Five AI Writing Detectors Tested: Most Work, Two Fail Badly

As AI-generated content proliferates across academic, professional, and social environments, the question of whether text was written by a human or machine has become increasingly urgent. Multiple commercial tools promise to identify AI authorship, but their actual performance varies dramatically.

A systematic test of five popular AI detection platforms—Pangram, Grammarly, GPTZero, Scribbr, and Copyleaks—reveals significant differences in accuracy. The evaluation used human-written article introductions alongside 150-word samples generated by ChatGPT, Gemini, and Claude based on identical prompts.

Why it matters

Organizations relying on AI detectors for academic integrity, hiring decisions, or content verification need to understand these tools' limitations. With detection accuracy ranging from 50 percent to 100 percent across platforms, choosing the wrong tool could lead to false accusations or missed AI content. The results suggest no single detector should be trusted in isolation for high-stakes decisions.

The top performers

Three platforms achieved perfect accuracy across all test samples. Pangram correctly identified all human-written text as 100 percent human and all AI-generated samples as 100 percent machine-written, displaying high confidence in each assessment. The tool even highlighted specific AI tells, such as the phrase "from the moment you."

GPTZero matched this performance, marking human samples with the declaration "We are highly confident this text is entirely human" and correctly flagging AI-generated content. The platform went further by identifying which specific sentences appeared most AI-like, though the patterns weren't always obvious to human reviewers.

Grammarly also achieved 100 percent accuracy, though with slightly less certainty. While human samples scored zero percent for AI characteristics, the AI-generated texts from Claude and Gemini registered as 68 percent and 66 percent AI-written respectively—correct identifications, but with less conviction than competitors.

The failures

Two platforms performed poorly. Scribbr, despite offering its detection service completely free without registration, correctly identified human writing but marked AI-generated samples from ChatGPT and Claude as fully human-written. The tool displayed high confidence in these incorrect assessments.

Copyleaks achieved only 50 percent accuracy on AI-generated content. While correctly flagging a Gemini sample as 100 percent AI-written, it rated a Claude sample as zero percent AI—a complete miss.

Practical implications

Every detector successfully identified human-written text, suggesting these tools may be better at confirming authentic human authorship than catching AI content. For organizations requiring reliable AI detection, the testing suggests using multiple checkers simultaneously rather than relying on a single platform.

Pricing varies considerably. Grammarly offers the lowest entry point at $12 monthly, while GPTZero starts at $23.99. Pangram begins at $20 per month. Scribbr remains free, though its poor performance on AI detection undermines that advantage.

These findings were first reported by Popular Science, which conducted the original testing.