Ask a chatbot whether the Golf you've found on Marketplace is a good buy and you'll get a confident, well-written answer in about four seconds. Ask it what that car's MOT history says and it'll tell you, just as confidently, that it can't see it. That gap is the whole story of using AI to check a used car, and it's worth understanding before you trust either answer.
Can AI check a used car?
Yes, but only as well as the information it's given. A general assistant like ChatGPT, Gemini or Copilot has exactly one source when you ask about a car: the words you paste in. It hasn't seen the MOT history, it can't look up the registration, and the finance, theft and write-off registers are licensed to paid checks, so it has no route to them at all. An AI car check built for the job starts by fetching those records and only then asks the model to read the advert against them.
That distinction matters more than which model is cleverer. The facts about a car live in a handful of official and industry records. A model that can't reach them is guessing from the seller's description, however good the prose sounds.
What can ChatGPT actually tell you about a car advert?
Quite a lot, as long as it's about the words rather than the car. Paste an advert in and a general assistant will:
- Decode the trade phrases. "Spares or repairs", "sold as seen", "HPI clear" and "FSH" all have specific meanings and it knows them. Our own decoder is at what used car advert phrases really mean.
- Spot the obvious tells in the writing. A price well under the market, a seller "selling for a friend", a listing with no registration and no photos of the interior. It'll list the used car advert red flags as well as most humans would.
- Write your questions for the seller. Ask it for ten questions to send before you travel and you'll get a sensible list.
- Tell you what the model is generally known for. Ask about a 1.2 PureTech and it'll mention the wet belt. Ask about a DSG gearbox and it'll mention the mechatronics unit.
All of that is useful and all of it is free. Use it. The problem starts with the next question.
What can't a general AI assistant see?
It can't see anything the seller didn't type. Here's the honest comparison of what each one is working from:
| What you want to know | General assistant (ChatGPT, Gemini, Copilot) | Car Advert Check |
|---|---|---|
| What the advert says, and how it says it | Yes, from the text you paste | Yes, from text or screenshots |
| Whether the mileage claim matches the record | No. It has never seen the MOT history | Yes. Every mileage reading from every MOT, pulled live from DVSA |
| What the advisories and failures were | No, unless you paste them in yourself | Yes, the full history, read alongside the advert |
| Whether the car has an outstanding recall | No | Yes, where the manufacturer has reported it to DVSA |
| Which faults this exact model and engine gets | General knowledge, unverified, sometimes the wrong engine | 791 documented faults, each with its typical cost, and 2,259 advert phrases sellers use to talk around them |
| Outstanding finance, theft marker, write-off record | No. Those registers aren't public | Provenance Check, from the same industry records every paid check uses |
| Whether it gives the same verdict twice | No. Ask twice, get two answers | Yes. Fixed scoring rubric, so the same facts give the same score |
| Unrecorded accident damage, repair quality | No | No. Nobody has this |
Two rows deserve a second look. The known-faults row is where a general assistant is most dangerous, because it answers fluently from training data and it's right often enough that you stop checking. Ask about a model it half-knows and it'll blend two engines together or quote a fault from the previous generation. The consistency row is the quieter problem: a chatbot reasoning freely about risk will land on "cautious optimism" one minute and "walk away" the next from the same advert, and you'll remember whichever one you wanted to hear.
How does an AI car check actually work?
It gets the records first, then reads the advert. Here's ours, step by step:
- You paste the advert and the registration. Text, or screenshots if you're on a phone, which is most people. We read the text, the photos and the details.
- We fetch the DVSA MOT history live. Every test the car has had, the mileage at each one, every advisory and every failure. It's the same free record you could read on GOV.UK, except that we're about to compare it with something GOV.UK has never seen: the advert.
- We load the known faults for that make, model and engine. Not a general impression of the model. Named faults, each documented, with what they cost to fix and the advert phrases that tend to be used to cover them. Rows we're not confident in are left out of every report.
- The model reads the seller's words against all of that. Does "genuine mileage" fit the MOT sequence? Does "full service history" square with the same oil-leak advisory three tests running? Is "just needs a sensor" one of the phrases that turns up on a known fault for this engine?
- You get one verdict, scored the same way every time. A risk score built from a fixed rubric, the biggest single finding, and the reasons. On a Provenance Check, outstanding finance, a theft marker or a Cat A or Cat B write-off force the verdict to its highest risk rating outright, and a Cat S or Cat N holds it at caution, whatever the rest of the car looks like.
The AI is the reader, not the source. That's the design decision that makes the answer trustworthy, and it's the one a general chatbot can't make for you.
What does that look like on a real advert?
Take a typical private advert (this one's made up, but you'll have seen its cousins): a 2016 Nissan Qashqai 1.5 dCi, "genuine 71,000 miles, full service history, drives spot on, new MOT, just needs a sensor for the warning light".
Ask a general assistant and you'll get a fair reading of the words: "just needs a sensor" is a phrase to probe, "drives spot on" means nothing, ask for the service invoices. All true, and all you'd have worked out yourself.
Run it through a check that has the records and the shape changes. The MOT history shows what the mileage was at every test, so "genuine 71,000" is either consistent with the sequence or it isn't, and you know which before you reply. "New MOT" is on the record too, along with any advisories the seller didn't mention. And "just needs a sensor" is checked against the Nissan Qashqai known faults, where the documented problems for that engine are listed with their costs, some of which run to a great deal more than a sensor. The verdict weighs all of it together rather than handing you a list.
Same advert, same words. The difference is what was in the room when they were read.
Is an AI car check accurate?
As accurate as its inputs, which is why the inputs matter more than the model. The MOT history is the official DVSA record. The known faults are documented, with sources. The advert is the seller's own words. The AI's job is the reading: noticing that three separate facts add up to something none of them says on its own, which is exactly the part humans skip once they've fallen for the colour.
Where it can be wrong is where the record is silent. A car with no MOT history yet, a fault that's never been documented, damage nobody ever reported: no check reads those, ours included. And no AI can look at the car. If the report says the advert and the record agree, you still go and see it, drive it and check the paperwork, in the order what to look for when buying a used car sets out. If you want the decision itself walked through, from the advert to the point where you either message the seller or don't, should I buy this car is the companion to this piece.
Why haven't we called it an AI car check until now?
Because "AI" has been stuck on everything from toothbrushes to tax software, and we'd rather you judged the report than the label. But we've been doing exactly this since the first report in March 2026, more than a thousand reports ago, and people have started searching for it by name. So, for the record: this is an AI car check, it was built for UK used-car adverts specifically, and the AI is the last step in the pipeline rather than the whole of it.
Which of these gaps can a check actually close?
Start with the honest bit: we don't hold records nobody else has. The MOT history is free to anyone with the registration. Finance, theft and write-off data come from the same industry registers every paid check draws on. What a general assistant lacks isn't secret data. It's any data at all.
So the comparison isn't really us against ChatGPT. It's a report built on the car's records against an essay built on the seller's description.
| The gap | Who closes it |
|---|---|
| Decoding the advert's wording | A chatbot does this well, and so do we |
| Testing the mileage claim against the record | The free scan, from the DVSA history |
| Knowing which faults this engine actually gets, and what they cost | The free scan names the biggest; the Smart Check explains every finding with repair costs |
| Turning it into questions for the seller and a figure to negotiate from | Smart Check |
| Finance, theft, write-off | Provenance Check, which includes the full Smart Check |
| Unrecorded damage, repair quality, how it drives | Nobody. You, at the car |
For someone who's just typed "ai car check", the place to start is the free scan, because it's the thing a chatbot structurally can't do: it reads the advert with the car's records open next to it. You get the verdict, the risk score and the biggest finding without a card or a password.
So what should you actually do? Keep asking the chatbot to draft your questions if you like, it's good at that. Then run the free scan on the car you're looking at, with the advert and the registration, and see whether the seller's words survive contact with the record. If they do and you want every finding explained, the report upgrades on the same link with nothing re-run, so nothing's wasted if the free verdict turns out to be "walk away".
