Can AI Detectors Tell Which Model Wrote It — ChatGPT vs Claude vs Gemini?

There’s a tidy little myth floating around the internet: that an AI detector can sniff a paragraph and announce “this was ChatGPT” or “that’s Claude” the way a sommelier names a vintage. It sounds plausible. It’s mostly wrong. The honest answer is that telling *human from machine* is already shaky, and telling *which* machine is a harder problem that almost no everyday tool actually solves.

Key takeaways

• Most mainstream detectors answer one question — human or machine — not which machine. Naming ChatGPT, Claude, or Gemini is a separate, much harder problem.
• A few tools advertise model attribution, but their accuracy drops fast on edited text, short passages, and any model released after they were trained.
• Detectors flag the generic statistical style that large language models share, not a brand signature — which is why a Claude draft and a ChatGPT draft can score nearly the same.
• Light editing, paraphrasing, or mixing two models scrambles any model-ID guess almost immediately.
• Treat a “this was written by GPT-4” claim as a probability with a wide error bar, never as proof.

Two very different questions

People collapse two questions into one, and that’s where the confusion starts.

The first question is binary: *did a human or a machine write this?* That’s what Turnitin, GPTZero, Copyleaks, and Originality are built to estimate. They output a probability — a percentage — that the text is machine-generated.

The second question is multi-class: *which* model, out of dozens, produced it? ChatGPT (and which version), Claude, Gemini, Llama, DeepSeek, Mistral, or something fine-tuned in someone’s basement last week. That’s not a yes/no call. It’s a “pick one of many,” and every new model release adds another option the tool has to know about in advance.

The first question is hard. The second is much harder. Conflating them is how the myth survives.

What detectors actually measure

Under the hood, most detectors don’t look for a secret ChatGPT watermark or a Claude logo. They measure statistical habits — chiefly how *predictable* the text is to a reference language model, and how *evenly* that predictability is spread across the passage. Low surprise plus low variation reads as machine. We walked through the exact mechanics in our explainer on perplexity and burstiness, and the short version matters here: these are *family* traits, not *individual* ones.

Why does that wreck model identification? Because today’s big models are trained on overlapping internet-scale data, tuned with similar reinforcement-learning recipes, and nudged toward the same polite, fluent, hedge-everything register. They converge. The “voice” that gets a paragraph flagged — smooth transitions, balanced sentences, careful qualifiers — is shared property across ChatGPT, Claude, and Gemini. A detector reading those shared traits sees “machine,” not “this particular machine.”

There are faint differences. A given model might lean on certain transition words, default to certain list structures, or hedge more often. Researchers have built classifiers that exploit exactly these tics, and some of that work reports decent model-attribution accuracy on clean lab data. But “clean lab data” is doing a lot of quiet work in that sentence.

Where model-ID tools fall apart

Picture a teaching assistant — call her Dana — who pastes a flagged essay into a tool that proudly labels it “GPT-4 generated.” Here’s everything that label is quietly hiding:

Short text. The essay’s flagged paragraph is 90 words. Model attribution needs a lot of signal to separate near-identical styles, and short passages barely have enough to call human-versus-machine, let alone brand. Confidence on a paragraph is mostly theater.

Edited text. The student ran the draft through a grammar tool and rewrote two sentences by hand. Every edit shifts the token-probability pattern the attribution model relies on. The “GPT-4” guess was built on statistics that no longer fully exist in the text.

Newer or unknown models. If the student used a model released after the tool’s training cutoff — or a fine-tuned variant — the tool can’t name what it’s never seen. It still has to pick *something* from its list, so it picks the closest match and presents it with the same confident UI. Detection accuracy already drops with new models; brand-level guessing drops faster.

Mixed sources. The student drafted in Gemini, polished in ChatGPT, and tightened a paragraph in Claude. Now whose fingerprint is it? Nobody’s, cleanly. The blend looks like a fourth thing.

So Dana’s “GPT-4 generated” label is, at best, a hint dressed up as a verdict. At worst it names the wrong product entirely — which is a real problem if anyone treats it as evidence.

“ChatGPT vs Claude vs Gemini”: a quick reality check

Suppose you ran the *same prompt* through all three models and fed each answer to a standard detector. What happens?

Usually, all three land in roughly the same zone — often all flagged as likely machine, sometimes all squeaking by if the prompt produced loose, varied prose. The detector isn’t ranking the brands. It’s reacting to how machine-like each output happens to be on that topic. Swap the topic and the scores shuffle, because difficulty of the subject affects predictability more than the model’s identity does.

That’s the tell. If detectors were really identifying models, the *brand* would drive the score. Instead, the *content and editing* drive it. Brand is noise.

The watermarking wrinkle

There’s a partial exception worth naming honestly. Some providers have experimented with *watermarking* — deliberately biasing word choices in a hidden statistical pattern so their own output can be recognized later. Google DeepMind’s SynthID-Text is the best-known effort; it nudges the model’s token-sampling during generation and ships a matching detector, and DeepMind open-sourced the scheme in 2024. In principle, a watermark is a brand signature, and a matching detector could say “this came from our model.”

Two big caveats. First, watermark detection only works for models that opted in, and only with the matching detector — it tells you nothing about a model that never watermarked. Second, watermarks are fragile: paraphrasing, translation, and synonym swaps degrade them, something DeepMind’s own write-up acknowledges. So even this isn’t a general “which AI wrote it” superpower. It’s a narrow, provider-specific, easily-disturbed signal — not the thing a classroom detector is using when it slaps a model name on an essay.

Why this matters if you’re the one being judged

If a tool tells a professor, an editor, or a client that your work was “written by ChatGPT,” understand what that claim is made of: a probability about *style*, often with a default model name attached for flavor, computed on text that may be too short and too edited to support any of it. OpenAI pulled its own classifier in July 2023 over low accuracy. A 2023 Stanford study found detectors systematically biased against non-native English writers. A brand label doesn’t escape those limits — it inherits them and adds a fresh layer of false precision on top.

The useful response isn’t to chase some model that “beats” attribution. There’s no such model, because the thing being measured is generic machine-style, not a brand. The useful move is to make sure your own thinking and voice are actually on the page — concrete examples, real specifics, sentences that vary, claims a generic model wouldn’t make. That’s what reads as human, and it’s also just better writing. If you want to dig deeper into how detection really works, browse the rest of the blog, or compare plans when you’re ready to put a draft to the test.

Frequently asked questions

Can any AI detector reliably tell ChatGPT from Claude from Gemini?

Not reliably. A handful of research tools and a couple of commercial products attempt model attribution, and on clean, long, unedited samples they can beat chance. But accuracy collapses on short text, edited text, or output from a model the tool wasn’t trained on. The everyday detectors — Turnitin, GPTZero, Copyleaks, Originality — mostly don’t claim model-level identification at all. They answer human-versus-machine, and even that is a probability, not a fact.

Why would a tool even want to name the model?

Mostly forensics and research — tracing where synthetic text in a dataset came from, or studying how models leave different traces. For the typical “did a person write this” use case, the brand barely matters, which is why most products skip it. Naming a specific model also creates legal and reputational risk if the guess is wrong, so vendors hedge.

Does editing the text break model identification?

Yes, quickly. Attribution leans on subtle token-probability patterns, and those are fragile. Paraphrase a few sentences, splice in your own examples, or run the text through a second model, and the fingerprint smears. It’s the same reason whole-text scores swing after editing — you’re changing the very statistics the tool reads.

If a detector says “GPT-generated,” is that proof a student used ChatGPT?

No. It’s a probabilistic guess about style, and the label is often just the tool’s default name for “looks machine-like,” not a confirmed source. OpenAI retired its own classifier in July 2023 for low accuracy, and the 2023 Stanford study found bias against non-native writers. A model name attached to a score doesn’t turn a guess into evidence.

Will model attribution get better over time?

Maybe somewhat, but it’s running uphill. New models ship constantly, providers update them quietly, and outputs keep getting edited and mixed. Each move shifts the target the tool was trained on. General detection already lags behind new models; brand-level identification lags even further.

The bottom line

Can AI detectors identify which AI wrote text? In the narrow, lab-clean sense — sometimes, a little. In the way the myth implies — pointing confidently at ChatGPT or Claude or Gemini from a flagged paragraph — no. They’re reading the shared accent of machine writing, not a brand stamp, and that accent fades the moment text is edited or mixed. So don’t fear a model name on a score, and don’t trust one either. Want to see what a detector actually keys on in your own writing? Run a draft through the free checker and watch which sentences light up.