DALL·E, Midjourney and Grok: Which Actually Watermark Your Images?
Key takeaways
- The gap is not visible versus invisible — it is metadata versus nothing. Most commercial tools write provenance data; many open-weight pipelines write none.
- DALL·E and ChatGPT images typically carry a C2PA manifest. That is the signal platforms read first.
- Midjourney has historically added no visible mark, which is why people wrongly assume its output is untraceable.
- Grok and Flux-family deployments vary enormously — the same model can mark heavily or not at all depending on who runs it.
- Stop guessing. Inspect the file: metadata is fully readable, and takes seconds to check.
Ask around and you will hear it stated as fact: DALL·E watermarks everything, Midjourney watermarks nothing, Grok is a mystery. Two of those three are wrong in ways that matter.
The real dividing line between generators is not whether you can *see* a mark. It is whether the file quietly carries a provenance record — and that is a completely different question with a completely different answer per tool.
This is a spoke of our AI image watermarks pillar guide.
Generator by generator
| Generator | Visible mark | Invisible watermark | Provenance metadata | Practical footprint |
|---|---|---|---|---|
| OpenAI / DALL·E / ChatGPT images | Historically on some surfaces | Not publicly confirmed for stills | C2PA manifest, commonly | Metadata is the real signal |
| Midjourney | No | Not publicly confirmed | Limited | Small file footprint, still classifier-detectable |
| Grok | Varies by deployment | Varies | Varies | Unpredictable — inspect the file |
| Flux-family / open-weight | Operator's choice | Usually none by default | Usually none | Smallest footprint |
| Adobe Firefly | No | Reported | Deep C2PA implementation | Heaviest provenance record |
| Google Gemini / Imagen | Product-dependent | SynthID | C2PA + generator tags | Heaviest overall |
Read down the metadata column. That is where commercial providers cluster and open-weight pipelines do not — and it is the column platforms read first.
Why "no visible watermark" fools people
A visible badge is honest about its presence. Metadata is not — it is invisible, it is structured for machines, and it survives every operation except re-encoding.
So the intuition "I see no watermark, therefore there is no watermark" is exactly backwards. The images most likely to be labelled on upload are often the ones with nothing visible on them at all, because they are the ones from big commercial providers who implemented C2PA properly.
The open-weight gap
Run a model on your own hardware and nothing in the pipeline marks the output unless you add marking yourself. No manifest, no generator tag, no watermark.
This is the structural weakness in watermarking as a policy tool: it covers participating providers only. It is also why the labelling laws and industry agreements put obligations on providers and deployers rather than relying on marking alone, and why platforms invest in classifiers that do not need a watermark at all.
Stop guessing — read the file
Metadata is fully inspectable. Drop an image into the metadata tool and you get an itemised list of what is actually present: C2PA manifests, EXIF fields, XMP tags, IPTC entries, ICC profiles, PNG text chunks and generator strings. No speculation about what a given generator "usually" does.
Invisible pixel watermarks are the exception: verification requires the issuing provider's detector, which is deliberately not public. Our SynthID guide explains why that asymmetry exists.
What removal can and cannot do
| Signal | Removal | Certainty |
|---|---|---|
| Visible badge | Crop or content-aware patch | Complete |
| C2PA manifest | Re-encode from raw pixels | Complete |
| EXIF / XMP / IPTC / PNG text | Re-encode from raw pixels | Complete |
| Invisible watermark | Optional resample and dither | Best-effort only |
| Visual classifier detection | Not addressable | None |
The one-minute habit
Before publishing any AI image, inspect it and read the before-and-after report. You will find out which of the rows above apply to your actual file, and you will stop relying on second-hand claims about what your generator does.
Read next: how platforms turn these signals into an AI label →
Frequently Asked Questions
Does DALL·E add a watermark to images?
Some OpenAI surfaces have applied a visible mark historically, but the dependable signal is metadata: images commonly carry a C2PA provenance manifest recording that they were generated. That manifest is invisible in the picture and readable by any platform on upload, which makes it the signal that actually matters.
Does Midjourney watermark its images?
Midjourney has not historically stamped a visible mark on output, and its metadata practices are limited compared with the major platform providers. That is why its images are often assumed to be untraceable — though style-based and classifier-based detection is entirely independent of any watermark.
Does Grok add a watermark to generated images?
It varies. Grok and other deployments built on open-weight model families inherit whatever marking the operator chooses to configure, so the same underlying model can produce heavily tagged output in one product and completely unmarked output in another. Check the specific file rather than trusting a general answer.
Do open-source AI image models embed watermarks?
Usually not by default. Open-weight pipelines run on hardware the operator controls, and marking is an opt-in step in the generation stack. That is precisely why watermarking alone cannot be a complete solution to provenance — it only covers the providers who choose to participate.
How can I check whether my AI image carries a watermark or metadata?
Inspect it. Metadata is fully readable, so a tool can enumerate every C2PA manifest, EXIF field, XMP tag, IPTC entry and PNG text chunk in the file and tell you what is there. Invisible pixel watermarks are the exception — those generally cannot be verified without the issuing provider's own detector.
If an image has no watermark, is it undetectable as AI?
No. Watermarks and metadata are only two of the signals. Platforms also run visual classifiers trained on generated imagery, and those work on the pixels alone regardless of what the file contains. An unmarked image is harder to flag with certainty, not impossible to flag.
Which generator leaves the smallest footprint?
Self-hosted open-weight pipelines, because nothing in the stack adds marking unless the operator configures it. That is a statement about what the file contains, not a promise of anonymity — classifier-based detection is entirely indifferent to your metadata.
AI Watermarks & Platform Labels: complete guide series
AI Image Watermarks in 2026: Every Visible Badge and Invisible Signal, Explained
A complete guide to AI image watermarks: which generators add visible badges, which embed invisible signals, and what actually survives editing.
AI Image Labelling Rules: The Laws, Treaties and Agreements Behind Every AI Tag
11 min readHow to Check Your Image for AI Watermarks and Tags Before You Post It
9 min readGemini and Samsung Galaxy AI Watermarks: The Marks on Your Phone Photos
9 min readHow Facebook and Instagram Know Your Image Is AI (The Full Detection Chain)
11 min readThe Meta AI Watermark: The Badge, the Hidden Tag, and How to Handle Both
9 min readThe Sora Watermark: What It Is, Why It Moves, and What Removal Really Means
10 min readVisible vs Invisible Watermarks: What Each One Survives
10 min read
Related Articles
AI Image Detection in 2026: How Detectors, Watermarks and Metadata Really Work
16 min readAI Image Metadata Explained: EXIF, XMP, C2PA and the Generator Tags That Expose You
12 min readAI Image Watermark Remover: The Complete 2026 Guide
10 min readAutomate Your Image Workflow with Batch Processing
6 min read