In May 2024, a car website ran a story about the worst vehicle someone had ever owned. The headline named the culprit: a “2023 BMW BRZ.” One problem – the BRZ has never been a BMW. It’s a Subaru, built with Toyota, and it has existed since 2012. The article had been scraped from a Reddit thread, run through an AI tool, and published without anyone noticing that the software had welded two unrelated usernames and car names into a vehicle that doesn’t exist, with total confidence. The story was picked up widely enough that one outlet’s headline about it just said “EMBARRASSING” in enormous letters, and it’s since become something of a case study among motoring writers about what happens when nobody checks the machine’s work.
That’s the funny version of the problem. The less funny version is quieter: reviews that read fine, sound plausible, hit all the right notes about ride quality and interior fit, and were never attached to an actual car, an actual test drive, or an actual owner. If you’re shopping for a vehicle right now, some of what you’re reading was very likely written that way – and increasingly, there’s a specific, if imperfect, way to check.
How much of this is actually happening?
More than a rounding error, and the number is moving in one direction. Originality.ai ran its detection model against Canadian car dealership reviews collected from 2020 through 2025 and found that 28.6% of 2025 reviews were likely AI-written, up from an average of 14.8% across the full five-year window – and the increase tracks almost exactly with ChatGPT’s late-2022 launch, staying under 3% before that. The same Originality.ai analysis of Canadian dealership reviews found something I didn’t expect going in: AI-flagged reviews weren’t uniformly glowing. Extreme ratings dominated – 71.1% were five stars, but 26.5% were one star – which tracks with two obvious motives: a dealership padding its own reputation, or a competitor, ex-employee, or annoyed customer generating a pile of scathing ones. Middling, three-star reviews, the kind a real mixed experience usually produces, were comparatively rare among the AI-flagged set.
It isn’t only the customer-review side. Automotive editorial has its own version of this, and it’s less hidden than you’d think. MotorTrend’s editor-in-chief has openly said parts of a recent piece were written with Hearst’s in-house ChatGPT tool, around the same time Hearst (MotorTrend and Car and Driver’s parent company) struck a licensing deal reportedly worth millions with OpenAI. Meanwhile, sites with no automotive pedigree at all have been publishing over 200 articles a day into Google Discover’s motoring feed, next to actual outlets like Autocar and Top Gear, with none of the editing effort spent hiding it. The Autopian’s David Tracy put it bluntly after finding an AI tool had convincingly imitated his own writing voice: “no fact in anything AI writes can be believed to be true” – not because the AI is lying, in his framing, but because it has no idea what’s true to begin with. It’s assembling plausible word sequences, not reporting on a car it drove.
Why this actually costs you money, not just annoyance
A car is usually the second-largest purchase most people make, and for a lot of buyers right now – with used-car prices where they are – it might be the largest. Reviews aren’t a nice-to-have in that decision; they’re load-bearing. Industry research from CDK Global found 70% of shoppers say good reviews and ratings matter when choosing where to buy, and Demand Local’s consumer research puts the read-before-you-buy figure even higher: 89% of consumers read reviews before a purchase, more than half read at least six of them, and 78% say they trust automotive reviews about as much as a recommendation from a friend. YouGov’s 2025 data on how Americans shop for cars has online reviews as the single most-used research source, ahead of dealership visits, brand websites, and social media, at 56%.
Put those two data points next to each other and the shape of the problem gets obvious: reviews carry as much weight as word-of-mouth, and a rising share of them were never spoken by a mouth at all.
What an AI detector is actually measuring
Worth being precise here, because “AI detector” sounds more magical than it is. Most tools – free AI detector ZeroGPT included – are trained classifiers that look at two statistical properties of text: perplexity (how predictable each word choice is, given the words before it) and burstiness (how much sentence length and structure vary across a passage). Large language models tend to pick high-probability words and produce evenly paced sentences. Humans ramble, trail off, and write one three-word sentence followed by a forty-word one. A 2026 benchmark from Reviewz.ai that tested detectors specifically on short-form reviews measured this directly: the standard deviation of sentence length across their human review sample was 8.4 words, versus 4.1 words for AI-generated ones. That gap – not any single magic word – is most of what a detector is scoring.
It is not reading for truth. A detector cannot tell you whether the torque figure in a review is correct, whether the writer actually drove the car, or whether the “common transmission problem at 100,000 miles” it mentions is real or invented. That distinction matters more than it sounds like it should.
The tools, tested against each other
Detector vendors publish their own accuracy numbers, and those numbers are consistently more flattering than what independent testing finds. The gap tends to run somewhere between 15 and 25 percentage points, and it exists because vendors mostly test against clean, unedited AI output – not the messy, lightly-edited, sometimes-paraphrased text that actually shows up in the wild.
| Detector | Access | Independent accuracy (general text) | False positive rate on human text | Where it breaks down |
|---|---|---|---|---|
| ZeroGPT | Free | ~71–80% | ~16–26% | Short reviews under 50 words; anything paraphrased |
| GPTZero | Free tier + paid | 52–96% (varies enormously by test) | 1–18% | Huge swing between vendor-controlled and independent tests |
| Originality.ai | Paid only | 76–98% | 5–15% | Non-native English writing patterns |
| Two detectors + a human read-through (ensemble) | – | ~95% | ~6% | Still the closest thing to reliable, per current testing |
Figures compiled from Reviewz.ai’s 2026 12-tool review-specific benchmark, the RAID benchmark, and Scribbr’s independent detector comparison, current as of mid-2026. No single tool in that table clears 90% on its own against real-world text – the ensemble row is the only one that does, and it does it by making three different, weaker signals vote.
Where every one of these tools quietly falls apart
A few failure modes show up across nearly every study I read for this piece, and they matter more than any individual accuracy percentage:
- Short text. Below roughly 250 words – which describes most customer reviews – accuracy drops toward a coin flip. A one-paragraph dealer review is close to unclassifiable.
- Paraphrasing. Running AI text through a tool like QuillBot before posting it can cut detection rates by more than half; one 2025 study on adversarial paraphrasing measured an average 87.88% reduction in detector effectiveness.
- Human editing of AI drafts. University of Chicago Booth research found that a person lightly editing an AI draft – the most common real workflow in 2026 – drops detector accuracy from over 90% to somewhere around 70–80%. This is also the exact workflow MotorTrend has publicly described using.
- Non-native English writers. A widely cited 2023 Stanford study published in the journal Patterns found that seven leading detectors falsely flagged over 60% of TOEFL essays written by non-native English speakers as AI-generated, while scoring near-perfect on native-speaker writing. The bias has narrowed since, not disappeared.
None of this makes detectors worthless. It makes a single score from a single tool a bad place to stop looking.
A real example, run through the process
Late in 2025, a tipster pointed The Drive toward a Ford dealership’s reviews on CarGurus. One five-star review described a family’s “amazing experience” buying an F-150 – and then, mid-review, pivoted to praising the truck’s third-row seating. The F-150 doesn’t have a third row; the Ford Explorer does. Read the review on its own, it’s plausible. Read a dozen from the same dealership together, and the pattern The Drive documented becomes obvious: identical sentence templates, clean grammar with no typos, always naming the year/make/model/trim (great for search visibility, incidentally), and stray extra spaces suggesting they’d been pasted out of some content tool. That’s the “statistical perfection” tell Reviewz.ai’s researchers flagged independently: real reviews have noise – typos, inconsistent capitalization, the occasional all-caps outburst. A batch of fifty reviews with none of that is itself a signal.
How to actually run the check
Paste the review text into a detector – ZeroGPT is free and doesn’t require an account, which makes it a reasonable first pass before you spend money on anything else. Then treat the score as one input, not a verdict, and look for the things detectors are bad at catching:
- Does it mention anything the car doesn’t have? Wrong trim, wrong seating configuration, a feature from a different model year. This killed the Huntley Ford reviews and the “BMW BRZ” story alike.
- Is there any friction? Real reviews mention the annoying stuff – a slow finance department, a scratch the dealer fixed, a weird noise at 40 mph. AI-written reviews tend to stay smoothly on-topic and rarely name a specific competitor model unless prompted to.
- Does the punctuation look human? Genuine reviews have typos, run-ons, and inconsistent capitalization. A cluster of reviews with flawless grammar and matching structure is a bigger red flag than any single AI score.
- Does the star rating cluster oddly? A page full of only 5-star and 1-star reviews, with almost nothing in between, matches the pattern found in the AI-flagged Canadian dealership data.
If a review fails two or three of these on top of a high AI-detector score, that’s a reasonably confident fake. If it only trips the detector and nothing else – especially if it’s under 50 words, or reads like it was written by someone who learned English as a second language – I’d treat the AI score itself with real suspicion.
The case against bothering with any of this
There’s a serious argument that detectors aren’t worth the effort at all, and it deserves more than a dismissal. Turnitin’s own chief product officer has acknowledged the platform intentionally skips flagging roughly 15% of AI-influenced text specifically to keep false accusations down – meaning its real detection rate sits closer to 85% than the 92–100% sometimes quoted. One outlet that spent two weeks reviewing the public evaluations of the leading tools put it about as skeptically as I’ve seen it put: “the best evidence for or against AI authorship is rarely the detector… We will quote the verdicts. We will not believe them.” Academic researchers studying AI-generated product reviews specifically found something similar – most detectors either fail to make useful predictions on this kind of short, informal text, or make them with so much noise that they’re barely better than guessing, and can’t reliably tell a fully AI-written fake review from a real one that a person cleaned up with AI afterward.
I think that argument is right about the limits and wrong about the conclusion. A tool that’s correct 75–85% of the time, used as one signal among several rather than a final ruling, still beats reading a review cold and trusting your gut – which is what almost everyone does today. The honest position isn’t “detectors work” or “detectors are useless.” It’s that no single free tool, ZeroGPT included, should be the whole process – but skipping the check entirely, when the data says nearly three in ten dealership reviews you’re about to read might not be real, isn’t the safer option either.
Quick answers to what people actually ask next
Does a low AI-detection score mean the review is definitely true? No. It means the writing pattern resembles human text – it says nothing about whether the facts inside it are accurate. A human can still lie, and a real customer can still get the trim level wrong.
Are dealerships breaking any rules by posting AI-written reviews? The FTC’s 2024 rule on fake reviews explicitly prohibits presenting AI-generated reviews as genuine consumer feedback, regardless of the industry.
Should I trust professional car-review sites more than customer reviews? Generally yes, but verify which ones actually road-test cars with named reviewers – outlets like Edmunds and Consumer Reports run in-house testing programs and disclose it, which is a stronger trust signal than polish alone.
How this article was put together. Figures on AI-generated dealership reviews come from Originality.ai’s 2025 analysis of Canadian dealership data; detector accuracy figures were cross-checked across Reviewz.ai’s 2026 review-specific benchmark, the RAID benchmark, and Scribbr’s independent 12-tool comparison, all current as of mid-2026. The Huntley Ford and “BMW BRZ” cases are drawn from reporting by The Drive and accounts from motoring-industry sources, respectively, both checked directly rather than via secondhand summaries. Detector accuracy is a moving target – generation models improve every few months, and these numbers should be rechecked in another six to twelve months rather than treated as fixed.
