The check before you sign.
AI can change how someone is described without changing the obvious facts. AffectiveHalo helps you spot those changes in notes, reports and letters before you sign or send them.
Compare the AI draft with your original notes, or check the draft on its own.
No account needed · Under a minute · Your document text is not retained
Her words, someone else's.
Synthetic exampleSource
I don't feel safe going outside on my own.
AI draft
The patient is reluctant to leave the house independently.
Finding
The person's description of feeling unsafe has become a professional characterisation of reluctance.
The AI saves the time. Your name is still on it.
That's a fair trade. AI gives back hours. But whatever the draft says about a person, you're the one who signs or sends it. The question is what your review can catch.
A fluent draft reads as right. The changes that matter are quiet: a possibility becomes a fact, one event becomes a pattern, someone's own words become your description, your judgement disappears. Facts survive; the person doesn't.
The document doesn't stay with you. Others act on it, inherit it and rely on it. Complaints, courts and the person themselves can ask to see it. What it says about them is what every later reader will believe.
The check makes your review real, in under a minute.
How it works
Paste the draft.
Add the AI-generated document and, optionally, the original source. Names are replaced in your browser before anything is processed.
See what needs attention.
AffectiveHalo highlights changes in how the person is portrayed, or potentially unsupported wording when no source is available. Flags come from rule-based checks, not a generative model.
Review and decide.
Each flag shows the exact words and why it was raised. Fix, accept or dismiss. It takes under a minute.
Keep your review record.
Download the date, the flags and what you decided. We keep the findings and a hash, never your text.
Findings support your judgement; they don't replace it.
See the full methodologyTwo ways to check
With the source
Compare the AI draft against your original notes, transcript or source material. AffectiveHalo identifies changes in how the person is portrayed: omissions, shifts in certainty, changes in attribution, and wording that turns their words into your interpretation.
Without the source
Examine the AI draft on its own. AffectiveHalo highlights wording that may deserve closer review: unsupported claims, labels without observations, and characterisations with nothing behind them.
Without the source, AffectiveHalo cannot establish what AI changed. It can only identify wording that may deserve closer human review.
Small word changes. A different person on the page.
With the source
Your conclusion, deleted.
Synthetic exampleSource
Low mood reported, but on balance I don't think this meets the threshold for a referral; review in two weeks.
AI draft
Low mood reported. Review in two weeks.
Finding
The draft keeps the observation and drops your decision. A reader will re-decide without you.
We call this representation drift. You'd call it 'that's not what I wrote.'
Without the source
A label where a person should be.
Synthetic exampleAI draft
Client is non-compliant with the sleep plan and minimises the impact of her drinking.
"Non-compliant" describes a category, not what your client did or said.
"Minimises" is presented as established, with no observation to support it.
The note doesn't describe what she did or said.
Without the source, we can't say what changed. We show what the note claims about the person, and where nothing in the note supports it.
Always about the person.
The check only looks at how a person is portrayed: their words, their certainty, what is attributed to them, how they are labelled, and what is said to cause their behaviour. Everything else in the note is left alone.
Who it is for
You sign or send it.
Therapists, counsellors, psychologists, social workers, nurses, physicians, coaches, and anyone writing about a person with AI.
AI saved you the writing.
The check makes the review real.
You sign or send with your review record.
You're responsible for the notes.
Team leads, quality and safety, governance, scribe rollout owners.
See where closer review may be needed.
Export what your oversight process asks for.
Your team's text is never stored.
What we've measured so far
395 of 400 altered notes flagged, on our synthetic mental-health benchmark of 500 cases. Read the benchmark
Measured on a synthetic benchmark using the with-source check. Initial testing has used synthetic records. Real-world evaluation is a next step. These results do not establish clinical safety or real-world performance. No equivalent performance figure is currently claimed for the without-source check.
Free to start.
Free
Free
Published examples, a monthly allowance of checks on your own notes (with or without a source), the review screen and downloads.
Check a noteOrganisation
Evaluation pricing on request
Review queue, trends, evidence packs, roles.
Start a team evaluationYou already pay for the AI tool to write the draft. This is the part that makes it yours.
Your own documents are the real test, which is why the first checks are free.
Your text is not ours to keep.
Names are replaced in your browser before anything is processed. Text is held in memory for the check and your review, then discarded. We keep the findings and a hash, never your text.
How people are documented shapes how they are understood. The check makes changes in how a person is portrayed visible in anything an AI writes about them, starting in mental health and social care.
Contact
Get in touch.
For team evaluations, API interest or research collaboration.
One example a week of what AI changes about a person.
Tell us if the flag is right.