how do you spot Claude writing?

Claude is Anthropic's family of frontier text and multimodal models, the second-most-used LLM after ChatGPT by consumer volume and the default in coding-assistant contexts (Claude Code, Cursor, and Windsurf all default to Claude in their best modes). as of mid-June 2026 the flagship is Claude Opus 4.8, released May 28, 2026. Anthropic operates with a consistent Haiku / Sonnet / Opus naming convention introduced with Claude 3: small / mid / flagship at every generation.

enterprise + API revenue is closer to parity with OpenAI than the consumer headline numbers suggest.

what Claude writing looks like

Claude has a distinct register. heavier use of em-dashes and en-dashes than most human writers. characteristic hedge phrases (“it's worth noting,” “that said,” “here's the thing”). structured headers and bullet points by default. rare use of contractions in formal mode. lists that pivot between parallel and non-parallel forms.

compared to GPT-4 / GPT-5 output, Claude's prose tends to be more cautious, qualifies claims more readily, and reaches for a metaphor where GPT would reach for a numbered list. it's also more likely to refuse hard questions or push back on uncomfortable framings, a stylistic tell when it appears in copy where the author would have answered.

how amige. detects Claude

text detectors use the same statistical signals on Claude that they use on every LLM: perplexity (Claude has uniformly low next-token surprisal compared to humans), burstiness (variance of perplexity across the document, also low), and cross-model perplexity ratios from techniques like Binoculars.

amige. routes the scan to the detectors strongest for this kind of text, fuses and calibrates their reads, and caps the confidence. on longer outputs the per-model attribution will often surface “Claude” as the top guess. on short outputs (under ~150 words), distinguishing Claude from GPT-class output is harder: statistical signal is too thin and stylistic tells haven't had room to surface, so amige. leans toward “uncertain” rather than guess.

detection difficulty over time

Claude has gotten harder to detect with each version. Claude 2 was a giveaway from its formality and hedging density. Claude 3.5 Sonnet was the first version that detection vendors said closed the gap on humans for the average case. short outputs from Claude 4-class models are routinely undetectable in independent tests.

the “agentic” / reasoning mode that mixes tool calls and chain-of-thought scaffolding into the output stream changes the surface form enough that some detectors regress. when one classifier in the panel is confident and another isn't, treat the split as informative rather than inconclusive.

for an amige. user: if amige. flags text as Claude, it likely passed your own “feels human” sniff test. the detection is statistical, not stylistic, so don't rely on “it doesn't sound like AI” as a counter. treat the verdict as “this text has the linguistic fingerprint of Anthropic's models,” which matters most in school, hiring, and journalism contexts.

controversy and context

less public scandal than OpenAI by a wide margin. Anthropic ships with a more cautious safety posture, more refusals on ambiguous queries, and less viral-deepfake exposure (Claude is not an image model).

notable, per publicly reported filings: music publishers sued Anthropic over training on song lyrics (filed 2023), and authors sued over training on copyrighted books (filed 2024). the authors' case (Bartz v. Anthropic) reached a settlement in 2025. that exposure is less existential for the business than the Stable Diffusion lawsuits are for Stability AI, but a real cost center that has shaped Anthropic's data-licensing strategy.

more on the underlying statistical signals in perplexity and burstiness, and how amige. routes, fuses, and calibrates all of it in the machine.