The protocol is forty-one pages of dense clinical prose for a first-in-human Phase 1 study: a healthy-volunteer single-ascending-dose and multiple-ascending-dose design for a novel small molecule, with sentinel dosing, a 3-plus-3-style escalation logic, intensive pharmacokinetic sampling, cardiac telemetry, and a stopping-rules appendix written by a clinical pharmacologist who has never had to explain it to a nineteen-year-old volunteer. The medical writer pastes the synopsis, the schedule of assessments, and the risk sections into the enterprise large language model with one instruction: produce an informed consent form at an 8th-grade reading level. Nine seconds later a clean, warm, readable draft appears. It scores a Flesch-Kincaid grade level of 7.8. It reads beautifully. It is also missing the element of consent that tells the volunteer whom to call if they are injured, and it has quietly softened the description of the dose-escalation design into a sentence that no longer conveys that this is the first time the drug has ever been given to a human being. Both of those are not style problems. One is a missing 21 CFR 50.25 required element, and the other is a comprehension failure that an Institutional Review Board (IRB) will read as a failure to disclose the true nature of the research. This lesson is about the discipline that turns a readable AI draft into a consent form that an IRB approves and a volunteer actually understands.

Why the ICF Is the Hardest Easy Document in the Dossier

The informed consent form looks like the simplest document a medical writer touches. It is short, it uses plain words, and it has no tables of efficacy results to reconcile. That surface simplicity is exactly why it is dangerous to hand to an AI without a control structure. The ICF carries two obligations that pull in opposite directions, and the tension between them is the entire craft. The first obligation is regulatory completeness: under 21 CFR 50.25, the form must contain a specific set of basic elements and, where applicable, a set of additional elements, and the absence of any required element is a deficiency regardless of how well the rest of the document reads. The second obligation is genuine comprehension: the form must communicate the research at a level the prospective subject can understand, which for a general adult population means roughly an 8th-grade reading level, and a document that is technically complete but incomprehensible has failed the ethical purpose of consent even if it passes a checklist.

A large language model is extraordinarily good at the comprehension surface and structurally blind to the completeness floor. It will lower the reading level on command, because simplifying prose is precisely the pattern it learned from millions of examples of plain-language rewriting. It will not, on its own, guarantee that all eight basic elements and the applicable additional elements of 50.25 are present, because the model is completing the pattern of a consent form, and consent forms in its training corpus vary in which elements they spell out explicitly versus assume. When the model drops the injury-compensation element, it does not experience an omission. It produces a fluent, finished-looking document, and the very fluency is what makes the gap invisible to a reviewer who reads for tone rather than for elements.

This is the central reframing for the ICF task. The writer is not asking the model to write a consent form. The writer is asking the model to perform a plain-language translation of clinical content, and then the writer is separately responsible for proving that the translated document satisfies a fixed regulatory element set. Translation and compliance are two different jobs, and the AI is good at one and silent on the other. The named author owns the gap between a readable document and a complete one.

The 21 CFR 50.25 Element Set the Model Cannot See

The eight basic elements of informed consent under 21 CFR 50.25(a) are the non-negotiable spine of the document, and every writer using AI for an ICF should be able to recite them as a checklist rather than trust the model to include them. They are: a statement that the study involves research, with an explanation of the purposes, the expected duration, and a description of the procedures, identifying any that are experimental; a description of reasonably foreseeable risks or discomforts; a description of any benefits reasonably expected; a disclosure of appropriate alternative procedures or treatments; a statement describing the extent to which confidentiality of records will be maintained; for research involving more than minimal risk, an explanation of whether compensation and medical treatments are available if injury occurs, and where to obtain further information; an explanation of whom to contact for questions about the research and about research-subject rights, and whom to contact in the event of a research-related injury; and a statement that participation is voluntary, that refusal involves no penalty or loss of benefits, and that the subject may discontinue at any time.

The additional elements under 50.25(b) apply when relevant: a statement that the treatment or procedure may involve unforeseeable risks; anticipated circumstances under which the subject's participation may be terminated by the investigator; any additional costs to the subject; the consequences of a subject's decision to withdraw and procedures for orderly termination; a statement that significant new findings will be provided to the subject; and the approximate number of subjects involved in the study. For a first-in-human Phase 1 study, several of these are not optional in spirit even if they are labeled additional: the unforeseeable-risks statement is the honest core of a first-in-human design, and the approximate-number-of-subjects element matters because cohort size is part of how a volunteer judges the maturity of the safety data.

The model cannot see this element set as a requirement. It can reproduce the shape of a consent form, and the shape usually includes most of these, but "usually" is the failure word. A first-in-human ICF that omits the research-injury contact, or that buries the experimental-nature statement inside a paragraph about procedures so that it no longer reads as a clear disclosure that this is research, will be returned by the IRB. The control that prevents this is not better prompting. It is an explicit 50.25 element checklist applied to the draft as a separate verification pass, element by element, against the actual text.

The Reading-Level Trap: The Number Is Not the Comprehension

Readability formulas are the most over-trusted tools in plain-language work, and the ICF task is where their limits bite hardest. The Flesch-Kincaid Grade Level and the SMOG (Simple Measure of Gobbledygook) index both estimate reading difficulty from surface features: Flesch-Kincaid weights average sentence length and average syllables per word, while SMOG counts polysyllabic words across a sample of sentences and is often preferred for health materials because it targets near-complete comprehension rather than partial. Both produce a single grade-level number, and both can be satisfied by a document that no human actually understands.

The trap is mechanical. A model asked to hit an 8th-grade level will shorten sentences and swap long words for short ones, and the score will fall. But "pharmacokinetic sampling" can become "blood tests to measure the drug" only if the model understands what the phrase means; if it instead shortens "pharmacokinetic" to "PK" the syllable count drops and the score improves while comprehension collapses. Worse, a formula rewards a document that is simple at the word level but conceptually opaque: a sentence like "You will be in the SAD or MAD cohort" is short, low-syllable, and scores beautifully, and is meaningless to a volunteer who has never heard of single-ascending-dose or multiple-ascending-dose designs. The number went down. The understanding did not go up. The formula cannot tell the difference, because it measures the surface of the prose and never the meaning.

So the reading-level tool is necessary and insufficient. The writer runs Flesch-Kincaid and SMOG on the draft, and a score above the target is a hard fail that sends the document back. But a passing score is not a pass; it is a precondition. The real comprehension test is whether a reader outside the field can restate the purpose, the principal risks, the voluntary nature, and the injury pathway after reading the form once. Many IRBs and many sponsors now require a teach-back or readability-panel step for exactly this reason, and the AI draft should be treated as the input to that human comprehension check, never as its output.

Translating the First-in-Human Design Without Softening the Risk

The hardest single passage in a Phase 1 healthy-volunteer ICF is the description of the dose-escalation design, and it is the passage where AI plain-language translation most reliably goes wrong in a way that matters. The clinical reality is stark: this is the first time the molecule has been given to a human, doses start very low and rise cohort by cohort, sentinel subjects are dosed first and observed before the rest of a cohort proceeds, and the entire architecture exists because the risks are genuinely unknown. An honest ICF must convey that unknown-ness without either terrifying the volunteer or, far more commonly in AI drafts, sanding it down into reassurance.

When a model simplifies "a first-in-human single-ascending-dose study with sentinel dosing and predefined stopping rules" it tends to produce something like "doctors will give you a carefully chosen dose and watch you closely." Every word of that is technically true and the net effect is a lie of omission, because it has erased the two facts the volunteer most needs: that no human has received this drug before, and that the dose they receive is part of a deliberate escalation in which earlier signals of harm change what later subjects get. The softening is not malice; it is the model's learned tendency to render clinical caution as bedside reassurance, because reassuring phrasing is statistically associated with patient-facing prose in its training data. The fix is a constraint at the source: the writer specifies the non-negotiable facts the passage must carry, in plain words, and verifies that the translation preserves them rather than smoothing them away.

This is where the writer's judgment is irreplaceable and where the AI is a drafting accelerant rather than an author. The model can produce five plain-language phrasings of "you are one of the first people ever to receive this drug" in seconds, which is genuinely useful, but only the writer knows that the phrase must survive into the final form intact, and only the writer can confirm that the surrounding simplification has not quietly relocated the unforeseeable-risk disclosure required under 50.25(b) into a footnote no one reads.

The Named Artifact: The IRB Submission Version of the ICF

The artifact this lesson produces is concrete and named: the IRB submission version of the informed consent form, the document that goes into the IRB package alongside the protocol, the Investigator's Brochure, and the recruitment materials for review at a convened meeting or by expedited procedure. Naming the artifact matters because it fixes the audience and the failure consequences. The audience is an IRB, which in the United States reviews against 21 CFR 50 and 21 CFR 56, and increasingly against the institution's own plain-language and health-literacy standards. The consequence of a deficient ICF is not a stylistic note; it is a deferral or a list of required modifications that delays study start-up, and in a competitive first-in-human program every week of IRB cycle time has a real cost.

IRB expectations have hardened in exactly the places where AI drafts are weakest. Reviewers look for the experimental-nature statement to be unambiguous and early, not embedded; they look for the risk section to be specific to this molecule and this design rather than generic; they look for the research-injury contact and compensation language to be present and accurate to the sponsor's actual policy, because an ICF that promises compensation the sponsor will not provide is worse than one that is silent; and they look for the voluntary-participation and withdrawal language to be clean and unconditional. A model can produce plausible boilerplate for every one of these, and plausible boilerplate is precisely the risk: the injury-compensation paragraph the model generates may describe a generic compensation scheme that contradicts the sponsor's specific clinical-trial-agreement and insurance arrangements, which is a factual error wearing the costume of standard language.

This is why the named artifact must be reconciled against named sources: the protocol for the design and risk facts, the sponsor's clinical-trial agreement and insurance documentation for the injury and compensation language, the contact roster for the three required contact pathways, and the institution's template requirements for local additions. The AI draft is reconciled to those sources the same way a Module 2.5 efficacy claim is reconciled to a table. A consent statement that cannot be traced to a real sponsor policy is treated as wrong until proven right.

Building the Verification Pass: Checklist, Then Readability, Then Comprehension

The defensible workflow for an AI-assisted ICF has three verification layers applied in a deliberate order, and the order is not arbitrary. The first layer is the 21 CFR 50.25 element checklist, run before anyone judges the prose, because an elegantly written document missing a required element is a non-starter and there is no point polishing it. The writer walks each of the eight basic elements and each applicable additional element against the actual draft text, marking each as present, absent, or present-but-inadequate, and an absent or inadequate element blocks progress until it is fixed at the source. This is a content-completeness gate, and it is the layer the model is most likely to fail silently.

The second layer is readability measurement: Flesch-Kincaid and SMOG run on the full body text, with a target at or below the 8th-grade level and a hard fail above it. This is cheap, fast, and objective, and it is also the layer most likely to give false comfort, so it is explicitly subordinate to the third. The third layer is comprehension verification, which the formulas cannot perform: a human reader from outside the clinical team, ideally representative of the target volunteer population, reads the form once and restates the purpose, the principal risks including the first-in-human nature, the voluntary character, and the injury pathway. Failure at the comprehension layer sends specific passages back even when the readability score passed, because the score measured the surface and the teach-back measured the meaning.

Around these three layers sits the audit trail, and for an ICF it is GxP-grade. The record captures the model and version, the system prompt and reading-level target, the protocol and policy sources loaded with their versions, the readability scores before and after revision, the completed 50.25 element checklist with dispositions, the comprehension-check outcome, and the named writer who signed the verification, all tied to the IRB submission version under document control. This is what makes the workflow survive not just the IRB but a later FDA bioresearch monitoring inspection that asks how the consent document was produced and verified. The discipline is identical to every other AI-assisted regulated artifact: load the real sources, translate, then verify against a fixed standard rather than against the draft's own confidence.

What This Means for the Writer on Monday

The ICF is the document where AI plain-language translation earns its keep most visibly and fails most quietly, and both facts argue for the same discipline. A model that turns a forty-one-page first-in-human protocol into a warm, readable 8th-grade draft in nine seconds removes hours of blank-page labor, and refusing that help is not a virtue. The virtue is in treating the draft as a translation to be verified, not a consent form to be trusted. The writer loads the protocol, the sponsor's compensation and insurance documentation, and the contact roster before drafting, because a consent statement is only as true as the source it traces to. The writer runs the 50.25 element checklist first, because completeness is a floor that fluency hides. The writer runs Flesch-Kincaid and SMOG, but treats a passing score as a precondition rather than a verdict, because the number measures the surface and not the meaning. And the writer routes the draft through a human comprehension check, because the entire ethical purpose of consent is understanding, and understanding is the one thing the formula cannot measure.

The deeper point is that the ICF makes the stakes of plain-language AI concrete in a way that an efficacy summary does not. A softened dose-escalation paragraph is not a citation error a reviewer catches at reference QC; it is a person agreeing to be the first human to receive a molecule without fully grasping that fact. The named author owns that gap absolutely. The model drafts the words; the writer guarantees that the words are complete under 50.25, readable in measure, comprehensible in fact, and true to the sponsor's actual policy. That guarantee is the work, and the nine-second draft is only where it starts.

Key Takeaways

• An ICF carries two obligations the model treats unequally: comprehension and completeness. The model excels at lowering reading level on command but is structurally blind to the 21 CFR 50.25 element floor, so a fluent draft can silently drop a required element like the research-injury contact while reading beautifully.
• The eight basic elements of 50.25(a) and the applicable additional elements of 50.25(b) are a checklist, not a model behavior. Run the element checklist against the actual draft text first, before judging prose, because an absent or inadequate required element is a non-starter no matter how polished the document.
• A passing Flesch-Kincaid or SMOG score is a precondition, not a verdict. The formulas measure sentence length and syllables, so a short, low-syllable sentence full of unexplained jargon like "SAD or MAD cohort" scores well and communicates nothing; only a human teach-back tests actual comprehension.
• The first-in-human dose-escalation passage is where AI most often softens risk into reassurance. Specify the non-negotiable facts the passage must carry, that no human has received the drug before and that the dose is part of a deliberate escalation, and verify the translation preserves them rather than smoothing them away.
• The named artifact is the IRB submission version, reconciled to real sources and captured in a GxP audit trail. Injury-compensation language must trace to the sponsor's actual clinical-trial-agreement and insurance policy, not to generic boilerplate, and the record must show the element checklist, readability scores, comprehension check, and a named signature under document control.