How we decode an AI tool

A decode answers the questions a directory listing won’t: what this thing really costs, who owns it now, where your data goes, whether it does what it claims, and whether it’ll still exist in two years. We lead with the catch because that’s the part you can’t find anywhere else — not because every tool is a scam. Several of the products we decode are genuinely good, and we say so.

We are not a certification body and we do not issue a seal. A seal is worth exactly the reputation of whoever issues it, and we have not earned that. What we publish is sourced analysis you can check yourself, which is a different and more honest thing.

Primary documents, first and preferably only: the product’s own privacy policy and terms of service, its pricing and App Store listings, regulatory orders, court filings, and company or acquirer announcements. Where we cite research we link the study and say who ran it.

Every decode carries a dated source list, and each entry states what that source actually establishes — so a future reader can re-check a specific claim without redoing the whole page. Each decode shows the date its facts were last verified. AI products change their terms often; a decode is a snapshot, and an old one may be wrong.

We do not treat other people’s reviews, listicles or affiliate roundups as evidence.

Four axes, each scored 1 to 5. There is deliberately no overall score. Averaging these would destroy the only useful information in them — a tool can be excellent value and terrible for privacy at the same time, and that tension is the finding, not a rounding problem.

5 — genuinely good, and documented. 4 — fine, with a caveat worth knowing. 3 — mixed, or the honest answer is “it depends what you use it for”. 2 — a real problem a normal user would not expect. 1 — the thing most people would consider a dealbreaker if they knew.

These are judgements, not measurements, and we label them as such on every page. The facts underneath them are sourced; the score is our reading of those facts. If you disagree with a score but not the sources, that is the system working.

When a policy is silent on something it obviously should address, we report the silence as a finding rather than inferring what it means. If an app whose entire product is a food-recognition model never says whether your photos train that model, that gap is worth knowing about — and stating it as a gap is both more accurate and more defensible than guessing at intent.

Where a company is being sued or investigated, we label allegations as allegations, include the company’s response and its defence, and say plainly when no court or regulator has actually decided anything. An open case is context, not a verdict, and we will not write it as one.

Where research is preliminary — a conference abstract that hasn’t completed peer review, say — we put that caveat in the body of the page rather than a footnote. On questions that affect health or money, the caveats are the credibility.

Right now, we don’t. aiwut carries no advertising, no affiliate links, no sponsored placements and no paid inclusion, and nobody has paid to appear here or to change what a page says.

That will probably change — affiliate links are the normal way a site like this pays for itself. If it does, we’ll disclose it on the page it applies to and say so here, before you read the review rather than after. A tool’s rating will not depend on whether it has a programme we can join. We would rather tell you that plainly now than quietly add links later and hope you don’t notice.

If something here is wrong — especially if you work at a company we’ve decoded — tell us and we’ll fix it and note the change. Sourced corrections get made quickly. Requests to remove accurate, sourced information do not.