Can detectors catch GPT-4o and Claude 3.5?
Newer models are harder to detect than older ones, and the gap keeps widening. Here is why GPT-4o and Claude 3.5 slip past detectors trained on GPT-3-era text, and what still catches them.
Updated October 31, 2026 ยท By the IsThisAI? team
Why newer models are harder to catch
Most AI detectors in wide use were trained on text from GPT-3.5-era models. They learned the statistical fingerprints of that generation: certain phrase choices, rhythm patterns, and predictability profiles. GPT-4o and Claude 3.5 write differently. Their word choices are less mechanical, their sentence variation wider, their tone more adaptable. Every model generation erodes the patterns the previous generation's detectors relied on.
This is not a flaw in any particular detector. It is structural. A classifier trained to spot 2023-era AI writing is fighting 2026-era AI writing with an outdated map.
What the benchmarks show
Independent testing has repeatedly found detection accuracy dropping as models advance. The pattern is consistent: high accuracy on the models the detector was trained against, declining accuracy on each subsequent generation. Detectors updated with newer training data recover some ground, which is why the detectors that retrain regularly outperform the ones frozen in 2023.
This is exactly why we recently added a ModernBERT-based detector trained on 2025-era AI text to our own ensemble. Fighting new models with old training data is a losing strategy, so we updated the training data.
Test it yourself: paste GPT-4o or Claude output into IsThisAI? and see what a modern detector makes of it.
Check your text free →What still gives new models away
Even the newest models have tells, though they are subtler:
- Default structure. Without specific prompting, both models still reach for the same essay scaffolding: balanced paragraphs, neat conclusions.
- Hedged certainty. Phrases like "it's worth noting" and "it's important to consider" survive across generations.
- Absence of lived detail. The models cannot have experiences. Text without specific, verifiable personal detail remains suspicious regardless of fluency.
- Prompt-shaped output. Text that answers a prompt a little too completely, covering every sub-question in order, reads as generated.
Where this is headed
The honest forecast: detection will keep working on casual, unedited AI use and keep failing on careful, edited, well-prompted use. Watermarking schemes could change the game if widely adopted, but none has stuck yet. For the foreseeable future, detectors are a filter for lazy misuse, not a lie detector for text.
Claude vs GPT: does the model matter?
Somewhat. Claude's default style is more conversational and less structured than GPT-4o's, which makes it slightly harder for detectors tuned on formal prose. But the difference is small compared to the difference prompting makes. A well-prompted GPT-4o beats a lazily prompted Claude on every detector we have tested. The prompt matters more than the model.
Why retraining is the whole game
A detector is a snapshot of AI writing at the time its training data was collected. As models evolve, the snapshot goes stale. The detectors that have held up best are the ones that retrain continuously on fresh model output. When evaluating any detector, ask when its models were last trained. A detector last updated in 2023 is fighting with one hand tied, no matter how good its architecture.
What about "humanizer" tools?
Tools that rewrite AI text to evade detection work, to a point, against older detectors. Against modern detectors trained on paraphrased and humanized text, their advantage shrinks. They also tend to degrade the writing: the output is often slightly off, with odd word choices that a careful reader notices even when the detector does not. They are an arms-race tactic with diminishing returns.
Bottom line
GPT-4o and Claude 3.5 beat most 2023-era detectors. Modern detectors trained on recent data do better but still struggle with edited output. If you rely on detection, keep your tools current and your expectations modest.
Sources
- Weber-Wulff et al., "Testing of detection tools for AI-generated text," International Journal for Educational Integrity 19, 26 (2023). https://link.springer.com/article/10.1007/s40979-023-00146-z
- Drayson et al., "AI Generated Text Detection using ModernBERT," arXiv:2502.15654 (2025). https://arxiv.org/abs/2502.15654
This guide is general information, not professional advice. AI detection results, including ours, can be wrong.