Can AI Detectors Detect GPT-5 and Claude 4?

Every new generation of language models prompts the same question: can detectors still identify AI-generated text? With the release of GPT-5, Claude 4, Gemini 2.0, and other advanced models in late 2025 and early 2026, this question has renewed urgency. The answer is nuanced: detection remains possible but has become more challenging, and the specific capabilities vary significantly across detection tools and content types.

How New Models Challenge Detection

Each generation of language models produces text that is statistically closer to human writing. GPT-5 and Claude 4 show higher perplexity variance, more natural burstiness patterns, and more diverse vocabulary distributions than their predecessors. These improvements narrow the statistical gap between human and AI text that detection tools exploit, making discrimination more difficult.

The latest models also demonstrate improved ability to follow style instructions, producing text that matches specific writing styles, registers, and tonal qualities more faithfully. This makes it harder for detection tools that rely on identifying a generic AI writing style, because the style can now be customized to match virtually any target voice or format.

However, fundamental differences between machine generation and human cognition persist. Even the most advanced models still optimize for statistical likelihood, creating subtle patterns that differ from the genuine unpredictability of human thought. The gap has narrowed but has not closed, and there are theoretical arguments that it cannot fully close given the fundamentally different processes underlying human writing and machine generation.

Detection Accuracy by Model

Testing across the major detection tools reveals a consistent pattern: detection accuracy decreases with each model generation but remains practically useful. For GPT-5 output at 500+ words, detection accuracy ranges from 83% to 90% across top tools, compared to 90-95% for GPT-4 and 95-98% for GPT-3.5. Claude 4 detection shows similar patterns, with accuracy ranging from 81% to 88%.

Gemini 2.0 presents an interesting case. Its output shows distinctive structural patterns that some detection tools exploit effectively, achieving accuracy above 90% despite the model's general sophistication. This suggests that detection difficulty is not purely a function of model capability but also reflects specific architectural choices and training approaches that create different statistical signatures.

EyeSift's detection engine has been updated to account for the characteristics of the latest models, maintaining accuracy above 85% on GPT-5 and Claude 4 output at adequate text lengths. The multi-method approach, combining statistical analysis with trained classifiers, provides resilience against the improvements in any single dimension that new models achieve.

The Text Length Factor

Text length has become even more important with the latest models. At 50-100 words, detection of GPT-5 and Claude 4 output is only marginally better than random guessing for most tools. At 250 words, accuracy reaches practically useful levels of 75-82%. At 500+ words, the full capability of detection tools can be leveraged. This increasing importance of text length reflects the fact that subtle statistical differences require more data to reliably detect as those differences become smaller.

For practical purposes, this means detection tools provide the most value for longer content: essays, articles, reports, and extended communications. Short-form content like social media posts, brief emails, and comment sections remain challenging to analyze reliably. Users should adjust their confidence in detection results based on the amount of text analyzed.

Detection Tool Adaptation

Detection tools have responded to new model releases with improved techniques. Training data is updated to include output from the latest models. New features targeting specific characteristics of new models are developed. Ensemble approaches are refined to weight different detection signals based on their effectiveness against current-generation output.

The response time matters. When a new model is released, there is typically a window of reduced detection accuracy before tools are updated. This window has shortened with each generation as detection providers have built more agile update pipelines. For GPT-5, major detection tools updated within two to three weeks of release, compared to months for earlier model generations.

The Bigger Picture

The detection-generation arms race is often framed as a competition that detection must eventually lose. This framing is misleading. Detection does not need to achieve perfection to be valuable. An 85% accurate tool provides enormous practical value for screening, deterrence, and risk assessment. The relevant question is not whether detection can achieve 100% accuracy but whether it provides sufficient signal to improve decision-making in its intended context.

Furthermore, detection is evolving beyond pure text analysis. Behavioral signals, writing process analytics, content provenance systems, and multimodal analysis all provide complementary detection capabilities that are not directly affected by improvements in text generation quality. The future of detection is not a single method keeping pace with a single generation model but an ecosystem of verification approaches that together provide robust content authenticity assessment.

The latest models are harder to detect than their predecessors, but they are detectable. Detection tools continue to provide practical value for organizations and individuals who need to verify content authenticity. The key is understanding the current capabilities and limitations, using tools appropriately, and integrating detection into broader verification workflows rather than relying on it as a standalone solution.