Claude writes carefully. Sonnet, Opus, and Haiku share the same fingerprint: thoughtful framing, polite hedging openings, structured bullet headers, parenthetical qualifications, and a heavier em-dash cadence than other large language models. TextSight runs a Claude-aware AI rewriter in three modes (Light, Balanced, Maximum) with closed-loop detector calibration on every pass. Free 1,500-word preview, no card required.
The same AI rewriter engine handles every major model, but the edit pass on Claude prose is different from the ChatGPT pass. Claude leaves a measured, structured fingerprint, and that fingerprint needs different rewrites.
The AI rewriter is not a paraphrase. It is a calibrated rewrite tuned against the patterns Claude leaves in its prose. The classifier reads which model produced the text from the prose alone, then routes the rewrite through the matching edit set. For Claude that means three things: rebalancing the hedge density, demoting the structured bullet pattern where prose reads better, and trimming the parenthetical clusters Claude is fond of.
A ChatGPT-tuned AI rewriter aimed at "delve, robust, leverage" vocabulary misses the Claude tells almost entirely. Claude rarely uses that vocabulary. Its tells live in structure and cadence: bullet headers with bold labels, "I'd like to," "I should note," "To be clear" openings, inline qualifications in parentheses, and a higher em-dash density than ChatGPT. A rewrite engine that does not know to look for those patterns leaves the output reading like Claude with slightly less filler.
Every AI rewriter pass runs against the same detector that scored the input. The rewrite is re-scored as it goes, and a Light pass that does not move the score retries automatically with adjusted weights. The detector and AI rewriter share one classifier signal set, so an output that satisfies one side satisfies the other. This is the coupling that prevents pretty-looking rewrites that still flag detectors.
The same three intensities work across every model, but their effect on Claude prose is specific. Picking the right one matters more than people realise.
Strips Claude's hedging openings and warm-up sentences. Keeps every factual claim, every code block, every named process. Score gain on typical Claude prose: 25 to 40 points. The safest mode for engineering writing, API documentation, internal memos, and any text where Claude's precision is the reason you used it.
Mid-depth rewrite. Cuts parentheticals down to roughly one per paragraph, demotes the bold-labelled structured lists into prose where prose reads better, and rebalances the em-dash and semicolon mix. Score gain: 40 to 55 points. Right for professional emails, policy drafts, internal communications, and most non-academic prose written in Claude.
Aggressive rephrasing. Will rework sub-claims and can shift emphasis on factual statements. The output surfaces a claims-shifted warning when the rewrite touches a factual claim so you can review before publishing. Score gain: 55 to 75 points. Suitable for social posts, blog intros, marketing copy, and anywhere Claude's measured precision is not load-bearing.
A useful Claude-specific default: Light on technical content, Balanced on professional writing, Maximum only on casual copy. Picking Maximum for an engineering blog is the most common mistake on this page's heat map.
Free tier handles a short essay or a handful of short emails. Paid tiers add volume, the Chrome extension, and the REST API. Full breakdown on the pricing page.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing →
Five Claude-specific tells the rewrite engine targets. None of these are present in raw ChatGPT output at the same density, which is why a ChatGPT-tuned AI rewriter misses them.
Claude defaults to numbered or bulleted lists where each item has a bold label, then a colon, then explanation. ChatGPT favours plain bullets without bold labels. The bold-label structured format is one of the clearest Claude signatures in 2026, so the AI rewriter demotes it to plain prose or to bullets without labels depending on context.
"I'd like to think about this," "I should note," "I'd be happy to," "To be clear," "Let me consider," "Actually," and "I'd suggest" are openings Claude uses to qualify its answers. ChatGPT rarely opens with first-person hedges. The AI rewriter strips these and rebuilds the opening as a direct claim or a question, depending on the surrounding context.
Claude inserts inline qualifications in parentheses two or three times per paragraph on average. Human writers use parentheticals roughly once every two paragraphs. The AI rewriter promotes some parentheticals to full sentences (where the qualification adds genuine information) and drops the ones that add nothing.
Claude uses em-dashes more than other large language models, often three or four within a few hundred words. The AI rewriter rebalances the punctuation: some em-dashes become commas or colons, some become full stops, and the surrounding sentences are reshaped so the prose does not feel choppy. The detector reads em-dash density as a model-attribution signal, so this is one of the higher-leverage edits.
"Let me think about this," "I want to flag a concern," "It is worth being careful here." Claude narrates its own reasoning in a way that human writers rarely do outside of explicit essay forms. The AI rewriter drops the meta-reasoning by default and reshapes the underlying claim so the prose moves rather than narrates.
All three Anthropic tiers share the same core fingerprint. Sonnet is the canonical Claude voice and the working default for the Balanced mode. Opus drafts longer paragraphs, richer chains of reasoning, and more nested parentheticals, so the same edit set runs a bit heavier on its output. Haiku produces tighter, shorter responses with fewer parentheticals and shorter structured lists, where Light mode is often sufficient. The humanizer reads the prose itself rather than asking you to pick a tier.
How TextSight frames the AI rewriter and where it sits relative to academic integrity policy.
The AI rewriter is positioned as a voice and calibration tool. The intended uses are: rewriting authored drafts so they read in the writer's own voice, fixing false positives where formal English overlaps with AI patterns, and removing repetitive structural tells from working drafts that a writer started with Claude and is finishing themselves.
Submitting AI-generated work as your own remains a policy violation at most institutions regardless of how it reads on detector output. The AI rewriter does not change that. What it does is help writers shape voice on prose they authored, and surface where their own prose accidentally lands in AI patterns so they can rewrite intentionally rather than guess.
For professional contexts (marketing copy, emails, blog drafts) the framing is simpler: take a Claude starting point, route it through authenticity, ship prose that reads like you. The AI rewriter is one step in a writing workflow, not a finishing step that licenses anything.
Every rewrite runs against the same detector that scored the input. That coupling is what keeps the output measurable rather than wishful.
Most AI rewriters on the market are paraphrase engines with no detector attached. They produce output that looks fine to a human reader but still flags every detector on the market because no one is checking. TextSight built the AI rewriter and the detector on the same classifier signal set, so a rewrite that satisfies the AI rewriter satisfies the detector by construction.
In practice that means a Light pass that does not move the Authenticity Score retries automatically with adjusted weights, up to the target threshold. The output ships with a fresh Authenticity Score on every pass so writers can see the move rather than trust marketing. When the score is high, it is high because the same classifier that scored the input is scoring the output, not because a paraphrase engine looks superficially different.
This is the structural reason TextSight ships detector and AI rewriter in one workflow rather than separately. Decoupling the two is how marketing-grade AI rewriters get away with claims that do not survive contact with real detection tooling.
Paste flow for one-off rewrites, Chrome extension for in-place editing, REST API for content pipelines.
The dashboard at app.textsight.ai. Paste Claude output, pick a mode, get the rewrite with a side-by-side score view. Right surface for one-off authenticity, for testing the modes, and for short essays. Available on every tier including the free plan.
Surfaces the AI rewriter inside Gmail, Google Docs, Notion, LinkedIn, and any contenteditable surface. Same three modes, same scoring, no copy-paste round trip. Right surface for in-place editing of professional emails and documents. Available on Pro and Business.
For teams routing Claude output through a content pipeline before publishing. Detect, rewrite, re-score, then ship. Available on Business with documented endpoints and a steady throughput target. Right surface for agencies that have standardised on Claude internally and want authenticity as a build step.
Detect Anthropic Sonnet, Opus, and Haiku output with sentence-level highlights.
See the detector →The ChatGPT-specific edit pass for the other half of the model market.
See the AI rewriter →The cross-model AI rewriter landing page with all three modes explained.
Read the overview →Free, Pro, and Business tiers side by side. Yearly billing saves 25%.
See pricing →No card. Three modes tuned for Anthropic's voice. Closed-loop detector calibration on every pass.