The form you sign is not the system you meet#

Picture a clinic appointment. A receptionist hands you a clipboard, you tick a box, and a doctor talks you through a procedure. For decades, that small ritual has carried a big moral weight. It says: you were told the risks, you understood them, and you chose.

Now add an algorithm. Perhaps it drafts your clinical notes while you talk. Perhaps it flags a shadow on your scan, or it sits inside a trial you have just agreed to join. The clipboard looks the same. What you are agreeing to has changed completely, and nobody may have said so.

That gap is the subject of a fast-growing body of research on informed consent AI, and the findings so far are uncomfortable. A 2026 audit of consent documents from 114 AI clinical trials found that 58% did not disclose the type of AI or what it was for. Meanwhile, a national US survey suggests the public mostly wants to be asked or at least told. This post goes through what the evidence shows, where it is thin, and what a better consent process might look like.

Informed consent is the principle that a patient should understand the risks, benefits and alternatives of a decision before making it. A 2026 systematic review calls it a cornerstone of medical ethics, and it is also a legal requirement in many countries.

Three things make AI awkward for that principle. The first is opacity. Many AI systems, especially those built on machine learning (software that learns patterns from large datasets rather than following hand-written rules), cannot easily explain why they reached a particular output. Researchers describe the algorithmic complexity, opacity and adaptive nature of AI systems as a direct challenge to traditional consent frameworks. "Adaptive" matters here: some systems change as they receive new data, so the thing you consented to in March may behave differently by September.

The second is scale of data. Consent for a knee operation concerns one procedure. Consent involving AI often also concerns what happens to your records afterwards, including whether they may be used to train or refine a model. The third is invisibility. A surgeon is in the room. An algorithm running quietly in the background of your electronic record is not.

None of this is brand new. The American Medical Association's ethics journal was already flagging consent as a core issue for medical AI in 2019, noting that the question of how to inform patients properly was still open. Seven years on, the question has become urgent because the technology is no longer hypothetical.

A quick note on terms. Readability scores such as SMOG (Simple Measure of Gobbledygook) estimate the years of schooling needed to follow a text. A score under about 13 is generally considered readable for a broad audience. Flesch-Kincaid Grade Level works similarly, and for patient materials the usual target is around eighth grade. I use these scores below because consent is useless if the words are out of reach.

The most direct evidence comes from a cross-sectional analysis, published in the Journal of Medical Internet Research in 2026, of consent documents from 114 AI-involving trials on ClinicalTrials.gov. The authors read each document and scored it against World Health Organization and US National Institutes of Health standards.

The headline numbers are stark. More than half of the forms (58%) failed to say what type of AI was involved or what it was meant to do, and 18.4% left out risks altogether. Only 14% were both short (under 15,000 characters) and readable (SMOG below 13). Just 11.4% used any visual aid. The forms for higher-risk trials were not easier to read than the rest.

Data handling was patchy too. When the forms addressed what happens to a participant's data after withdrawal, the answers varied: 44.7% said nothing, 26.3% promised destruction, 25.4% allowed continued use, and only 3.5% offered the participant a choice.

Gaps in consent forms from 114 AI clinical trials Bar chart: 58% did not disclose AI type or use; 18.4% omitted risks; 14% were both brief and readable; 11.4% included visual aids; 3.5% offered a data choice on withdrawal. What AI trial consent forms leave out (n = 114) No AI type or use stated 58% Risks omitted entirely 18.4% Brief and readable (met both) 14% Included visual aids 11.4% Offered data choice on withdrawal 3.5% Orange = shortfalls; green = good practice rates. Source: Su et al., J Med Internet Res 2026.

There is a caveat worth stating. The study looked only at publicly posted consent documents, and the authors note that final on-site versions may differ. Registry entries can lag behind what a hospital actually hands to a participant. Even so, a 58% gap is hard to explain away.

What do patients know, and what do they want to be told?#

Understanding is the other half of consent, and here the evidence is smaller and more local. A 2026 survey of 55 patients at an NHS orthopaedic fracture clinic found that 72.7% were aware of AI in healthcare in general, but only 10.9% said they had a strong understanding of ambient AI, the kind that listens to consultations. Patients expressed moderate trust in their clinicians and raised concerns about transparency, privacy, consent and accuracy. The sample is tiny and drawn from a single clinic, so treat it as a signal rather than a national picture.

For a bigger picture, the University of Michigan's TIERRA team ran an online survey of 3,000 US adults in November 2025, weighted to national demographics, with a margin of error of 3.2 percentage points. Each respondent answered one of three question sets, so the base for each result is smaller than 3,000. The survey is a university research news release rather than a peer-reviewed paper, which is worth bearing in mind.

Use of AIWanted explicit permission each timeHappy with notification onlyNeither
Clinical decisions (diagnosis, treatment choice)54%32%14%
Clinical note-taking and after-visit summaries46%38%16%
Administrative tasks (billing, scheduling)45%39%16%

Source: University of Michigan Medical School, TIERRA, January 2026.

Two things stand out. Only a small minority (14% to 16%) wanted neither permission nor notice, so silence is the least popular option. And even for back-office tasks, about 45% wanted to be asked. People are not drawing the line where hospital policy might.

The ambient scribe test case#

If you want to watch this debate play out in real time, look at AI scribes. These are tools that listen to a consultation, transcribe it and draft a clinical note, typically using speech recognition and a large language model (an AI trained on huge amounts of text to predict and generate language). They are spreading quickly because they promise to give clinicians their evenings back.

The consent question is surprisingly nuanced. In April 2026, NHS England issued guidance that, according to trade press coverage, says explicit consent is not required for using ambient scribes in individual care. Clinicians must instead tell patients at the start of each interaction that the tool is in use, and patients must be able to object: if someone dissents, the tool cannot be used. I could not retrieve the guidance document itself, so this summary rests on that secondary report and is worth checking against the original.

That approach (transparency plus a right to say no) sits closer to the "notified" column of the Michigan table than the "permission" column. Whether it satisfies a public that mostly wants to be asked is an open question.

There is a deeper problem, laid out in a 2026 Personal View in The Lancet Psychiatry. Roth and Ayers argue that an AI scribe actually bundles three different operations00201-4): generating a clinical narrative, generating a clinical observation such as a mental state examination, and generating diagnostic reasoning. Each fails in a different way and demands a different level of oversight. A single "we use a note-taking tool" notice cannot honestly describe all three. The authors say this has implications for liability, consent and regulation, and they call for frameworks that reflect those differences.

Here is the twist. The same technology that complicates consent might also repair it. A 2026 systematic review in JMIR AI pooled 33 studies from 2020 to 2025 on AI in the consent process, covering patient education (18 studies), AI-written consent documents (10) and AI-assisted consent conversations (5).

The results are encouraging but need careful reading. Large language models produced accurate content, though their readability still missed targets: the best model in the review scored a Flesch-Kincaid Grade Level of 10.59, above the recommended eighth-grade level. AI-generated documents improved reading ease scores by 44% to 122% and cut the required grade level by 10% to 47%. Among the randomised trials of AI-assisted consent, patients showed better comprehension of procedural risks than with physician-led consent alone, and some reported lower anxiety and higher satisfaction. The authors still conclude that further research is needed on ethical concerns.

A single randomised trial in interventional radiology gives a flavour. In it, 122 participants with no prior exposure to the procedures read either a ChatGPT-4o-generated or a standard consent form. The AI group scored 82.9% on a comprehension test versus 77.3%, with the biggest relative gain among people with lower educational attainment. Attitudes did not differ. Note that these were volunteers, not patients facing a real decision.

Not every result is rosy. A small in silico evaluation of AI-drafted consent forms in oral and maxillofacial surgery found acceptable readability (grade 6.0 to 8.0) but weaker scores for ethics, safety and transparency than for general quality. Only four drafts were assessed, so it is a pilot at best. Taken together, these studies support a cautious position: AI can help explain things, but a clinician still has to check what is said and make sure the hard parts (uncertainty, alternatives, data use) are not smoothed away.

do patients really know what they agree to when ai joins their care 2

Regulation is catching up. The EU AI Act's Article 50 transparency duties, which a reproduction of the legislation says apply from 2 August 2026, require providers to design systems that interact directly with people so that those people are told they are dealing with AI, unless that is obvious. Information must be given clearly, at the latest at the first interaction. Article 50 is a general transparency rule rather than a clinical consent standard, so it sets a floor, not a ceiling. Medical devices and high-risk systems face separate obligations that I have not covered here.

That leaves a design question. The traditional model treats consent as a single event: sign once, proceed. A 2026 bioethics paper titled Informed Consent in the Age of Clinical AI: Beyond the One-Time Authorization Model argues, judging by its title, for something more continuous. I was unable to access the full text, so I cannot summarise its specific proposals. The idea fits the evidence above, though. If systems adapt over time, and if patients' preferences differ by use case, a one-off signature is a poor fit.

Based on the studies cited here, a practical checklist for hospitals and developers would include:

  • Say what the AI is and what it does, in plain language, since most trial forms currently do not.
  • Offer a real opt-out that works, as the NHS guidance requires for scribes.
  • Tell people what happens to their data, including after they withdraw.
  • Test forms for readability and comprehension, not only legal completeness.
  • Treat different AI functions differently rather than using one blanket notice.

Frequently asked questions#

Do doctors have to tell me when they use AI? It depends on the country, the tool and the setting. In England, current NHS guidance reported in April 2026 says clinicians must tell patients at the start of each interaction when an ambient scribe is used, even though explicit consent is not required (Digital Health). Rules for other AI uses vary and are still developing.

Can I refuse to have an AI scribe in my appointment? Under the NHS England guidance as reported, yes: if a patient dissents, the tool cannot be used (Digital Health). Elsewhere, ask your clinic what its policy is.

Do most people want to be asked before AI is used in their care? In a 2025 US survey of 3,000 adults, 54% wanted explicit permission each time for clinical uses such as diagnosis, and 46% for note-taking (University of Michigan TIERRA). Only 14% to 16% wanted neither permission nor notice.

Do consent forms for AI trials explain the AI properly? Often not. An audit of 114 trials found 58% of consent documents did not state the type of AI or its intended use, and 18.4% omitted risks entirely (Su et al., 2026).

Can AI write better consent forms than humans? Sometimes. In one randomised study, readers of a ChatGPT-generated form scored 82.9% on comprehension against 77.3% for a standard form (Koç et al., 2025), but a pilot evaluation found weaker coverage of ethics, safety and transparency (Jain et al., 2026). Human review remains essential.

What happens to my data if I withdraw from an AI study? It varies widely. In the audited trials, 44.7% of consent forms said nothing, 26.3% specified data destruction, 25.4% allowed continued use, and just 3.5% let participants choose (Su et al., 2026).

Does the EU AI Act require hospitals to tell patients about AI? Article 50 requires systems that interact directly with people to be designed so those people are informed they are dealing with AI, unless this is obvious, and applies from 2 August 2026 (text of Article 50). It is a general rule, so specific clinical obligations sit elsewhere.

References#

  1. Su H, Xiao F, Chau H, et al. Informed Consent Disclosures and Minimum Requirements in AI Clinical Trials: Cross-Sectional Analysis. J Med Internet Res. 2026;28:e94504. https://doi.org/10.2196/94504 (peer reviewed)
  2. Dababneh S, Hébert N, Meloche L, et al. Enhancing Patients' Informed Consent Through AI: Systematic Review. JMIR AI. 2026;5:e93501. https://doi.org/10.2196/93501 (peer reviewed)
  3. Kathiramalai S, Davis T, Schaller G. Patient Attitudes Towards Ambient Artificial Intelligence in Clinical Consultations. Cureus. 2026;18(9):e115822. https://doi.org/10.7759/cureus.115822 (peer reviewed)
  4. Roth AS, Ayers NB. AI scribe functions in psychiatric practice: clinical oversight, consent, and regulation. Lancet Psychiatry. 2026;13(9):804-810. https://doi.org/10.1016/S2215-0366(26)00201-400201-4) (peer reviewed)
  5. Koç U, Güneş YC, Çolakoğlu MN, et al. ChatGPT-generated informed consent forms in interventional radiology: a randomized controlled evaluation of comprehension, attitudinal responses, and readability. Eur J Radiol. 2025;195:112624. https://doi.org/10.1016/j.ejrad.2025.112624 (peer reviewed)
  6. Jain A, Deshmukh R, Bhutekar U. Evaluation of an Artificial Intelligence Language Model for Generating Informed Consent Forms in Oral and Maxillofacial Surgery: A QUEST Framework Analysis. J Maxillofac Oral Surg. 2026;25(4):1118-1126. https://doi.org/10.1007/s12663-026-03085-7 (peer reviewed)
  7. University of Michigan Medical School, TIERRA. AI in Healthcare: Notification and Consent Preferences Based on U.S. National Survey. Research news, 5 January 2026. Link (institutional report; not peer reviewed)
  8. NHS England guidance on ambient scribes, as reported by Digital Health, 2 April 2026. Link (secondary report; the primary guidance should be consulted)
  9. Regulation (EU) 2024/1689 (AI Act), Article 50, via the Artificial Intelligence Act website. Link (reproduction of the legal text)
  10. American Medical Association. What to tell patients when artificial intelligence is part of the care team (summary of AMA Journal of Ethics, February 2019). Link
  11. Informed Consent in the Age of Clinical AI: Beyond the One-Time Authorization Model. 2026. DOI link (cited by title only; full text not accessed)

You may like these posts

See all