Claude Opus 5.5 uses 95% fewer em dashes, but its answers are getting longer

Refract AI Intelligence Digest

BLUF

New AI models are adapting their output to evade detection patterns, complicating content authenticity verification.

NEWS

Analysis of Claude Opus 5.5 reveals a significant reduction in stylistic markers like em dashes compared to previous versions, alongside simpler vocabulary and shorter sentences. Despite these stylistic changes, the model generates longer responses overall. This trend indicates ongoing optimization to mimic human writing patterns more closely.

Why I Care

Security teams and content moderators face increased challenges in identifying AI-generated disinformation or phishing attempts as detection heuristics become less reliable. Organizations relying on AI content filters may experience higher false negatives, impacting trust and safety protocols.

Next Steps

Security leaders should update AI detection policies by Q4 2026 to account for evolving linguistic patterns. Teams must implement multi-layered verification methods rather than relying solely on stylistic analysis for threat assessment.

Anthropic's Claude Opus 5.5 appears to be changing how it writes, with new analysis showing fewer obvious AI writing patterns, shorter sentences, and simpler wording compared with Opus 5. [...]
Back to Blog Listing

Source: BleepingComputer ·

This digest was generated by Refract AI Collective to help the public sector security community stay informed.