Archive — US — techcrunch
Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)
A study by Graphite found that while AI writing models have evolved, they still exhibit distinctive habits or 'tells' that differ from human prose. The study identified 13,000 phrases more prevalent in AI writing, with specific models like Opus 5.5 using phrases like 'this matters' significantly more than humans. Despite efforts, certain patterns persist, raising questions about the ability of developers to fully control narrative styles in AI-generated content.
Key points
- Graphite conducted research comparing AI-generated and human-written content.
- The study identified 13,000 phrases that occurred at least twice as often in AI writing.
- Opus 5.5's notable tells include overuse of 'this matters' (116 times more) and 'dependable' (23 times more).
- OpenAI's Astra uses phrases like 'another dimension' and 'may provide' more frequently.
- All models have significantly reduced the use of em-dashes in their writing.
- Despite the reduction in certain 'tells', the overall number remains constant, with new patterns emerging.
- Researchers are skeptical about the control labs have over these linguistic tells due to the complexity of LLMs.
Why it matters
- The findings align with ongoing conversations about AI writing authenticity and its implications for industries relying on content generation.
- Trends in AI continue to evolve, with large language models (LLMs) aiming for human-like writing; however, issues with unique tells persist.
- The emphasis on human-like clarity in AI writing models, such as from Anthropic and OpenAI, reflects broader trends in the IT industry towards user-friendly interfaces.
Graphite's study reveals that while AI writing models have made strides in mimicking human prose, they still reveal unique linguistic patterns. This ongoing challenge emphasizes the complexity of achieving truly natural writing in AI, suggesting that complete elimination of tells might remain elusive.