Opus 5.5 uses the word 'dependable' 23 times more than humans

By Splitfeed EditorialLast updated Oct 1, 2026, 19:20 UTC

On October 1, 2026, an analysis identified statistical word-frequency patterns that distinguish output from Anthropic's Opus 5.5 model from human-written text. The most striking finding is that Opus 5.5 uses the word "dependable" at a rate 23 times higher than in comparable human writing samples. The phrase "this matters" was also flagged as a recurring tell specific to the model's output. The methodology focuses on raw frequency counts rather than meaning, grammar, or sentence structure, representing a different approach to AI detection than semantic or stylistic analysis.

The findings sit within a broader and increasingly urgent effort to develop reliable methods for identifying AI-generated text as large language models become more fluent and harder to detect through casual reading. Word-frequency fingerprinting, if it holds up to scrutiny, could offer a lightweight and scalable detection signal. What remains genuinely in dispute is how durable such signals are: Anthropic or any model developer could potentially adjust training to flatten these quirks once they are publicly known, and it is not clear whether the patterns persist across different prompting styles, languages, or task types. Whether Anthropic has responded to or disputed the methodology has not been reported.

With no published reactions from any camp at this stage, it is possible to sketch the expected fault lines. Pro-AI voices would likely argue that surface word-frequency anomalies are trivial, that they do not reflect any meaningful deception, and that models can be tuned away from such patterns quickly. Anti-AI commentators would probably treat the findings as evidence of the difficulty in trusting AI-generated content at face value, and as a reason to demand stronger disclosure requirements. Those in the middle ground would typically frame this as a useful but limited tool, one that works only until developers adapt, and would call for more robust and standardised detection frameworks.

The argument will sharpen if Anthropic responds to the methodology or if follow-up analysis tests whether the patterns survive prompt variation or a model update. Any public statement from Anthropic, or a peer review of the detection method, would clarify how seriously this particular fingerprint should be taken as a detection signal.