What it does, and what it will not do

MHABot turns structured observations into a mental state examination. That is the whole of it. The restraint is the point.

How the note is built

You work through the domains of the examination, selecting what you found. Each selection is attached to a fixed phrase, and those phrases are assembled into prose by rules that do not change from one note to the next.

Nothing is inferred. Nothing is weighted. There is no model deciding what you probably meant. Every sentence in the output traces back to something you selected, and if you select nothing, nothing is written.

What the tool does add is grammar: joining clauses, agreeing verbs, collapsing three short sentences into one that reads properly. It changes the punctuation, never the findings.

The three things it refuses to do

It does not diagnose

No diagnosis, no differential, no suggestion of what the presentation might be. Software that proposes a diagnosis is, under the UK Medical Devices Regulations, very likely a medical device, with all that follows. More to the point, a documentation tool that reached a clinical conclusion would be doing the one part of the work that is unambiguously yours.

It does not score risk

No number, no traffic light, no low, medium or high. NICE guideline NG225 advises against using scales to predict suicide or repetition of self-harm, and against global risk stratification to decide who is treated or discharged.

The evidence behind that recommendation is worth knowing. Work from the National Confidential Inquiry into Suicide and Safety in Mental Health found that the large majority of patients who went on to die by suicide had been judged low risk or no immediate risk at their final contact with services. MHABot supports a narrative risk formulation instead, which is what current guidance asks for.

It does not recommend a section

Detention under the Mental Health Act is a statutory act reserved to specified professionals — an Approved Mental Health Professional and the required medical recommendations — following a face-to-face assessment. No software has standing to make that recommendation, and any that offers one has misunderstood what it is for.

Where the note goes

Nowhere. The note is assembled by code running in your browser. The PDF and the Word file are generated there too. Our database contains no field capable of holding an assessment, a patient identifier, or any clinical text.

This is not a policy we could quietly change. It is how the thing is built: there is no endpoint that receives a note, because none was written.

The trade-off is honest, and you should know it before relying on this. Notes live in your browser's storage. Clearing your browsing data deletes them and we cannot recover them. Export anything you need to keep.

There is one exception, and only if you turn it on: emailing a note to your own address sends the text through the server so it can be delivered. It is held in memory for the moment it takes to send and written to no log or table. If you never press that button, no clinical text leaves your device.

No patient identifiers

The intake asks for an age band, an area, and the setting. No name, no date of birth, no NHS number, no address. You should not add any, and the tool gives you no field that invites it.

This is partly data protection and partly clinical: an assessment written without identifiers is one you can paste into a record system, discuss in supervision, or use for teaching without further thought.

The consistency checks

As you work, your entries are compared against a fixed set of rules, and where two of them sit oddly together the tool says so. Elevated mood recorded alongside flat affect. A person recorded as mute whose tone of voice has also been described.

Each flag shows the rule it applied, so you can judge it rather than take it on trust. They are prompts to look again, never clinical opinions, and many flagged combinations are entirely real findings that simply deserve a deliberate sentence. An absence of flags means nothing about whether your findings are correct.

Who it is for

Qualified clinicians — psychiatrists, doctors on psychiatry rotations, mental health nurses, psychologists, approved mental health professionals — and students and trainees working under supervision.

It is not for members of the public and not for self-assessment. The options assume you know what the terms mean, and the tool will happily produce a fluent note from findings that are clinically wrong. It writes; it does not check your clinical judgement.

On the word "AI"

Generation is deterministic and template-based. The same selections always produce the same note. There is no model deciding what to write, which is precisely why the output cannot invent a symptom you did not record — the failure mode that makes clinicians wary of ambient scribes.

If a language model is ever used to smooth the prose, it will be optional, off by default, constrained to rephrasing without adding facts, and validated against your input before anything is shown. That line is not going to move.