Mostly, yes, but the margin is thinner than it was a year ago. A January 2026 study published in the Journal of Advances in Information Technology ran nine AI detectors, including our own AI Detector, against text from ChatGPT, DeepSeek, Gemini, and Grok. One commercial tool caught all of it: 100% accuracy across every model tested and the human-written control set. A free tool in the same lineup caught as little as 63%. Then the researchers paraphrased the AI text before rescanning it, and the leaderboard reshuffled again. A score that looked rock solid on raw output nearly cut itself in half.
That spread matters more this summer than it did last year, because the model doing the writing has changed again, and detectors have to relearn its fingerprints every time it does.
The model lineup moved again this summer
Anthropic shipped Claude Sonnet 5 as its new default model on June 30, 2026, and it now edges out the company's own Opus 4.8 on professional-deliverable benchmarks. Opus 4.8, released in late May, still holds the top spot on overall intelligence rankings, and Anthropic's Fable 5 has become the pick for coding tasks. OpenAI moved its ChatGPT default to GPT-5.6 on July 9. Google's Gemini 3.1 Pro has led most major benchmarks since February, with a 2.5-million-token context window that dwarfs anything else on the market, while Gemini 3.5 Flash handles the bulk, lower-cost end of the lineup. DeepSeek and Moonshot AI's Kimi K3, released July 16, round out the field.
Every one of those releases resets the clock for detection vendors. A detector trained on last quarter's model outputs is, by definition, guessing when it meets a model that did not exist when the training data was collected.
What the 2026 study actually found
The JAIT study is worth taking seriously because it is independent, peer-reviewed, and specific about numbers instead of vague about "accuracy." Researchers ran Turnitin, ZeroGPT, detecting-ai.com, GPTZero, QuillBot, Grammarly, Sapling, Copyleaks, and Originality.ai against the same four models and a set of human-written control texts.
On untouched, unaltered AI text, the paid classifiers clustered towards the top of the pack, with Originality.ai maintaining a perfect 100% rate across all the models. The free detectors lagged considerably behind, with some detection rates dipping as low as 63%. The real plot twist came with the software companies' use of paraphrasing and re-writing the same AI writing in the style of non-native English writers, then running the models again. Turnitin saw its accuracy drop to 45.7%. Grammarly fell all the way down to 19%. Copyleaks, GPTZero, and Sapling were the three products that kept their accuracy intact when put under that duress.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
An independent academic comparison choosing to include detecting-ai.com alongside the field's established names is its own kind of validation, a different one than a vendor's own benchmark page. On our own testing, the same detector is rated at 99% accuracy with a false positive rate under 1%, though we would rather you weigh that against independent studies like this one than take our word for it alone.
Why the newest models slip past older detection logic
Detectors have historically leaned on two signals: perplexity, or how predictable the word choices are, and burstiness, or how much sentence length and rhythm vary. Older AI models scored low on both, which made them easy to flag. Newer models are trained with heavy human-preference tuning, and human raters tend to reward smoother, more natural-sounding rhythm. That training pressure nudges output toward the same perplexity and burstiness ranges as human writing.
The pattern is visible even in edge cases outside the mainstream chatbot field. Diffusion-based language models like LLaDA have posted perplexity scores nearly identical to human text, 43.03 versus 44.62 in one measurement, while still showing lower burstiness than a human writer would. That combination is exactly the profile that slips past a detector tuned on older assumptions.
There is a second, more structural problem: generalization. Research on detector performance shows that per-model accuracy on hard cases can sit above 0.90 for a model the detector was trained on, then drop below 0.80 the moment a genuinely new model shows up. That is not a bug a vendor can patch once. It is the reason different detectors give you different scores on the exact same paragraph: each one is only as current as its last retraining cycle.
Where detection still holds up
None of this means detection has stopped working. Commercial tools with frequent retraining cycles, Copyleaks, GPTZero, and Sapling among them in the JAIT results, kept catching most AI text even after paraphrasing. The gap opens widest with free tools running older models and with any tool that has not published an update in a while.
It is worth checking a detector's own model-support notes before trusting a same-day claim about a brand-new release. GPTZero's published support list, for example, names GPT-4.1, o3, Gemini 2.5, and Claude Sonnet 4, a full model generation behind Sonnet 5 and Gemini 3.1. That is not a knock on GPTZero specifically. Every detector, including ours, runs a training cycle behind the newest release by definition. A vendor's marketing page claiming "works on every 2026 model" the week that model ships deserves the same skepticism you would give a nutrition label with no ingredients list. Independent numbers also tend to run lower than vendor claims: one comparison found Copyleaks scoring 76% to 91% in outside testing against a vendor-stated 99.52%, a reminder to weight third-party tests over a company's own benchmark page.
What this means if you are checking text today
The practical takeaway is not "detection is dead." It is "treat a single score with the same caution you would treat a single data point anywhere else." That is truer than ever with a model that shipped in the last few weeks.
Before you trust a detector score
- Check whether the tool's own documentation lists the specific model you suspect, not just "ChatGPT" or "AI" generically
-
Run a second detector if the writer used a model released in the past month or two -
Treat a borderline score as inconclusive rather than as a verdict, especially for non-native English writers, who see higher false-positive rates across nearly every tool -
If the text may have been paraphrased or run through a humanizer, expect accuracy to drop further and weight the score accordingly, see our breakdown of what "undetectable AI" claims actually hold up -
For high-stakes calls like academic integrity or hiring decisions, pair the AI score with a source-based plagiarism check and a human read before acting on it
Our own AI Detector retrains against new model output on a rolling basis and uses sentence-level highlighting so you can see exactly which lines triggered a flag rather than trusting one blended score. The technology page walks through the dual-model architecture behind that, including the Paraphraser Shield layer built specifically for the obfuscation problem the JAIT study measured. None of that erases the honest limit here: no detector, ours included, can promise certainty on a model that shipped last week. What changed in 2026 is not that detection broke. It is that the gap between "reads AI text well" and "reads this month's AI text well" got wide enough to matter, and knowing where detectors still miss is now part of using one responsibly.