Do AI Detectors Work on Claude and Gemini? 2026 Cross-Model Tests
The honest answer is yes, current detectors catch Claude and Gemini output, but not as reliably as they catch ChatGPT. Independent 2026 cross-model tests put unedited GPT-5 text in the 85 to 92 percent detection range, Gemini 2.5 in the 80 to 88 percent range, and Claude 4 the lowest at 72 to 84 percent. The gap is not random: it tracks how much each model's output looks like the data detectors were trained on, and how long vendors have had to study it. Below we break down what that means for each model, why the gap exists, and what it implies if you are choosing a detector or checking your own work.
The short answer: detectors cover all three, unevenly
Across every 2026 benchmark we reviewed, all three frontier models are caught on a majority of unedited samples. None is 'undetectable.' The spread is roughly 20 points between the easiest and hardest, which is large enough to change outcomes in any single high-stakes check but not large enough to call any of them safe from scrutiny.
The practical takeaway is that model choice is a weak strategy. A well-humanized draft from GPT-5 scores about as low as a well-humanized draft from Claude, because post-processing matters far more than the source model. Vendors that train on a broad model mix, such as Originality.ai, which reports an ensemble covering 15-plus frontier models, hold up better across the board than those tuned to one family.
GPT output is the easiest to catch
GPT-5 sits at the top of the detection range for a simple reason: detectors have had the longest and largest exposure to OpenAI text. Turnitin, GPTZero and Originality.ai all built their classifiers on GPT-3.5 and GPT-4 samples years before the newer models arrived, so the statistical fingerprint of ChatGPT-style output is the one they know best.
That self-reinforcing loop means GPT-5 is still flagged at 85 to 92 percent on default settings even after its diversity improvements, while older GPT-3.5 text is caught around 88 percent. If you are testing whether a detector works at all, GPT output is the most forgiving demo; it is also the least representative of how the tool will behave on other models.
Claude is the hardest to flag
Claude 4 is consistently the lowest-detected of the three, with catch rates near 72 to 84 percent and some individual detectors dropping below 50 percent on Claude in 2026 testing. The cause is measurable: Claude produces more varied sentence lengths and a broader vocabulary, which pushes the burstiness and perplexity signals detectors rely on closer to human writing.
This advantage is temporary. As Claude's market share grows, vendors feed its output into retraining sets and the gap narrows within months. A classifier that misses a third of Claude essays today may catch most of them next quarter, so any claim that a model is 'undetectable' should be read as a snapshot, not a permanent property. Tools with a transformer classifier that generalizes across languages, such as Copyleaks, tend to lose less ground across model families than single-family specialists.
Gemini sits in the middle, caught by structure
Gemini 2.5 lands between the other two at roughly 80 to 88 percent, but it gets flagged for a different reason. Where ChatGPT trips detectors at the word level and Claude at the rhythm level, Gemini's tells are structural: rigid hierarchies, default bullet lists, and mechanical transitions between sections that newer document-level detectors now analyze.
For institutions this matters because Gemini is free inside Google Workspace, so students with a school account have zero-friction access. A detector that excels on ChatGPT but stumbles on Gemini will miss a meaningful share of submissions. Broad-coverage institutional tools such as Turnitin, already embedded in most university workflows, are the ones judged on cross-model performance rather than a single model.
Detection always trails the model release
The asymmetry is structural: improving a generator is an algorithmic change that can invalidate detectors overnight, while improving a detector requires retraining on millions of new samples. Independent analyses estimate detectors lag each major model release by three to nine months, and vendors ship version updates to close the gap.
Originality.ai's v4.5, released March 2026 specifically to address GPT-5 patterns, lifted its GPT-5 accuracy from about 60 to 78 percent, a vivid example of the catch-up cycle. Winston AI and others publish defensible, dated reports for exactly this reason: a score from six months ago describes a different model than the one shipping today, so documenting the vendor and date beside every result is the only honest way to compare.
What this means for you
If you are choosing a detector, do not assume one tool covers every model your writers might use. Check which models a vendor actually tests, prefer broad-coverage ensembles for institutional or classroom use, and treat any single score as one input, not a verdict, because false positives cluster on formal and non-native writing regardless of source model.
If you are checking your own work, run a free, low-false-positive self-check such as Scribbr before you submit, and remember that disclosure of AI use is a separate question from whether the text reads as machine-written. Detection measures statistical texture, not authorship, so a properly cited draft can still trip a detector and a hidden one can still pass.
Bottom line
Yes, detectors catch Claude and Gemini, but less reliably than ChatGPT: 2026 tests put GPT-5 at 85 to 92 percent, Gemini 2.5 at 80 to 88 percent, and Claude 4 the lowest at 72 to 84 percent. The gap tracks training exposure, closes within months of each model release, and matters most for institutions that must cover multiple models. Prefer broad-coverage tools like Originality.ai and Turnitin, and run a low-false-positive self-check such as Scribbr before submitting; never read a single score as proof of authorship.