Is it watermarked?
ChatGPT Watermark: Is Anything Actually Watermarked?
Three different things get called the ChatGPT watermark, and since October 5, 2026 all three exist somewhere. Since May 19, 2026, images generated by ChatGPT have carried Google DeepMind's SynthID pixel watermark stacked on top of OpenAI's own C2PA metadata. Audio got SynthID on July 31, 2026, though without any C2PA manifest. Text was the exception until 2026-10-05: OpenAI had built a text watermark by 2024, held it back, and explained why. On October 5, 2026 it announced textGrain, an invisible text watermark rolling out over the coming weeks to ChatGPT and Codex in the EU only, with an opt-in for API customers worldwide that stays off by default. Nobody outside a group of approved researchers can check text for it.
Correction,
The verdict box, a takeaway, the images section and two answers said images from the OpenAI API carry C2PA plus SynthID, and one answer said images are marked everywhere. OpenAI's help page, read 2026-10-06, excludes gpt-image-1 and gpt-image-1.5 from API coverage, so API images from those models should not be assumed to carry either signal; the page now says so wherever API images are mentioned. The shelved-watermark section also said non-native English phrasing looks more machine-like to watermark detectors. A keyed watermark detector does not respond to writing style; the 2024 concern was that watermarking would stigmatise non-native speakers who use ChatGPT as a writing aid, and the page now says that.
Correction,
This page absorbed the site's two other OpenAI pages, the head-term evidence file and the removal query, which split one subject across three URLs. Grafting their facts in surfaced three claims all three pages had wrong. All three said OpenAI marks audio the way it marks images, with C2PA metadata plus SynthID. It does not: OpenAI attaches no C2PA manifest to audio, and this site's own status database recorded that on 2026-09-03. All three also dated OpenAI's C2PA steering-committee membership to the May 2026 image rollout; OpenAI's own page places the membership in 2024 and describes only its "C2PA Conforming Generator Product" status as recent. And all three named ChatGPT Voice and GPT-Live as the watermarked audio surfaces; OpenAI's own page says "ChatGPT and the OpenAI API" and names neither product. The verdict box, the takeaways and the audio section now match the primary sources.
Correction,
The answer box graded Sora output as confirmed for both C2PA metadata and SynthID. That contradicted this site's own evidence: the full OpenAI file records a November 2025 audit that found visibly watermarked Sora 2 videos carrying no detectable C2PA metadata, and unwatermarked Pro-tier videos that did carry it, and the provider table already marked Sora video as inconsistent. The box now separates images and supported audio, which are confirmed, from Sora video, which is contested. The same box also offered media verification as confirmed without the caveat stated further down this page, that any upload can strip the manifest and a clean result proves nothing; it now says verification holds only on an unmodified original. The shelved text watermark carried three different evidence labels on this one page and is now labelled Reported, matching the rest of the corpus.
Short answer
- ChatGPT text
- Partial: textGrain in EU ChatGPT and Codex (announced 2026-10-05, rolling out over the coming weeks), API opt-in worldwide, off by default; no public detector
- ChatGPT text outside the EU
- No, not at launch
- ChatGPT and API images
- Yes, C2PA plus SynthID, except gpt-image-1 and gpt-image-1.5 images through the API
- ChatGPT and API audio
- Yes, SynthID only, no C2PA
- Sora video
- Signals inconsistent in testing
- Can you verify media?
- Only on an unmodified original
- Can you verify text?
- No, the detector is limited to approved researchers
- Is there a text watermark to remove?
- On EU and opt-in API text, yes; paraphrase and translation weaken it
- An AI detector saying 87%
- Not watermark evidence
OpenAI has documented provenance on generated media for years and, until 2026-10-05, described text provenance as a goal rather than a shipped feature. textGrain changes that for EU ChatGPT and Codex text and for API customers who opt in. It does not change what an ordinary user can check: the detector is not public, and OpenAI says the absence of a detected watermark does not prove a human wrote the text.
Key takeaways
- New on 2026-10-05: OpenAI announced textGrain, its first shipped text watermark. It is an invisible change to the randomness in the model's word choices, rolling out over the coming weeks to eligible ChatGPT and Codex text in the EU across all plans. OpenAI says it is not making text watermarking a global default at launch.
- API customers anywhere can opt in for select models through Organization settings > Data controls > Text provenance. It stays off by default, and enabling it does not give the customer access to the detector.
- The detector is not public. OpenAI is limiting it at first to approved researchers and expert organizations, case by case, and says it reports only whether an OpenAI watermark is detected, without identifying the user or revealing prompts or conversations.
- OpenAI's own numbers, at a 1% target false positive rate: about 80% of 200-token passages and about 95% of 400-token passages detected. Replacing 10% of words with synonyms cut detection of 400-token passages from about 92% to 66%, and replacing 25% cut it to 17%. Substantial paraphrasing or translation can make the watermark undetectable.
- Confirmed: since May 19, 2026, images from ChatGPT, the OpenAI API, and Codex carry both C2PA Content Credentials metadata and Google DeepMind's SynthID pixel watermark, checkable through OpenAI's Verify tool, except images from gpt-image-1 and gpt-image-1.5 through the API, which OpenAI's help page excludes and which should not be assumed to carry either signal.
- Confirmed, with one correction: audio from ChatGPT and the OpenAI API got SynthID on July 31, 2026, two days before EU AI Act Article 50 took effect, alongside a Content Provenance API. Audio carries SynthID only. OpenAI attaches no C2PA manifest to it, so an audio file has no metadata layer to lose.
- History, reported: by 2024 OpenAI had built an internal ChatGPT text-watermarking method reportedly around 99.9% effective and held it back, citing circumvention risk, a concern that it would stigmatise non-native English writers who use ChatGPT as a writing aid, and internal survey data suggesting roughly 30% of users would use ChatGPT less if it shipped. OpenAI's own published posts made the same arguments. That history ended with textGrain on 2026-10-05.
- Contested: OpenAI says every Sora video carries both visible and invisible provenance signals. A November 2025 independent investigation found the reverse in places, and video is the one modality OpenAI gives nobody any way to check.
- Removal, honestly: outside the EU, ChatGPT text carries no watermark to remove. Inside the EU, OpenAI's own figures show synonym swaps, paraphrasing and translation weaken or erase textGrain, while copy and paste does not touch it. Image C2PA metadata comes off by accident on any ordinary upload, and the only reported way to defeat SynthID on an image was regenerating the whole image through a different model at a real cost to quality.
- Community discussion, unresolved: a June 28, 2026 OpenAI Developer Community thread found that a PNG's valid C2PA metadata gets stripped down to bare file chunks the moment it is re-uploaded into a ChatGPT conversation, even though the original file still verifies fine. No OpenAI staff response was found.
- Hidden Unicode characters, non-breaking spaces, repeated phrases, and writing-style habits get called ChatGPT watermarks constantly online. OpenAI has never named any of them as a mechanism, and says textGrain inserts no hidden characters or invisible spaces.
Signal breakdown
OpenAI: what carries a mark, and what doesn't
The clearest split between modalities of any provider: images carry two independent provenance layers, audio carries SynthID alone, and text is marked where an API customer opts in, with EU ChatGPT and Codex text rolling out from 2026-10-05.
- Chat and API textConditional
textGrain, announced 2026-10-05: EU ChatGPT and Codex text over the coming weeks; global API opt-in, off by default
- ImagesMarked
C2PA Content Credentials since 2024, plus a SynthID watermark since 2026-05-19. API coverage excludes gpt-image-1 and gpt-image-1.5, so their API images should not be assumed to carry either
- AudioMarked
SynthID, since 2026-07-31 (two days before Article 50 applied)
- Sora videoContested
OpenAI lists Content Credentials only; independent testing found them inconsistent; Sora is discontinued
Detector: Public for images and audio: the Verify tool, now at openai.com/research/verify/, and a Content Provenance API accept those two. The textGrain text detector is limited to approved researchers, by application.
Every state above traces to a primary source with a verification date in the status database, and the provider detail is the rest of this page.
Removal reality check
OpenAI: what is actually there to remove
What exists to remove
- Chat and API text. textGrain, announced 2026-10-05: EU ChatGPT and Codex text over the coming weeks; global API opt-in, off by default
- Images. C2PA Content Credentials since 2024, plus a SynthID watermark since 2026-05-19. API coverage excludes gpt-image-1 and gpt-image-1.5, so their API images should not be assumed to carry either
- Audio. SynthID, since 2026-07-31 (two days before Article 50 applied)
What can be verified
Public for images and audio: the Verify tool, now at openai.com/research/verify/, and a Content Provenance API accept those two. The textGrain text detector is limited to approved researchers, by application.
For an image or an audio file you can actually run the check. For text you cannot: unless you are an approved researcher there is no detector to run, and OpenAI says the absence of a detected watermark does not prove human authorship.
What are the three layers people call the ChatGPT watermark?
ConfirmedSort out which one you are asking about first. Two of the three are on every image or audio file, they break in opposite ways, and the third, for text, applies only in the EU or where an API customer opted in.
| Layer | What is documented |
|---|---|
| C2PA Content Credentials metadata | documented by OpenAI, attached to generated images, and the most fragile layer by far. |
| SynthID | documented for images since May 19, 2026 and for audio since July 31, 2026. It lives in the pixels or the waveform, not in a metadata field. |
| textGrain, the text watermark in chat output | announced 2026-10-05: an invisible change to word-choice randomness in ChatGPT and Codex text in the EU, and in API text where the customer opts in. Until 2026-10-05 no OpenAI text watermark was live; an earlier system was built and shelved. |
The first two are media layers and do not fail the same way. Metadata can be lost to format conversion, re-saving, screenshotting, or a plain re-upload, gone in a single step.
Losing the metadata tells you nothing about whether SynthID is still present. SynthID is embedded in the content itself and is designed to survive common transformations better than metadata does.
That difference is the whole reason a metadata-stripping tool is not a watermark remover, and the reason OpenAI keeps repeating that "no detection method is foolproof."
Are ChatGPT images watermarked?
ConfirmedTwo provenance layers on every ChatGPT image, built by two rival companies, plus the one routine habit that erases your ability to check either.
OpenAI says images generated with ChatGPT, Codex, and the OpenAI API include C2PA Content Credentials metadata and a SynthID watermark, verifiable through a public tool at openai.com/research/verify/ (OpenAI's help page now links it as openai.com/verify). OpenAI's help page, read 2026-10-06, excludes models "planned for imminent deprecation (gpt-image-1, gpt-image-1.5, sora-2)" from API coverage, so API images from gpt-image-1 or gpt-image-1.5 should not be assumed to carry either signal. The rollout is dated May 19, 2026, and OpenAI describes reaching "C2PA Conforming Generator Product" status as the recent part of it.
Its C2PA steering-committee seat is older than that. OpenAI's own page places the membership in 2024, alongside adding Content Credentials to DALL-E 3, so the 2026 announcement widened what gets marked rather than starting the relationship.
The detail worth pausing on is whose watermark it is. SynthID is Google DeepMind's, so two competing labs now run the same pixel-level signal on their own output.
OpenAI describes the two layers as complementary: "C2PA helps content carry detailed context, SynthID helps preserve a signal when metadata does not."
OpenAI's own caveats deserve quoting rather than glossing.
- "No detection method is foolproof."
- C2PA metadata "can be stripped, lost through uploads and downloads, or broken by transformations."
- Provenance signals are "not a guarantee that content is accurate, unedited, legally owned, or presented in the correct context."
So a result of no signal detected does not mean the image was not made by ChatGPT. It can simply mean the metadata did not survive whatever happened to the file after it left OpenAI's servers.
The Verify tool accepts images and audio only. It deliberately stops short of a definitive conclusion when it finds nothing, for the same reason.
Is ChatGPT audio watermarked?
ConfirmedAudio got a watermark on July 31, 2026, three weeks after a system card that never mentioned it. It did not get the metadata layer images have.
On July 31, 2026, two days before EU AI Act Article 50 transparency rules took effect, OpenAI extended SynthID watermarking to supported audio from ChatGPT and the OpenAI API. It added a Content Provenance API the same day, POST /v1/content_provenance_checks, which accepts images and audio up to 50 MiB with audio capped at 60 seconds.
Audio did not get the treatment images got, and it is worth being precise about this because the shorthand "same approach" is wrong. OpenAI marks audio with SynthID and attaches no C2PA Content Credentials to it. Two of its own surfaces agree, both listed below: the July 2026 provenance update says supported audio "now includes SynthID watermarking" and never mentions C2PA, and the help-center modality table reads "Audio | SynthID watermarks."
That has a practical upside. An audio file has no manifest to strip and no issuer to read, so the fragile-metadata caveat that dominates the image story simply does not apply to it.
GPT-Live's own system card, published three weeks earlier on July 8, 2026, makes zero mention of SynthID, C2PA, or watermarking. The audio provenance rollout was announced separately from, and after, the model's own safety documentation rather than built into it from the start.
Does every Sora video carry a watermark?
Community discussionOpenAI says every Sora video carries visible and invisible signals. One tester found the opposite pattern, and no OpenAI tool accepts video to settle it.
A November 2025 investigation by Ethan Le Sage, published on LessWrong and cross-posted to the EA Forum, checked Sora 2 videos against both the Content Credentials Verify tool and the open-source c2pa-rs CLI. Some videos carrying the visible OpenAI watermark had no detectable C2PA metadata at all, while some Pro-tier videos without the visible watermark did carry it.
That is close to the reverse of OpenAI's public claim that every video generated with Sora includes both visible and invisible provenance signals. Rechecked on September 3 and again on September 25, 2026, the post is unedited, carries no author correction, and its three comments are all from November 2025, none of them from OpenAI.
Two limits keep this at community discussion rather than a confirmed failure. It rests on a single independent tester using third-party tooling, and SynthID is a separate pixel-level mechanism that neither tool measures, so an absent C2PA manifest does not establish an absent watermark.
OpenAI's help page, re-read on 2026-10-06, lists video as carrying Content Credentials only and notes that Sora is discontinued.
Is ChatGPT text watermarked now?
Official announcementIn the EU, it is being added over the weeks after OpenAI's 2026-10-05 announcement. Elsewhere, no, unless the text came through an API account that opted in.
Partly. On October 5, 2026 OpenAI announced textGrain, an invisible change to the model's word choices, and it is rolling out over the coming weeks to eligible ChatGPT and Codex text in the EU across all plans. API customers anywhere can opt in for select models; it is off by default. The detector is not public, and OpenAI says the absence of a detected watermark does not prove human authorship.
| Where the text comes from | Watermarked? |
|---|---|
| ChatGPT or Codex, EU user | Yes, rolling out over the weeks after 2026-10-05, all plans |
| ChatGPT or Codex, outside the EU | No, not at launch |
| OpenAI API, any country | Only if the organization opts in, for select models; off by default |
| OpenAI models through cloud partners | Planned "in the coming weeks"; availability may vary by output and by partner |
For how textGrain works, OpenAI's detection and robustness figures, and the full list of its limits, see this site's page on OpenAI's textGrain watermark. The AI watermark status tracker shows how every other provider compares.
What happened to OpenAI's shelved 2024 text watermark?
ReportedOpenAI shelved a reportedly 99.9%-effective text watermark and said why. That history ended on 2026-10-05, when it shipped a different system on narrower terms.
Leaked internal documents, reported by the Wall Street Journal and covered by Tom's Hardware on August 5, 2024, describe an internal ChatGPT text-watermarking method reportedly around 99.9% effective, which the company had weighed releasing for about a year. OpenAI chose not to ship it.
The reasons given, per that reporting, were specific.
- The method could be circumvented, which invites a false sense of certainty.
- False positives: 99.9% accurate still means a meaningful volume of wrongly flagged writing at ChatGPT's scale.
- Watermarking could stigmatise non-native English writers who use ChatGPT as a writing aid, since their legitimately assisted text would carry the mark.
- Internal survey data suggested close to 30% of users would use ChatGPT less often if the feature launched.
This is not only a leak. OpenAI has separately and publicly discussed text-watermarking research and its risks, naming false positives, circumvention, and disproportionate impact on non-native English speakers, in its own content-provenance posts. The leaked rationale and the published one make the same arguments.
That was a more detailed public rationale than Anthropic, Google, or Meta offered about their own text-watermarking decisions, and it made the missing ChatGPT text watermark a considered tradeoff rather than a technical gap. See the cross-provider map for what each of those providers has actually said.
Until 2026-10-05 nothing changed that. OpenAI's 2026 writing on provenance described text as a future goal, and a corroboration pass on August 17, 2026 against OpenAI's developer content-provenance guide found it documented images and audio and did not mention text. Secondary reporting also said OpenAI's Responsible AI Provenance Guide stated the company did not embed watermarks in text outputs; that could not be verified against OpenAI's own site.
textGrain is a different system, and it shipped on narrower terms than a global default. Outside the EU, ChatGPT text is not marked at launch. The detector is restricted to approved researchers rather than offered to the public. OpenAI also publishes its own failure figures: detection falls sharply after synonym swaps, and paraphrasing or translation can erase the mark.
What can removing a ChatGPT watermark actually mean?
Community discussionFor text outside the EU there is nothing to remove. For EU text, OpenAI's own numbers say rewording weakens the mark. For an image, the metadata falls off by accident and the pixel watermark has only been beaten by regenerating the picture.
Start with the text case, because it is the query people type. Until 2026-10-05 there was no live ChatGPT text watermark, so a tool claiming to remove one was making a claim about a mechanism OpenAI had built and shelved. Outside the EU that is still the position for ChatGPT text.
For EU ChatGPT and Codex text, and for API text from accounts that opted in, there is now a target, and OpenAI has published what weakens it. On 400-token English passages at a 1% target false positive rate, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. OpenAI's help page says the mark survives light edits, copy and paste, and screenshots depending on content, and that substantial paraphrasing or translation can make it undetectable.
Short text was weakly marked before anyone touched it. OpenAI says short passages are unreliable, code is harder to watermark, and factual or precise answers and verbatim reproduction carry a weak signal.
None of this can be checked by the person doing the editing. The detector is limited to approved researchers, so a tool that claims to remove textGrain cannot show a before-and-after result from OpenAI's detector unless its makers were approved. And by OpenAI's own statement, an absent watermark does not prove a human wrote the text, so removal cannot turn text into evidence of human authorship.
Cleaning hidden Unicode characters out of copied ChatGPT text is ordinary formatting hygiene, worth doing on its own terms. It does not affect textGrain, which OpenAI says inserts no hidden characters or invisible spaces.
For images the two layers come apart completely. The metadata is the part that breaks first, usually by accident: the June 2026 developer thread above shows an ordinary re-upload doing it, and format conversion, re-saving, or a screenshot do the same.
SynthID is the harder half, and public discussion shows a real asymmetry. The Hacker News thread about a CLI watermark-removal tool, posted May 19, 2026, pulled 387 points and 258 comments, making it the highest-engagement watermark discussion found there. It reported a hard technical limit: the tool cleanly strips the visible watermark, and defeating SynthID reportedly required regenerating the image entirely through Stable Diffusion XL, degrading quality in the process.
Commenters split on whether that is good news or bad news.
- The privacy and resistance framing: removal tools push back on tracking built into generated media.
- The abuse framing: the same techniques enable deepfake scams, with one commenter raising a Kentucky political deepfake ad reported to have cost $1.7 million and reached 49 million views.
- The sidestep: newer watermark-free local models make the whole removal question moot for anyone who just does not use a watermarking provider.
A separate discussion on the same site, from the day of OpenAI's SynthID announcement, drew 332 points and 180 comments. One commenter claimed Stable Diffusion at 10 to 15 percent denoising strength was enough to defeat SynthID, saying they tested it the day Nano Banana Pro shipped. Nobody published a reproducible test alongside that claim, so it stays community discussion, not a documented result.
For text, a July 2026 thread titled "Text AI watermarks will always be trivial to remove" asked whether foundation-model labs already use watermark removal internally to keep synthetic text out of their own training data. It proposed low-denoising image-to-image regeneration as a practical bypass, which is an image technique, not a text one. Nobody in that thread demonstrated a removal method against ChatGPT text, because there was no live ChatGPT text watermark until 2026-10-05.
What gets mistaken for a ChatGPT watermark?
Rumor/speculationFour things the internet keeps calling a ChatGPT watermark. OpenAI's documentation names none of them as a mechanism.
- Hidden Unicode characters such as zero-width spaces.
- Non-breaking spaces and other invisible whitespace.
- Repeated phrases and stock openers.
- Writing-style habits, punctuation tics, and sentence rhythm.
None of these appear in OpenAI's help-center article, or anywhere else in its documentation, as an actual watermarking mechanism. The text watermark OpenAI did ship, textGrain, works through word-choice randomness, and OpenAI says it inserts no hidden characters or invisible spaces.
So a stray zero-width space or non-breaking space in ChatGPT text is not the watermark, and its absence does not mean the text is unmarked. Cleaning invisible characters out of pasted text is still worth doing for formatting reasons, it just does not touch textGrain.
FAQ
Is ChatGPT output watermarked?
It depends entirely on the modality. Images from ChatGPT, the API, and Codex carry C2PA Content Credentials plus a SynthID watermark, except gpt-image-1 and gpt-image-1.5 images through the API, which OpenAI's help page excludes. Supported audio carries SynthID, without a C2PA manifest. Sora video is supposed to carry both visible and invisible signals, and independent testing has found that inconsistent. Text depends on where it came from: under a rollout OpenAI announced on 2026-10-05 and is completing over the coming weeks, eligible ChatGPT and Codex text in the EU carries OpenAI's textGrain watermark, API text carries it only if the customer opted in, and ChatGPT text elsewhere carries none.
Is my ChatGPT text watermarked outside the EU?
No, not at launch. OpenAI says it is adding textGrain to ChatGPT and Codex text in the EU only and is "not making text watermarking a global default at launch." The exception is text produced through an OpenAI API account whose organization opted in, which can be marked wherever the customer is.
Can I check text for the OpenAI watermark?
Not unless you are an approved researcher. OpenAI is limiting the textGrain detector to approved researchers and expert organizations at launch, case by case, and is not making it public. Enabling watermarking on an API account does not grant detector access either. Any website offering to detect OpenAI's text watermark is not using OpenAI's detector unless OpenAI approved it.
Does the OpenAI watermark identify me?
OpenAI says no. The detector reports whether it detects an OpenAI watermark "without identifying the user or revealing their prompts or conversations," and OpenAI lists identifying the user among the things a detection cannot do. It also cannot measure how much a person contributed or establish who owns the text.
Does copy and paste remove the ChatGPT text watermark?
No. textGrain is in the choice of words, not in added characters, and OpenAI says it inserts no hidden characters or invisible spaces. Copying and pasting carries the words over unchanged. What weakens it, per OpenAI, is changing the words: synonym swaps, substantial paraphrasing, or translation.
Does API text carry the watermark?
Only if the organization turns it on. Since 2026-10-05 API customers worldwide can opt in for select models under Organization settings > Data controls > Text provenance, with per-project overrides. It is off by default, there is no API parameter for it, and OpenAI has not published the list of eligible models outside the Settings screen.
Does ChatGPT's image watermarking prove ChatGPT text is watermarked too?
No. Image provenance (C2PA plus SynthID) and textGrain for text are separate systems built for different modalities. Images are marked in ChatGPT, Codex and the API except gpt-image-1 and gpt-image-1.5 images through the API, and supported audio in ChatGPT and the API. Text marking applies only to API output where a customer opts in, with EU ChatGPT and Codex text rolling out.
Why did OpenAI wait until 2026 to ship a ChatGPT text watermark?
It reportedly had one by 2024, about 99.9% effective by internal measure, and held it back over circumvention risk, a concern that it would stigmatise non-native English writers who use ChatGPT as a writing aid, and survey data suggesting roughly 30% of users would use ChatGPT less if it launched. The system it shipped on 2026-10-05, textGrain, is a different one, and it shipped narrowly: EU only for ChatGPT and Codex, opt-in on the API, and no public detector. OpenAI's post cites the EU AI Act requirement that generated text be identifiable in a machine-readable way.
Can you remove a ChatGPT watermark?
For ChatGPT text outside the EU there is nothing to remove. For EU or opt-in API text, OpenAI's own figures show that replacing 25% of words with synonyms cut detection of 400-token passages to 17%, and that substantial paraphrasing or translation can make the mark undetectable; nobody without detector access can confirm the result. On an image, the C2PA metadata layer comes off through routine handling and often disappears without anyone trying. Defeating the SynthID layer is a different matter: the busiest public report of it required regenerating the whole image through a different model, at a real cost to quality.
Does cleaning hidden characters remove a ChatGPT watermark?
It removes hidden Unicode characters if they are present, which is useful formatting hygiene regardless. It does not prove those characters were ever an OpenAI watermark, since OpenAI says textGrain inserts no hidden characters, and it has no bearing on textGrain or on image and audio SynthID and C2PA signals.
Does re-uploading an image to ChatGPT remove its watermark?
It removes part of it. A June 2026 developer forum thread found that re-uploading a PNG into a ChatGPT conversation strips its C2PA metadata down to bare file chunks, while the original file still verifies elsewhere. SynthID is a separate pixel-level signal and is not addressed by that at all.
Whose watermark is actually on a ChatGPT-generated image?
Both OpenAI's and Google's. OpenAI attaches its own C2PA Content Credentials metadata, then stacks Google DeepMind's SynthID pixel watermark on top, an arrangement OpenAI itself frames as complementary layers rather than a single proprietary system.
Can I check whether a ChatGPT image or audio clip still carries its mark?
Partly, and the two layers differ in who can check them. Any C2PA manifest viewer or the open-source c2pa-rs CLI reads the metadata layer yourself, and it often has not survived a re-upload or a re-save. The SynthID layer can only be checked by asking OpenAI's own tooling: the Verify tool and the Content Provenance API, both of which take images and audio and neither of which takes video.
Can I trust a result of no watermark detected on a ChatGPT image?
Not fully. OpenAI's own documentation says no detection method is foolproof and that metadata can be stripped by uploads, downloads, or file transformations. One observed case: re-uploading a genuinely OpenAI-generated image into a ChatGPT conversation strips its C2PA metadata down to bare file chunks, even though the original file still verifies elsewhere.
Next steps
- Read the full account of textGrain: who gets it, how the scheme works, OpenAI's detection numbers, and how it compares with SynthID and Claude's watermark. OpenAI's textGrain watermark
- See why the two layers on a ChatGPT image fail in completely different ways before you trust either one. C2PA vs SynthID
- See the broader evidence on which watermarks have actually been broken in published research, not just in forum claims. Can AI watermarks be removed?
- Paste text you are unsure about into the free in-browser cleaner to see whether it contains invisible Unicode characters. Hidden character cleaner
- Understand why a watermark and an AI detector answer completely different questions before you rely on either. AI watermark vs AI detector
- Check what the EU actually requires of providers and deployers now that the transparency rules apply. EU AI Act and AI watermarking
Sources and citation status
- OfficialOpenAI: Our approach to EU text provenance rules (2026-10-05)
- OfficialOpenAI Help: Provenance signals in OpenAI-generated content
- ResearchtextGrain: Entropy-Calibrated Watermarking for Language Model Text (OpenAI technical report, 2026-10-05)
- ReportingTechCrunch: OpenAI will start watermarking ChatGPT's text in the EU (2026-10-05)
- OfficialOpenAI: advancing content provenance
- OfficialOpenAI: source of what we see and hear online
- OfficialOpenAI API changelog (no watermarking or provenance entries, Oct 2023 to Aug 2026)
- ReportingTom's Hardware: OpenAI built a text-watermarking method for ChatGPT (via WSJ reporting)
- ResearchKirchenbauer et al., On the Reliability of Watermarks for Large Language Models (ICLR 2024)
- CommunityLessWrong: OpenAI does not appear to be applying watermarks honestly
- CommunityOpenAI Developer Community: are C2PA metadata chunks stripped from images uploaded to ChatGPT?
- CommunityHacker News: Remove-AI-Watermarks, a CLI and library for removing AI watermarks from images
- CommunityHacker News: OpenAI adopts Google's SynthID watermark for AI images with verification tool
- CommunityHacker News: Text AI watermarks will always be trivial to remove