AI Writing Tells: New Quirks in Frontier Models

What happened

Graphite, a marketing analytics firm, conducted a study to identify linguistic patterns that distinguish AI-generated text from human-written content. The study tested four major AI models, revealing unique tells for each, such as specific words and phrases, and noted that while some well-known tells are being reduced, new ones emerge with each model update.

Why it matters

The findings are crucial for content authenticity, intellectual property verification, model governance, and competitive differentiation. They also provide insights for improving AI detection tools and understanding the evolving landscape of AI-generated text.

• AI models continue to evolve, producing new linguistic tells even as they reduce old ones.
• Model-specific tells can be used to differentiate AI-generated text from human writing.
• Detection tools must adapt to new patterns, using multi-layered approaches like lexical, phrasal, and syntactic analysis.
• The arms race between AI developers and detection engineers is shaping the future of content authenticity and model governance.
• The study provides actionable insights for improving AI detection and understanding model behavior.

The bigger picture

The study highlights the importance of detecting AI-generated text for safety, trust, and competitive advantage. It also influences model development, detection tool design, and industry practices around AI transparency and content authenticity.

The Graphite study underscores the ongoing evolution of AI-generated text and the need for continuous adaptation in detection methods. It highlights the importance of linguistic tells in distinguishing AI output from human writing and their implications for trust, compliance, and competitive advantage.

Technical details

• Control corpus: 10,000 pre-ChatGPT articles representing diverse journalistic styles.
• AI models rewrote summaries of control articles, preserving core facts.
• Over 13,000 phrases were flagged as appearing twice as often in AI-generated text.
• Claude Opus 5.5 uses 'dependable' 23× more frequently and phrases like 'this matters' 116× more often.
• Astra uses 'another dimension' and 'corrective framing' >100× more often than human text.
• Gemini 3.1 Pro nearly eliminated em-dashes through token-level punctuation filters.

FAQ