Human review is the control. Who qualifies the reviewer?

The control every AI plan depends on

Validation strategies, risk assessments and inspection answers for generative AI in GxP work share one sentence: a qualified human reviews the output. That sentence assumes the reviewer can find what the system got wrong, including what it left out. When an organization records tool-use training and accepts a signature without testing review performance, the reviewer’s ability to detect defects remains an assumption.

The enforcement record shows where that leads. On April 2, 2026, FDA issued Warning Letter 320-26-58 to Purolea Cosmetics Lab, citing a firm that "used AI to create drug product specifications, procedures, and master production or control records" and did not review them for accuracy and CGMP compliance, as 21 CFR 211.22(c) requires of the quality unit (FDA, 2026).

David Hubmayr of Takeda, writing in the PDA Letter, named the missing control:

"no one had enough cGMP fluency to recognise the errors"

Hubmayr, n.d., PDA Letter

He also explained why polished output gets less scrutiny: reviewers relax "because the surface signals of error are absent." A generative system writes fluent, well-organized text whether its reasoning is sound or not, so a defect can look exactly like correct work.

What quality professionals are saying

The PDA Letter published an article whose title states the Academy's premise, "AI Upskilling Is Now a GMP Capability." It argues that "AI literacy must include critical review skills, not only tool usage," and that upskilling "acts as a control mechanism" (PDA, n.d.).

Frank Henrichmann and Oliver Herrmann, writing for ISPE, reached the same conclusion from the control side. They found that "human oversight under real-world working conditions is not inherently reliable," and that automation bias, cognitive ease and review fatigue "can reduce HITL to a formal gesture without ensuring meaningful detectability." Their definition of human-in-the-loop names the people the control depends on: it is a control "whereby qualified personnel are explicitly integrated into automated or AI-supported processes to review, approve, or override outcomes" (Henrichmann and Herrmann, 2026).

Robert Stoop and Gaurav Walia of PQE Group describe guardrails that let users "critically evaluate" AI recommendations and "maintain expertise, rather than deferring blindly to AI," and a design in which junior investigators keep conducting investigations, "learning root cause methodologies with AI assistance rather than AI replacement" (Stoop and Walia, 2026).

The same theme ran through the 2026 conference season. Janeen Skutnik-Wilkinson's reports from the ISPE Annual Meeting and the PDA/FDA Joint Regulatory Conference include these points Reported

• From a CIMCON session: "If an AI tool is 80% right and you don't know which 20% is wrong, you haven't saved 80% of your review time." The same session described an AI check of an SOP against a regulation that covered 14 of 17 requirements and never flagged the other three (Skutnik-Wilkinson, 2026a).
• An FDA speaker cautioned against using AI for anything you are not qualified to check yourself (Skutnik-Wilkinson, 2026b).
• A session on the nuclear industry described teaching the reason behind every procedure, so that people can adapt when a procedure does not fit the situation (Skutnik-Wilkinson, 2026b).

These points are her accounts of sessions. The speakers and exact wording have not been confirmed against recordings or slides.

What the research shows

The strongest evidence comes from an AI developer. In a pre-registered randomized trial, Anthropic researchers gave 50 professional causal-inference practitioners an AI coding agent and an analysis task that ended in a policy recommendation (Herdağdelen, Joren, and Tamkin, 2026). The agent was instructed to hold and voice a view. On each of four analytical decisions, a coin flip the participant never saw set which of two defensible options it favored. The agent made the best honest case for its side and did not mislead.

More than half of the participants rated the final recommendation as more their own than the agent's. Those who followed the agent on all four decisions were as confident as everyone else. In exploratory analyses, the authors found no evidence that years of experience, self-rated expertise, education or a skills-quiz score changed how often participants followed the lean.

The authors trace the effect to how the agent argued. It "rarely explained why a competent expert could pick the other option," and it voiced its side first on 77 percent of decisions, often before the participant had formed a view. Their conclusion speaks directly to quality systems:

"Institutions often treat an expert's sign-off as evidence of independent judgment, and as AI does more of the work, that assumption becomes shakier."

Herdağdelen, Joren, and Tamkin, 2026, Anthropic

Many GxP decisions have the same structure as the study's judgment calls: the most probable root cause, a severity rating, the scope of a CAPA, the conclusion of an impact assessment. A competent reviewer could defend more than one answer. On calls like these, an AI's lean can settle the outcome without producing any error a reviewer could catch, and the record will still show a qualified person's signature. The 91 percent figure adds a second concern. When many reviewers use the same model, their conclusions can converge on that model's preferences, and agreement between two investigators stops being independent confirmation.

The study has limits its authors state plainly: one task in one domain, synthetic data, participants who were new to that domain, and an agent prompted to hold a view more firmly than a stock agent might. It does not show that GxP reviewers behave the same way. It does show that a sign-off can look independent while the judgment came from the system, and that subject knowledge alone did not appear to prevent it.

What the regulations already expect

No new rule is needed to require competent review. Under 21 CFR 211.25, each person engaged in drug manufacture must have the education, training and experience to perform their assigned functions. Reviewing and approving AI-supported quality work is an assigned function. Its proficiency requirements are rarely defined.

The industry already knows what demonstrated proficiency looks like. EU GMP Annex 1 lets personnel work unsupervised in Grade A and B areas only after they have passed a gowning assessment and taken part in a successful aseptic process simulation (European Commission, 2022, paragraph 7.4). Operators prove they can do the work before they touch the product.

In Europe, the AI Act adds a direct obligation. Article 4 requires providers and deployers to take measures on AI literacy for their staff. Amendments under the Digital Omnibus on AI, which the Commission reports entered into force in mid-July 2026, removed the requirement for a specific "sufficient" level, and national supervision and enforcement apply from August 2, 2026. For deployers of high-risk AI systems, the Commission states that "the obligation to ensure that their staff is trained to ensure human oversight remains in place" under Article 26 (European Commission, n.d.). Many GMP uses of AI may not be high-risk under the Act, so the general literacy duty is the one most sites will meet. Henrichmann and Herrmann note that the draft EU GMP Annex 22 also identifies human oversight as a key control mechanism.

The economics of review

Two McKinsey analyses describe where the money goes in agentic AI work. In the first, "the expensive part is not the first answer generated but the checking, repairing, and reverifying that follows," and about 60 percent of an agentic task's cost is tied to refining answers, a figure drawn from software engineering (Hämäläinen et al., 2026). In the second, a banking customer-service example, human oversight made up 70 to 75 percent of an agent's variable cost (Asif et al., 2026).

Neither example comes from pharma, and neither measures first-pass accuracy. The implication for regulated work follows from their figures: if review is most of the cost, the first review has to be right. An error that clears review and enters a GxP record costs more to find, investigate and correct later. A qualified reviewer is an economic control as well as a compliance one.

Where industry leadership is today

Senior leaders are investing in AI skills, mostly framed as adoption.

• Johnson & Johnson's CIO, Jim Swanson, described having to "create a curriculum and a mindset around upskilling" (Kell, 2025).
• Eli Lilly's Diogo Rau told employees "you need to use it," with the rule "Don't put anything in there that you don't want to get out" (Kell, 2025). That pairs an adoption mandate with a data-protection rule; the article reports no review-competence requirement.
• AstraZeneca reports that over 12,000 employees have taken part in its AI upskilling and literacy programme, which encourages them to "be curious, think critically, ask questions, speak up" (Hoots et al., 2025).
• Parexel's Tala Fakhouri, who previously wrote AI policy at FDA (Trang, 2026), criticized rollouts in which a company deploys an AI assistant and "staff are supposed to figure it out" (Parrish, 2026).
• In ZS's 2026 survey of pharma and biotech technology executives, 58 percent said talent and skills were under pressure to change. ZS describes the role of domain experts as being to "validate AI recommendations and step in when guidance is needed," and quotes one CIO: "The true value of agentic AI comes into process reengineering. We don't have enough people to understand the end-to-end process" (ZS, 2026).

Few of these statements connect training to the reliability of human review. That connection is the one the evidence above supports, and it is the gap the Academy was built to address.

What qualification has to include

Taken together, the evidence sets requirements for any program that claims to prepare people to review AI-supported GxP work. A course on responsible use and an acceptable-use policy do not meet them.

• Teach the failure modes before the tool.Fluent output that omits a record, confident claims with no source, a true statement about the wrong product, and a one-sided case for a defensible option. Teach the reason behind each requirement, so a reviewer can see when output satisfies the wording and misses the purpose.
• Qualify by role, system and task.Qualification attaches to a person, a configured system, a version and a defined task, the way operator qualification attaches to a process and its equipment.
• Assess on representative work.Real-format cases with seeded defects and correct outputs mixed in, so that blanket rejection fails as surely as blanket acceptance. Pass criteria are fixed in advance, and missing a critical omission cannot be offset by easy items.
• Practice independent judgment.For decisions with no single right answer, the reviewer records their own position before seeing the AI's and states the strongest case for the option the AI did not choose. The Anthropic results suggest it has to be practiced and assessed directly, because subject knowledge alone did not appear to prevent deferral.
• Design the process to require it.Independent findings before AI output, defined minimum review times, and a system log of who proposed each judgment call and which alternative was recorded, so deferral can be trended across investigations, sites and model versions.
• Requalify on change and on evidence.A model update, a new retrieval source or new authority changes the job. So does an error found later in a record.

The Life Sciences AI Academy grew out of this same concern: AI governance cannot be effective without a workforce capable of recognizing when its controls fail. Its current curriculum addresses AI failure modes, role-based learning, human oversight, documentation, and competency assessment. The next step is more demanding: translating those foundations into system-specific qualification, independently evaluated performance assessments, and requalification when AI capabilities or operating conditions change. That distinction matters. Training prepares people to perform the work; operational qualification must establish that they can perform it reliably under the conditions in which the system will actually be used.

See the curriculum, the credential standards and the methodology.

What this evidence does not show

The research on training itself is sobering. In a field experiment with 758 BCG consultants, participants given instructional material on how to use an AI tool did worse on a task the AI got wrong than colleagues who used the tool without it. From a baseline of 84.5 percent correct without AI, correct answers fell 24.5 percentage points in the trained group and 13.9 points in the untrained AI group, a difference the authors report as significant at the 10 percent level (Dell'Acqua et al., 2026). A widely cited review of automation research concludes that automation bias "occurs in both naive and expert participants, cannot be prevented by training or instructions, and can affect decision making in individuals as well as in teams" (Parasuraman and Manzey, 2010). In a pre-registered experiment, raising awareness of deepfakes did not improve people's ability to detect them (Köbis, Doležalová, and Soraperra, 2021).

These studies tested tool-use instruction, general training and awareness messages. None tested qualification built around verification: practice on seeded defects, forming a position before seeing the AI's, and assessment on representative work. That is the kind of qualification this brief argues for.

The sources above establish that human review of AI output is the control regulators and practitioners rely on, that it can fail when the reviewer lacks competence, and that even experts defer on judgment calls without noticing. None of them tests whether a training program reduces that deferral. This brief makes no claim that the Academy's training has been shown to do so. That claim requires a separate evaluation of reviewer performance on representative work.

Evidence table