sense. That slide is where accountability fails, because the signature still carries its full weight while the verification behind it has hollowed out. The discipline is to hold the verification standard constant regardless of how good the draft looks, because the cover page makes no allowance for the draft having been persuasive.

The QPPV Owns the PSUR, and the Model Cannot Share the Load

The pharmacovigilance world makes the principle especially vivid, because European pharmacovigilance law personifies accountability in a single named role. The Qualified Person for Pharmacovigilance, the QPPV, is an individual the marketing authorization holder must designate, who is personally responsible for the establishment and maintenance of the pharmacovigilance system and accountable for the content of safety deliverables, including the Periodic Safety Update Report. The PSUR, governed under ICH E2C(R2) as the periodic benefit-risk evaluation report, is a document on which the QPPV's accountability rests, and that accountability is personal and non-transferable by design.

AI now drafts large portions of a PSUR: it can assemble the line listings, draft the signal and risk evaluation sections, and produce the benefit-risk narrative far faster than manual writing. This is genuine value in a function drowning in writing volume. But none of it shifts the accountability one millimeter off the QPPV. When the PSUR makes a benefit-risk statement, that statement is the QPPV's professional judgment, attested under their name, and the fact that an AI assembled the supporting text does not make it the AI's judgment, because the AI has no judgment to lend. The QPPV who signs a PSUR is certifying the benefit-risk conclusion as their own, and they must own it as fully as if they had written every word.

This is why the most consequential PV judgments cannot be delegated to a tool no matter how capable it becomes. A causality assessment, a signal validation, a benefit-risk conclusion, these are determinations the named accountable person makes and defends, and the AI's role is to assemble and accelerate the supporting material, never to make the call. The QPPV who treats an AI-generated benefit-risk narrative as the conclusion rather than as a draft of the conclusion has misunderstood their own role, because the role exists precisely to place a named human judgment between the data and the regulatory record, and nothing the model produces can occupy that place.

Sponsor Accountability Does Not Transfer to a Vendor

The principle scales from the individual author up to the organization, and at the organizational level it takes a form that the FDA-EMA Guiding Principles make explicit: sponsor accountability does not transfer to a vendor or a tool. When a sponsor uses a vendor's AI system to help produce regulated content, the sponsor remains accountable to the regulator for that content. The vendor is not the one who answers the Information Request, defends the submission in a pre-approval inspection, or bears the consequence of a deficient filing. The sponsor is. Buying an AI tool, however sophisticated, does not buy a transfer of responsibility, because responsibility to the regulator runs from the regulated party and cannot be outsourced along with the drafting.

The FDA-EMA Guiding Principles of Good AI Practice in Drug Development, released on 14 January 2026, state this directly in the accountability principle: the use of AI does not diminish the sponsor's accountability, and that accountability does not transfer to the AI system or its provider. The principle exists because the natural temptation runs the other way, toward treating a validated vendor tool as if it absorbs some of the responsibility for what it produces. It does not. A vendor can be contractually liable to the sponsor for failing to meet a specification, which is a private commercial matter, but the sponsor's accountability to the FDA or the EMA for the truth and adequacy of the submission is undiminished and undivided. The regulator does not accept "the vendor's model produced it" as an answer, any more than it accepts "the model said so" from an individual writer.

This closes a loop that runs all the way through the program's treatment of tools. Vendor evaluation, validation, and contractual terms matter enormously, and later levels develop them in depth, but none of that machinery shifts the ultimate accountability. It manages risk on behalf of an accountable sponsor; it does not relocate the accountability. The sponsor that deploys AI across its regulatory, clinical, and safety functions is choosing to accelerate its work while retaining full ownership of every output, and the organizations that succeed with AI are the ones that internalize this rather than hoping the tool will carry weight it structurally cannot carry.

Why the Tool Cannot Absorb Judgment No Matter How Good It Gets

A natural objection arises as the models improve: surely, the reasoning goes, a frontier model good enough to draft a near-perfect benefit-risk narrative is good enough to be trusted with the conclusion, and at some point the human verification becomes ceremonial. This objection misunderstands what accountability is by confusing capability with standing. The reason the conclusion must be a human's is not that the model is too weak to reach it; it is that the model has no standing to be answerable for it. Answerability is not a function of accuracy. A model that was right ninety-nine times still cannot be questioned about the hundredth, cannot explain its reasoning under oath, and cannot bear a consequence, so it cannot occupy the role the regulation defines, however capable it becomes.

This is why the named-author principle does not erode as models improve; it is invariant to model quality, because it rests on a property the model will never acquire. A better model changes the verification economics, it produces fewer errors to catch, so the author's checking finds less to fix, but it does not change who must do the checking or who answers for the result. The temptation to let a highly capable tool's output stand unverified is therefore the most dangerous form of the accountability trap, precisely because the output is so good that the verification feels redundant. It is not redundant; it is the act by which a person converts a probabilistic draft into an attestable record, and that conversion is required no matter how few corrections it produces.

There is a deeper point here that the whole program rests on. Regulation assigns accountability to people because people are the only entities that can be held to account, and that assignment is a feature of how the system enforces truth, not a temporary limitation waiting for better technology. A future in which a model is accountable is not a more advanced version of the current system; it is a different system with no one to question when a submission is wrong. The named author is not a placeholder for an automated future. The named author is the mechanism, and the mechanism does not become obsolete because the drafting got faster. It is exactly as essential when the model writes ninety-nine percent of the draft as when it writes none.

The Disclosure and Documentation Corollary

If the human owns the output regardless of how it was produced, then owning it well requires being able to show how it was produced, which is where accountability connects to documentation and transparency. An author who is accountable for an AI-assisted document should be able to account for the AI's involvement: what tool was used, what it drafted, what sources it was given, and what the human did to verify the result. This is not a confession that AI was used as though it were a fault; it is the ordinary documentation that lets an accountable person demonstrate that their accountability was discharged rather than merely asserted.

This is why the program treats the audit trail and the emerging practice of disclosing AI involvement as part of the accountability picture rather than a separate compliance chore. A named author who can produce the record of how a draft was generated and verified is an author whose signature means what it says. One who cannot is an author asserting accountability they have no evidence of discharging, and that gap is exactly what an inspection or an Information Request is designed to find. The documentation does not create the accountability, which exists the moment the name goes on the cover page, but it is how the accountable person proves, after the fact, that the verification behind the signature was real.

The transparency principle and the accountability principle therefore work together. Transparency is the sponsor being able to explain how AI contributed; accountability is the sponsor owning the result regardless. An author who is transparent about AI use and accountable for the output has the complete posture the FDA-EMA framework envisions, and it is a calm posture, not a defensive one, because it rests on having actually done the verification the signature certifies.

The Capstone of the Responsible-AI Thread

This lesson is the capstone of the responsible-AI thread because every prior principle resolves into it. Bias matters because the named author owns a recommendation even when a biased model produced it, so they must interrogate the data behind it. Confidentiality matters because the named human, not the vendor, is accountable for protecting the information, so they must control where it goes. And the mechanics of how models work matter because the author owns the gap between a fluent draft and a true one, so they must verify what the model cannot. Each principle is a different facet of a single underlying fact: accountability lands on a person, and the AI changes how the work is done without changing who is responsible for it.

The equipped posture this level has built toward is exactly this clarity. The professional who understands that the name on the cover page is theirs, that the QPPV owns the PSUR, and that sponsor accountability does not transfer to a vendor is not anxious about AI and is not naive about it. They use the tools to move faster, draft more, and reduce the blank-page burden that has always slowed regulated writing, while holding firmly to the one thing the tools cannot touch: the human judgment, attested under a human name, that stands between the data and the regulatory record. That is the whole of responsible AI in life sciences, and it is why the writer's name still goes on the cover page even when sixty percent of the draft came from the machine. The model drafts. The named author certifies. That sentence is the foundation everything in the rest of this program is built upon.

Key Takeaways

• The name on the cover page is a human's no matter how much of the draft the model produced, because accountability must land on a person who can be questioned, can attest, and can bear consequences, and an AI tool can do none of those things; it has no knowledge, judgment, or legal standing.
• Accountability in an AI-assisted workflow means verifying claims against sources, defending every statement, and accepting consequences, none of which the drafting tool reduces. The author authors by owning, not by typing, and the signature attests "I have verified this," not "I typed this," regardless of how much the model wrote.
• The QPPV owns the PSUR, and AI cannot share the load, because a benefit-risk conclusion is a named human's professional judgment attested under their name; the AI assembles and accelerates the supporting material but cannot make the call, since it has no judgment to lend.
• Sponsor accountability does not transfer to a vendor or a tool, and the FDA-EMA Guiding Principles (14 January 2026) say so explicitly. A vendor may be contractually liable to the sponsor, but the sponsor's accountability to the FDA or EMA for the truth and adequacy of the submission is undiminished and undivided; "the vendor's model produced it" is not an answer.
• Documentation and transparency let the accountable person prove the verification behind the signature was real. Being able to account for what tool was used, what it drafted, what sources it had, and how the human verified it is how accountability is demonstrated rather than merely asserted. The model drafts; the named author certifies.