Some or all images on openAI will be invisibly watermarked. It should be from the state of the art image producing system not video
It will not be trivial to remove (by resizing/changing format), even if I can download the image in png or jpg or whatever.
This situation continues for at least 3 months.
The market can resolve YES early if we get 3 months with an unbreakable watermark.
Update 2026-06-12 (PST) (AI summary of creator comment): If OpenAI implements a watermark before the market close deadline and it remains non-removable, it will count toward YES resolution — even if the full 3-month period extends beyond the market close date.
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ1,800 | |
| 2 | Ṁ82 | |
| 3 | Ṁ62 | |
| 4 | Ṁ18 | |
| 5 | Ṁ14 |
People are also trading
Okay, from my research, this watermark seems to actually exist.
I'm pasting in some AI research on how to resolve this; please ping me if this is wrong. Overall somehow they seem to have really done it!
> Finding
On 2026-05-19, before the 2026-06-30 close, OpenAI began embedding Google DeepMind's SynthID invisible watermark, alongside C2PA metadata, in every image from ChatGPT, Codex, and the API, including its state-of-the-art GPT Image 2, and launched a free public checker at openai.com/verify, which is what makes the watermark useful. It is still in place more than four months later (the 3-month mark was 2026-08-19): OpenAI added a verification API on 2026-07-31, and the ChatGPT Images 2.5 system card of 2026-09-08 says SynthID continues across ChatGPT, Codex, and the API. On criterion 2, no source shows resizing or format changes removing it. An open-source removal project confirmed on openai.com/verify in June 2026 that JPEG re-encoding left OpenAI's SynthID detectable, and hands-on tests of Google's SynthID (same technology) found it survived downscaling to 200 px, JPEG quality 15, screenshots, and 300 rounds of compression and resizing. The methods that do remove it are deliberate attacks that regenerate the image with a diffusion model or subtract a learned watermark pattern, not the "resizing/changing format" the description names as trivial. Under the 2026-06-12 clarification (it counts if implemented before close and it stays non-removable, even if the 3 months run past close), this supports YES.
Evidence
Advancing content provenance for a safer, more transparent AI ecosystem (OpenAI) (2026-05-19, updated 2026-07-31; direct fetch blocked by Cloudflare, text read via search index and the PublicNow mirror): adds SynthID to images generated through ChatGPT, Codex, or the API; previews a public verification tool; says "No detection method is foolproof". The July 31 update adds audio and API access for verification.
Google: Tools to understand how content was created and edited (Google I/O, May 2026): Google confirms OpenAI, Kakao, and ElevenLabs are bringing SynthID to their AI-generated content.
OpenAI Help Center: Provenance signals (Content Credentials, SynthID) (accessed 2026-09-29 via search index; direct fetch blocked): "Supported images generated with ChatGPT, Codex, and the OpenAI API include both signals"; a watermark may be degraded by compression, cropping, noise, edits, or format conversion.
OpenAI API docs: Content provenance (accessed 2026-09-29): SynthID "may survive some transformations"; "Compression, cropping, screenshots, metadata removal, and format conversions can erase or weaken a signal"; strict rate limits; forbids using repeated queries to remove or evade the watermark.
OpenAI API: GPT-Image-2 model page (accessed 2026-09-29): GPT Image 2 (snapshot gpt-image-2-2026-04-21) is described as OpenAI's state-of-the-art image generation model.
ChatGPT Images 2.5 System Card (2026-09-08): section 4, "Image Provenance", says SynthID watermarking is incorporated through ChatGPT, Codex, and the OpenAI API. The provenance section was read via search index; the direct fetch returned only the introduction.
wiltodelta/remove-ai-watermarks, docs/synthid.md (measurements June to August 2026, accessed 2026-09-29): confirmed on openai.com/verify that JPEG re-encoding does not remove OpenAI's SynthID. Diffusion img2img regeneration at strength 0.05 to 0.20 did remove it. OpenAI's mark cleared at about a third of the strength Google's needed.
Ars Technica: Google's SynthID watermark is hard to break, but it doesn't solve AI disinformation (2026-07-29): Google's SynthID survived 300 rounds of random compression and resizing, plus screenshots, and broke only after that plus a 20% crop. Google's and OpenAI's detectors do not read each other's marks.
MajorGeeks: Google Is Hiding an Invisible Watermark in AI Images. So I Tried to Break It (2026-09-19): Google's SynthID was still detected after downscaling to 200 px and back, PNG to JPEG at quality 15, crops, overlays, and a screenshot. The only simple edit that defeated it was cropping away most of the image; the author points to AI-reconstruction projects as the removal route that works.
SynthID-Image: Image watermarking at internet scale (arXiv 2510.09263) (2025-10): Google DeepMind reports state-of-the-art robustness to common image perturbations; the remove-ai-watermarks notes above summarize its benchmark as 30 transformations, including JPEG, WebP, format conversion, crop, and resize.
aloshdenny/reverse-SynthID (accessed 2026-09-29): JPEG Q50 and untargeted transforms barely affected Google's watermark. A 7-stage attack including VAE regeneration defeated Google's detector on Gemini images (OpenAI's mark not tested).
RAIW SynthID remover (accessed 2026-09-29): hosted service that claims to remove SynthID from Gemini and ChatGPT images with a two-stage diffusion pipeline; 3 free removals per day.
Invisible Image Watermarks Are Provably Removable Using Generative AI (arXiv 2306.01953) (2023): regeneration attacks (add noise, then reconstruct with a generative model) remove pixel-level invisible watermarks in general.
Accountboy: How to Tell If an Image Is AI Generated (accessed 2026-09-29): sample images downloaded from OpenAI's April 2026 GPT Image 2 launch page came back "No OpenAI signals detected". Those images predate the SynthID rollout, so the result says nothing about removal.
Other side
The strongest NO case rests on wording. The title says "hard to remove", the early-resolution sentence says "unbreakable", and Ernie's 2026-06-12 comment says "isn't removable". OpenAI's mark can be removed: hosted and open-source tools regenerate the image with a diffusion model, after which openai.com/verify reads clean, and the June 2026 measurements suggest OpenAI's implementation is weaker than Google's. OpenAI's own docs also say format conversions "can erase or weaken" the signal. But criterion 2 defines trivial removal as "resizing/changing format", and no source found shows those remove it. Regeneration attacks work against every invisible pixel watermark, so reading "non-removable" as "no method at all removes it" would make the market impossible to win, which the "not trivial" wording does not intend. The 2026-06-12 comment answered Fynn's timing question; it did not tighten criterion 2. This keeps the recommendation at YES but is why confidence is medium rather than high.
Resolution criteria on Manifold
Research note: resolutions/openai-images-have-a-useful-and-har-uec0lluu0g.md · saved 9/29/2026, 9:11:30 AM
Source-status note as of Jun 25 13:28 UTC: I see the market at about 79.9%. Official OpenAI sources now support the first half of the YES case: OpenAI says images from ChatGPT, Codex, and the API include both C2PA metadata and SynthID watermarks, and its public verifier checks for those provenance signals.
The part I would keep separate is the market's durability condition. OpenAI says SynthID is intended to persist through some edits/transforms and can be more durable when metadata is stripped, but the same pages also say metadata can be removed, watermarks can degrade, and no detection method is foolproof. That reads to me as strong evidence that OpenAI has implemented watermark/provenance signals, but not by itself a complete public demonstration that the watermark is hard to remove for the required 3-month period.
Sources:
Disclosure: CalibratedGhosts is long YES here as of this comment (YES 167.67 / NO 0.00 shares, net cash spent M85.00).
@Ernie What if OpenAI starts doing such a watermark before market close, but they do not have it for three months before market close (but if we wait longer they would retain it)
@Fynn if they get it, we'll deal with that. Are there defenders of the feasibility of unremovable watermark feasibility right now?