How to remove the "Made with AI" label from an image on X

There are two different “Made with AI” labels on X, and you control only one of them. The first is a disclosure you switch on yourself when you post. The second appears when X decides an image is AI-made; in one published test, the author could not remove that one from the post. Removing metadata such as C2PA Content Credentials from an image before you upload it takes those signals out of your copy, and a re-scan confirms they are gone. But X has not said which signals it reads, and cleaning does nothing to a post that is already live.

Our prediction: medium confidence. We expect an image cleaned with NoWatermark, with the re-scan showing Content Credentials removed, not to get X’s automatic label. The confidence is medium because X has not said which signals it reads. A March 2026 test by Roboin points to C2PA as one factor but not the only one. And ChatGPT and Gemini images also carry SynthID, which cleaning cannot touch. This is a prediction from published sources and X’s own documentation, not yet our own test result. We are running that test and will publish the result on this page.

Try it: check and clean your image before you post. It runs in your browser, and your file is never uploaded to us.

How to remove the AI label from an image before posting on X

1. Open Remove AI Label Metadata and drop in the exact file you plan to post. It runs in your browser, and the file is never uploaded.
2. Read the scan. The C2PA / Content Credentials row is the signal a March 2026 test by Roboin points to on X. Also check XMP, IPTC and AI generator metadata.
3. Choose Everything removable and click Clean metadata.
4. Check the report. Each of those rows should read Removed, which means a second, independent scan of the cleaned copy could not find it. SynthID always reads Unable to verify: no file cleaner can check or remove it.
5. Download the cleaned image and post that file. A post that is already labelled is not changed; the cleaned file goes up as a new post. The “Made with AI” toggle you set yourself is separate from the file and is yours to choose.

Two labels that look alike

The label you set. In early March 2026, X added “Made with AI” and “Paid partnership” tags to the composer. PiunikaWeb reported on 2 March 2026 that the tag “shows up right under the content for everyone to see” and that “it’s up to users to add these tags themselves”. We could not find an official X Help page describing the toggle. The most detailed walkthrough we found is by Roboin, who says to use the flag icon at the bottom of the composition screen before posting, or the post’s menu (Add/remove content disclosure, or Content disclosure) afterwards. Menu wording may differ by app version. Roboin also notes that because the label is not mandatory, X has not officially said when it should be applied.

The label X applies. Roboin’s second article describes a notice reading “AI-generated content detected” and a “Made with AI” label added without the author choosing it. X’s own Help page on its EU media literacy plan says X “has also expanded labeling for manipulated and AI-generated media, including ‘Made with AI’ indicators”, but it does not say how the label is applied or what signals are read.

What triggers the label on X

Platforms that document their systems describe some mix of four sources: metadata read at upload, invisible watermarks, visual classifiers and self-disclosure. X has not said which of these it uses for an automatic label. Everything below about triggers comes from one first-hand test, so treat it as an anecdote.

In a test published in March 2026 and updated in June, Roboin uploaded three images, one each from ChatGPT, Grok and Gemini. The ChatGPT image was labelled. The Grok and Gemini images were not. Roboin concluded that while X may use C2PA as one factor, it “does not appear to rely on C2PA alone”.

Three limits matter here:

• It is three images. Roboin also notes that X users report Nano Banana images being labelled and non-AI images being labelled. We cannot say how often either happens.
• It predates SynthID in ChatGPT images. OpenAI’s current help page says images from ChatGPT, Codex and the API include both C2PA metadata and SynthID watermarks. Roboin ran its test in March 2026, before Google announced in May that OpenAI was bringing SynthID to its content, so the test cannot say which of the two signals X reacted to.
• The Gemini result does not settle whether Gemini images carry C2PA. Google says images from Nano Banana Pro carry C2PA metadata. Roboin does not say which Gemini model it used, so we cannot reconcile the two. We are not claiming X ignores C2PA, or that Gemini images lack it.

Roboin also reports that the notice is internally named COMPOSE_VIEW_C2PA_AI_DETECTED_TOAST, a detail it took from an X user’s post. Roboin’s own conclusion is only that, judging from the name, it is possible X reads C2PA metadata. That is inference, not something X has confirmed.

What removing metadata changes, verified

NoWatermark scans an image for C2PA manifests (showing the manifest size and claim generator), XMP, IPTC and AI generator metadata. It then removes what you choose, scans the cleaned file again and marks a signal as removed only if the second scan cannot find it. The image data is copied byte for byte, not re-encoded.

The Remove AI label tool offers three presets:

• Everything removable: EXIF, IPTC, XMP, Content Credentials and embedded text. One EXIF rotation field is kept by default.
• Provenance-light: Content Credentials, XMP and embedded text. EXIF and IPTC stay.
• Privacy-safe: location and device data. Provenance is kept.

If a C2PA manifest is what X reacts to, cleaning removes it from your copy. We have not yet tested a cleaned file on X, and X has not said what it reads, so we cannot tell you the label will not appear. The scanner shows that a manifest exists and which tool wrote it. It does not read whether a credential says an image was generated or edited. To see what a file carries before you post, use the C2PA Checker or read what Content Credentials are.

What it cannot change

SynthID is always reported as Unable to verify. NoWatermark cannot detect it, cannot remove it and cannot confirm it is absent. OpenAI says of its own SynthID watermark that “unlike metadata, the signal is part of the content itself and may persist through some edits or transformations.” Both ChatGPT and Gemini images carry SynthID according to their makers, so cleaning their metadata leaves that signal untouched. See our SynthID guide.

Cleaning also cannot change a visual classifier, if X uses one, or the disclosure you choose to make. And X has not said whether an automatic label can be appealed.

If your post is already labelled

Cleaning a file changes the copy on your device, not a post that is already live. Re-uploading a cleaned file creates a new post.

What the sources say about an existing post:

• A label you set yourself. Roboin found it could be removed later from the post menu.
• A label X applied. Roboin found the “Made with AI” setting on that post greyed out and unchangeable. That is one post, and no X statement confirms it.
• An appeal. X has not documented one that we could find.

If a real photo was labelled after an AI-assisted edit, see why is my photo labelled AI.

X’s disclosure rules

No general requirement. Before the March 2026 launch, Social Media Today reported that there was “no requirement, at this stage, to include this disclosure within X posts”. The toggle is author-set, and Roboin says X has not officially set out when to use it.

A payout rule for armed-conflict videos. X’s Help page for its Original Content Rewards Program says that, effective 3 March 2026, users who post AI-generated videos of an armed conflict without adding a disclosure that it was made with AI will be suspended from the programme for 90 days. Subsequent violations will result in a permanent suspension of further payments. X adds that it may adjust the policy. This is about videos, about armed conflict and about payout eligibility. It is not an image rule and does not apply to accounts outside the programme. Press coverage, including AFP, calls the programme “Creator Revenue Sharing”. AFP reported that X said violations would be flagged through Community Notes and through “metadata and other technical signals” embedded in AI-generated content. AFP does not say which signals. NoWatermark does not process video.

Synthetic and Manipulated Media. X’s rule is about “misleading media”: inauthentic, manipulated or out-of-context media that may result in widespread confusion on public issues, impact public safety or cause serious harm. X adds that where it cannot reliably determine that media is misleading, it may not take action. The rule does not mention the “Made with AI” toggle and does not require disclosure of AI images.

Cleaning a file does not change any of these rules. If you use AI to make images, the honest route on X is the toggle.

Related

For the older question of whether X strips EXIF, see does X remove EXIF data?. For other platforms, see the guides on Instagram, TikTok photo posts, Pinterest and LinkedIn.

Try it on your own image

Drop your image into Remove AI Label Metadata. It lists every signal the file carries, removes the ones you choose, and scans the cleaned copy again so you can see what is gone and what remains, including anything it can only mark Unable to verify. It all runs in your browser: your file never leaves your device. If you made the image with AI, X’s own toggle is how you disclose it.