When a *chatbot triages* in healthcare, who answers for the advice?

The argument in Skåne

A group of doctors in Skåne has sent an open letter to the Swedish Medical Association, asking it to take a stronger stance against AI systems in healthcare that add to doctors' workload rather than lightening it ^[1]. They single out 1177's chatbot and plans for automated triage as examples of systems being rolled out before the profession has had a proper say.

The chatbot on 1177 Vårdguiden gives advice to people who are unwell. Automated triage would go further, sorting patients by urgency without a doctor or nurse making the call. The doctors' worry is that these systems shift responsibility onto clinicians who have to pick up the pieces when the AI gets it wrong, without giving them the tools or the time to check what the AI actually did.

It is a bit like being handed a to-do list written by someone else and being told you are accountable for everything on it, even the bits you did not write and cannot verify.

What the AI Act says about healthcare decisions

The AI Act puts some AI systems for health into its high-risk category, in Annex III ^[2]. That list includes AI for triage and prioritisation of patients in emergency care. For those systems, the requirements apply from 2 December 2027: risk management, data governance, technical documentation, automatic logging under Article 12, human oversight and a conformity assessment before the system goes live.

Chatbots that give general health advice, without deciding who gets treatment first, sit in a greyer area. If the chatbot only points people towards information or suggests they ring 1177, it is harder to say it counts as high-risk under the current wording. That does not mean it carries no risk in practice, just that the AI Act's strictest rules may not apply.

Either way, Article 26 sets out duties for anyone using a high-risk AI system ^[3]. They must use it according to the instructions, have human oversight, monitor how it performs and inform people affected by decisions. For a hospital or a region using AI to triage patients, that would mean making sure a clinician can step in, and telling patients when an AI has been involved in deciding how urgent their case is.

The evidence question: what did the chatbot say?

When a patient follows a chatbot's advice and things go wrong, the first question is what the chatbot actually said. Did it tell them to wait, to ring back or to go to A&E? The patient may have a screenshot. The organisation running the chatbot needs its own record, one it can show a supervisor, an ombudsman or a court without having to ask anyone to take its word for it.

Most chatbot platforms keep logs, but a log the operator can edit is easier to challenge. If the only proof is a database entry that could have been changed yesterday, the patient's screenshot starts to look like the stronger piece of evidence.

Article 12 of the AI Act requires high-risk systems to log events automatically, over the system's lifetime ^[4]. It does not say the logs have to be tamper-proof, but it does say they are there for traceability. When something is questioned months later, a log you can prove has not changed is more useful than one you cannot.

What a digital witness does in this case

Here, the thing to seal is each answer the chatbot gives, as it is sent to the patient: the question, the answer, the model version, the time and the notice that the patient was talking to an AI, as required by Article 50 from 2 August 2026 ^[5].

Sealing takes a fingerprint of the record, not the content, so nothing about the conversation leaves the healthcare provider's systems. The fingerprint goes to WitniumChain, which returns a witness ID. To check the record later, anyone the provider shares it with calculates the fingerprint again and compares it with the sealed one. If they match, the record has not changed since it was sealed. Change a single character and it will not match.

What it shows is that the record of what the chatbot said, and when, has not been altered. What it does not show is whether the advice was correct, whether the patient followed it or whether the organisation met its duty of care. Those are separate questions. But when a case turns on what the AI actually said, having a checkable record of your own is better than relying on the patient's screenshot.

How it would work

1. The chatbot sends its answer to the patient and, at the same time, seals a record of the exchange.
2. Months later, the patient or a supervisor asks what advice was given.
3. The organisation shares the record and the witness ID. Anyone can check that the record matches the sealed fingerprint.

You can get started with the REST API, the SDK or the MCP server. The content stays with you. The first 2,000 witnesses a month are free.

What this means for you

If you are running a chatbot that gives health advice, or planning one, ask:

• Could we show, six months from now, exactly what the chatbot said to a particular patient on a particular day?
• Does our logging cover the AI's answer, the model version that generated it and the time it was sent?
• Can we prove the log has not been edited, even by someone with administrator access?
• Have we told patients they are talking to an AI, and kept a record of that notice?

If you are a doctor or a nurse being asked to oversee an AI system, ask:

• What can I see about what the AI did before I have to make my own decision?
• If a patient deteriorates after following the AI's advice, what record will exist of that advice, and who keeps it?

For healthcare leaders, the question is whether the AI reduces the clinical workload or just shifts the accountability. The AI Act's requirements for high-risk systems are there to make sure oversight is real, not just on paper.

What to watch

• 2 August 2026: Article 50 of the AI Act applies. People interacting with an AI system must be informed that they are doing so, with certain exceptions ^[5].
• 2 December 2027: High-risk requirements under Annex III apply, including automatic logging under Article 12 for triage AI in emergency care ^[2][4].

The debate in Skåne is about workload and professional autonomy, but it is also about something older: when you delegate a decision, you need to know what was decided and be able to show it. AI does not change that. It just makes the question harder to answer if you have not thought about it early.