Which kind of AI suits the job?
Six questions about your situation — not about how models are built. You will leave with what your own answers rule out and why, what is still worth considering, and a prompt you can put to whatever system you are thinking of using.
No account, and your answers are never stored — they are worked out and handed straight back, never written to a database or a log. Being exact, because this tool argues you should check claims rather than accept them: your answers stay in this browser tab until you clear them; and this site counts page visits like any other, which sets a cookie and records the page, the time, your browser’s user-agent string, an identifier derived from your address, and where you came from. Neither of those touches what you typed here.
Your result also opens two further paths — a question sheet for a supplier you have in mind, and a check of what is already switched on in software you run — and both are worth more once you have seen what your own answers rule out.
What your answers mean
Ruled out, and why
Each of these follows from something you said, and the answer it followed from is shown so you can check the working. Some of them rest on an assumption you may be entitled to reject — that you would be building something from scratch, or that you need a polished result rather than a rough one. Where that is so, the reason says which assumption it is making. If it does not hold for you, the exclusion does not either. This is a set of arguments, not a verdict.
Still worth considering
A shortlist, not a recommendation. Nothing here is ranked, and more than one may suit.
Put this to the system you are considering
Copy this and paste it into whatever you are thinking of using. Run it against two or three and compare what comes back — that comparison is the useful part, not any single answer.
Have you a particular product in mind?
Then the useful output is not our opinion of it. Kinds of system can be ruled out, because that is reasoning about architecture. A named product can only be asked, because nobody has the facts without asking — including us. So this gives you the questions, and what a deflecting answer looks like.
You will mostly be grading answers they have already published — a trust page, a data processing agreement, the documentation. Where the questions get sharp, it is because your own answers made them sharp.
This stays in your browser. It is not sent anywhere — the questions come from your six answers, so the name is not needed to work them out.
These questions cite New Zealand and European provisions so you can see where each one comes from. That is background for asking, not legal advice, and we are not lawyers. If an answer matters to a decision you are making, take it to someone who is.
What this gives you is a record of what you asked and what they answered. It is not a security assessment, and a small organisation cannot realistically make one — the Privacy Commissioner has said as much. A written undertaking is a remedy and a paper trail; it is not itself a safeguard. Both of those are worth having, and neither is the same as knowing the product is safe.
A note on your own reading: once a set of questions circulates, suppliers learn to answer it well, which is not the same as being well built. The patterns below each question tell you what a deflection looks like — they grade the shape of an answer, not its substance. An answer in good form can still be one you need to check.
Before you email anyone: ask the paperwork first
Most of these questions have already been answered somewhere the supplier has published — the terms, the privacy policy, the data processing agreement, the sub-processor list. Gather what you can find, then paste this and those documents into whatever AI assistant you already use. It will not judge the supplier and is not asked to. It finds the passage that answers each question and quotes it, so you can check the words are really there.
It is set up around a known weakness: this kind of reading misses far more than it invents. So it must quote everything it claims to have found, and where it finds nothing it has to hand you the terms to search for yourself — because a blank is usually its failure rather than the document's silence.
Why we built this, and who we are
This tool names no product and recommends none. It is published by My Digital Sovereignty Ltd, a New Zealand company that also builds and sells AI systems — which is a reason to check its reasoning rather than take it on faith. Every deduction above shows the answer it came from, so you can. It is also why the rules are written to be able to rule us out: across every combination of answers, the kind of system we sell is excluded about half the time — measured at 51.1% across all 6,400 of them, and asserted in the test suite as a band that fails if it drifts out of 40–60%, in either direction.
Our values are published in full, on both sites, and they are the same values: the Tractatus Framework values and My Digital Sovereignty’s values.
Questions & answers about this tool — including the obvious one: you build and sell AI systems, isn’t this just a way of selling them? Also what happens if you want to make sure AI is not used, and why some of the reasons above rest on assumptions you may be entitled to reject.
We put this tool's own document prompt to our own published terms and printed the result — five of seven questions answered, two not.
If the question is the other way round — how to establish that a machine was not used, or what a disclosure should say — that is The Marks It Leaves, with its own questions and answers.
New to this? Understanding AI: a reading guide assumes no prior knowledge. The reasoning behind this tool is set out in The Least Performant AI.
This is not a new instrument so much as an old one made usable. The question the prompt ends on — name one thing this system will refuse to do — is the test published in A value that never says no. And the reason nothing here is scored is the rule set out in Many rooms, one map: a single number is the move that work spends its length refusing. We hold ourselves to it too — don’t trust us, check the working.