Does ChatGPT Watermark Its Text? (Maintained Reference)

Short answer: as of August 2026, there is no confirmed, deliberately deployed watermark in ChatGPT's text output. The famous April 2025 "watermark" was a training artifact that OpenAI fixed within days. Deliberate text watermarking systems exist — OpenAI has built one and chosen not to deploy it; Google has deployed one for Gemini — and image/audio provenance marking is now widespread. The details matter, so here is the full record with dates and sources.

This page is maintained and dated. We publish it alongside a free invisible-character inspector that runs in your browser. We do not sell an AI detector, a "humanizer," or anything else that would benefit from you believing a particular answer.

The April 2025 incident: U+202F narrow no-break spaces

In April 2025, users noticed that OpenAI's newly released o3 and o4-mini models inserted U+202F (narrow no-break space) characters between words in longer outputs — an invisible character that looks identical to a normal space. Because the character survives copy-paste, many concluded ChatGPT had started watermarking its text.

The documented sequence:

• April 20, 2025 — Rumi, an education-technology company, published the first widely-cited analysis, "New ChatGPT Models Seem to Leave Watermarks on Text," showing systematic U+202F insertion in o3/o4-mini output.
• April 22, 2025 — OpenAI told Rumi the characters were "a quirk of large-scale reinforcement learning" — an emergent habit picked up from training data (U+202F is standard in French typography and common on the web), not a deliberate watermark.
• By April 23, 2025 — the behavior was gone from the models' output.

Two things about this incident are worth keeping straight:

1. The characters were real, checkable facts — anyone could paste output into a character inspector and see them. That part was never in dispute.
2. "Watermark" was an inference. An intentional watermark that can be removed by find-and-replace would be a poor watermark; the training-artifact explanation fit the evidence and the rapid fix.

What OpenAI has actually built (and not shipped)

Separately from the 2025 incident, OpenAI confirmed (via 2024 reporting by the Wall Street Journal and OpenAI's own blog updates) that it had built a statistical text-watermarking system for ChatGPT — reported as roughly 99.9% effective at scale — and chose not to deploy it. Reported reasons included survey results (about 30% of users said they would use ChatGPT less if output were watermarked), fairness concerns for non-native English writers, and the ease of removal through paraphrasing or translation. OpenAI said it preferred exploring provenance metadata for media instead.

In November 2025, OpenAI also addressed the most famous visible "tell": ChatGPT's overuse of em dashes. Custom instructions asking the model to avoid them finally took effect, and newer models tone them down by default. The em dash was never a watermark either — it is ordinary punctuation that careful human writers have used for centuries.

What IS deliberately watermarked today (August 2026)

• Google Gemini text — Google's SynthID Text alters token selection during generation to embed a statistical signal. Detection requires Google's detector and is specific to SynthID; it is not a general "AI detector."
• Images — provenance marking is now mainstream: C2PA Content Credentials are signed at capture on recent Google Pixel and Samsung Galaxy devices, attached by DALL·E/ChatGPT image generation, Adobe tools, and Meta/TikTok labeling pipelines. SynthID also marks Google-generated images, audio, and video invisibly.
• Claude (Anthropic) — Anthropic's support documentation describes provenance marking for generated media and evolving practices for text; its stated approach has focused on C2PA-style credentials for supported media rather than a user-detectable text watermark.
• EU AI Act, Article 50 — as of August 2026, providers of generative AI systems in the EU are required to ensure AI-generated content is marked machine-readable — the regulatory driver behind much of the above.

What invisible characters prove — and don't

Finding hidden Unicode in a text tells you something true about the text's journey through software, not about its author:

• Common innocent sources: copying from web pages (zero-width spaces from markup), Word and Google Docs (no-break spaces, smart punctuation), PDFs (soft hyphens), spreadsheets (line separators), messaging apps.
• AI-related sources: interface formatting layers and occasional model quirks like the April 2025 incident.
• Deliberate uses: invisible fingerprinting of leaked documents, Unicode steganography (variation selectors and the tags block can encode arbitrary bytes), and hidden prompt-injection payloads aimed at AI agents.

And the converse: clean text proves nothing about human authorship. Removing invisible characters does not defeat AI-text detectors either — those rely on statistical properties of the writing, not hidden bytes. Any tool that promises "undetectable" output by stripping characters is selling you a false model of how detection works.

Check your own text

Our invisible character inspector shows every hidden character in pasted text — code point, Unicode name, category, and likely origin — and cleans them with protections for emoji and non-Latin scripts. It runs entirely in your browser; nothing is uploaded.

Sources

• Rumi, "New ChatGPT Models Seem to Leave Watermarks on Text" (April 20, 2025) — rumidocs.com
• OpenAI statement to Rumi on the RL training quirk (April 22, 2025)
• Wall Street Journal, "There's a Tool to Catch Students Cheating With ChatGPT. OpenAI Hasn't Released It" (August 2024)
• OpenAI, "Understanding the source of what we see and hear online" (blog, updated 2024)
• TechCrunch, "OpenAI fixed ChatGPT's em-dash problem" (November 14, 2025)
• Google DeepMind, SynthID overview — deepmind.google/models/synthid
• C2PA specification — spec.c2pa.org; Content Credentials verify tool — contentcredentials.org/verify
• Anthropic support, "How Claude marks AI-generated content" — support.claude.com
• EU AI Act, Article 50 transparency obligations (application from August 2026)

Changelog

• 2026-08-11 — Page created: April 2025 U+202F timeline, OpenAI's shelved watermark, November 2025 em-dash fix, SynthID/C2PA status, EU AI Act Article 50 note.