The short version
Use ChatGPT on a used car listing the way you'd use a sharp friend who has never seen the car: paste the text, let it flag the story and the contradictions, take its question list to the seller, and verify every photo read it gives you. Never hand it the bare link and trust what comes back, and never ask it about recalls, value, or history — it will answer, and it doesn't know. Run the four steps once on a real listing and you'll see the split for yourself: a strong reader of words, a shaky reader of pictures, and a confident inventor of facts. Keep it in the first job.
Paste a used car listing into ChatGPT, ask "is this a scam?", and you'll have a confident, well-organized answer in about ten seconds. That is the appeal and the problem in one. A general chatbot is very good at sounding like it checked a car, and there's a real gap between reading an ad and checking one. Whether the answer is worth anything depends on what you actually gave it and what you ask next.
So run it as a test rather than a verdict. Below is the four-step version — the link, the text, the photos, then a question about the car itself — with what a general chatbot typically does at each step and the exact points where it stops being a tool and starts being a storyteller. Fifteen minutes with a listing you're genuinely considering will teach you more than any review of the model.
Step one: paste the link
Start with what most people do first: paste the Facebook Marketplace or Craigslist URL and ask what the chatbot thinks. Watch closely, because this is where the most misleading result lives. Depending on the version you're using and whether web browsing is turned on, one of three things follows. It says it can't open the page — honest, and fine. It opens the page but hits a login wall, a consent prompt, or a stripped-down version with no photos, and reasons from that — partial, and it may not say so. Or it doesn't open the page at all and answers anyway, producing a plausible description of a car it has never seen.
That third case is the trap. A Marketplace or Craigslist link carries almost nothing in the URL — an item ID, sometimes a city — and a model asked about a link it can't fetch will sometimes fill the gap with a generic sedan at a generic price and a generic list of concerns. It reads like analysis. It's fiction, and nothing in the formatting tells you which one you got.
The test is simple: ask it something it could only know by opening the page. The interior color, the exact mileage, the last sentence of the description. If it can't answer, or answers wrong, then everything else it said about that link counts for nothing, and you move to step two.
Step two: paste the text
Now copy the ad's actual text — title, price, mileage, location, the full description — and paste it in. This is where a general chatbot earns its keep. Language models are pattern readers, and scam listings are written to a pattern. Give it the words and ask three things: what's suspicious, what's inconsistent, and what should I ask the seller. Expect it to do these well:
- Spot the borrowed story — the deployment, the divorce, the late relative's car in storage, the shipping company that will "hold your money" — and explain why each one keeps showing up in fakes
- Catch contradictions inside the ad: a mileage in the headline that disagrees with the body, a trim that model year never came in, "one owner" next to "selling for a friend"
- Flag payment language — a deposit to hold the car, wire transfer, gift cards, a payment app's friends-and-family mode — and any push to move the conversation off the platform
- Notice dealer boilerplate in a supposedly private listing, which is the signature of a curbstoner
- Turn every flag into a question for the seller, which is honestly the most useful thing it produces all session
Where the text read runs out
Two limits, and they matter. First, it has no market data. When it calls the price "suspiciously low" it's inferring that from the story around it, not from comparable listings — you still have to search the comps yourself, in your area, for the same year and trim. Second, it can only find contradictions that are on the page. A clean, boring, correctly written ad gets a clean, boring result whether or not the car behind it exists, and the better scams are written by people who know that.
Treat a text read as a first filter with a good memory, nothing more. It will flag the tired-sounding listing you'd have flagged yourself after coffee, and it will do it every time without getting bored. It will not tell you the car is real.
Step three: upload the photos
Save the listing's photos and upload them alongside the text. The current models read images, and a used car ad is mostly images, so this ought to be the strongest step. In practice it's the most uneven one. Ask it to read the odometer, find the VIN, and say whether the photos agree with each other and with the ad — and check its work:
- An odometer reading from a sharp, straight-on dash photo is usually right; from a small or angled one, a single digit can come back wrong, and a wrong digit is worse than no reading
- A VIN read through the windshield or off the door-jamb sticker can come back with a character swapped. Always confirm the VIN the seller texts you against a free decoder and the NHTSA recall lookup before you rely on it
- It can notice a warning light on the dash, a plate blurred in every shot, a bumper a shade off from the fenders, or an interior worn far past the claimed miles — things you could see too, if you looked as patiently
- It will not, on its own, search the internet for the photos. Stolen pictures are the classic fake, and finding the original ad takes a reverse image search through Google Lens or TinEye — do that by hand
- Ask it to describe each photo before you trust its read of it. If the description is wrong — wrong color, wrong body style, a detail that isn't there — so is the analysis built on it
Step four: ask about the car itself
Here's where the test turns. Ask a follow-up any real buyer would ask: "What are the common problems with this year and model?" or "Does this car have any recalls?" or "What's it worth?" A general chatbot will answer all three fluently, and it has no way of knowing any of them for this car. It can't look up recalls by VIN. It can't see a title record, a lien, or an odometer history. It doesn't know what this trim is selling for in your zip code this month. What it produces is a blend of whatever it read about the model in general, phrased as if it looked something up.
Sometimes that general knowledge is decent — widely documented model issues do get named. But you can't tell from the answer which parts are real and which are pattern completion, and a recall it invents arrives with exactly the same confidence as one that exists. The fix is to route those questions to the sources that actually hold the data: the NHTSA recall lookup by VIN for recalls, NICB VINCheck for theft and total-loss records, official title records through an approved NMVTIS provider, and real comparable listings for price. Use the chatbot to write the questions. Don't let it answer them.
What a purpose-built check does differently
None of this makes a general chatbot useless. It's a tireless reader of ad text, and that alone catches the lazier half of the fakes. But the four steps show the shape of the gap: it can't reliably open the link, it reads photos unevenly, it never searches for them, and it answers questions about the car from memory. A tool built for this one job is designed around each of those failures — it fetches the listing itself so there's nothing to hallucinate, pulls the photos rather than asking you to upload them, runs the reverse image search, compares the asking price against real comps, and reports what it verified separately from what the ad merely claims. A free browser listing check like SiftCar's takes the ad's link and does that in one pass.
The line to hold with any of these tools, chatbot or purpose-built, is the same: they read the ad, and the ad is the seller's version. What any AI can and can't see in a listing ends at the edge of the photos. The checks that reach the car are still yours — the VIN decode, the records, the test drive, and a mechanic under it before any money moves.