AI can write the prose of a psychological report. It cannot write the report. The difference is that a report is a professional opinion attributed to a named person, and the reasoning that produces that opinion is the part software has no business generating.
I build one of these tools, so treat this as an interested party drawing a line rather than a neutral survey. But it's the line every psychologist I've spoken to draws first, and it's the one I'd want a buyer to understand before spending anything.
It will produce the sentences. Whose reasoning is in them is a different question.
What can it actually do?
Three things, reliably.
Assemble structure. Sections, headings, the order your documents follow. Genuinely useful and completely uncontroversial.
Turn supplied findings into sentences. You provide results, history, observations. It produces the descriptive prose that carries them. This is the bulk of a long document and it carries no judgment.
Hold consistency across length. Keeping terminology and phrasing steady over twenty pages is tedious for a person and easy for software.
Put together, that's a meaningful share of a write-up psychologists describe as ten to fifteen hours. Not the whole thing — but the part that's least interesting to do and most repetitive to redo.
What can't it do?
The part your name is attached to.
It can't decide what an uneven profile indicates. It can't weigh a discrepancy between two sources and say which carries more evidential weight. It can't judge whether an observation changes how a result should be read. Those are the acts that make a document a professional opinion, and every psychologist I've asked has said the same thing about them: it's what they trained for and they will not delegate it.
There's a practical reason beyond the principle. A tool that produces conclusions hands you something to verify rather than something you wrote — and verifying a conclusion you didn't reach is usually slower than reaching it, and always riskier with your signature underneath.
What about the parts in between?
This is where it gets genuinely interesting, and where I'd be sceptical of anyone claiming a clean answer.
Some sections are neither pure description nor pure judgment. A summary selects what matters. A recommendation follows from an interpretation. Software can draft the shape of these, but the selection embedded in them is a clinical act.
Our answer is to scaffold rather than author: lay out what needs addressing, leave the substance to the clinician. It's a narrower promise than the category usually makes and it's the one we can actually keep. Where the line falls in practice is in why the clinician stays the author.
Some sections are neither pure description nor pure judgment. That's where care is required.
How would I test the claim?
Take a case with an uneven profile — the interesting kind, not a tidy one — and give the tool everything except one piece of information a report would normally contain.
Then look for two things. Did it produce a confident sentence covering the gap, or did it flag the absence? And can you trace each claim in the output back to something you actually supplied?
A tool that invents to fill a hole has told you it will do that on a real file, where you might not notice. That test takes twenty minutes and it's more informative than any demo. The full protocol, without touching a real client file, is in trialling an AI report tool without risking a client file.
So what's the honest answer?
AI can take the typing. It cannot take the thinking, and you should be wary of anyone selling you the second thing — including me, if I ever start.
If the typing is what's eating your evenings, that's a real problem worth solving and it's what ours is built for. If your bottleneck is somewhere else, no amount of drafting will touch it, and you're better off knowing that before you buy.
— Ian
