SynthID works, within a scope most people misunderstand. It reliably answers one question — "did a participating AI model generate this?" — and it answers nothing else. Since May 2026 that scope got considerably wider, which makes the boundary worth understanding precisely.
What SynthID actually is
SynthID is Google DeepMind's invisible watermarking system. Rather than writing to metadata, it modifies the generated content itself — pixel values for images and video, frequency spectra for audio, token sampling distributions for text. Because the signal lives in the content rather than beside it, it survives the things that destroy metadata: cropping, compression, filters, format conversion, and screenshots.
It is no longer Google-only
For its first years SynthID marked only Google's own model outputs — Imagen, Veo, Lyria. That changed in May 2026, when OpenAI joined the C2PA steering committee and began embedding SynthID alongside C2PA Content Credentials in images generated through ChatGPT, Codex, and its API. Two of the largest generators now apply the same invisible mark, which moves SynthID much closer to an industry default than a vendor feature.
This is a genuinely good development, and it is worth being clear about what it changes and what it doesn't. It means far more AI-generated imagery now carries a durable origin signal. It does not mean the internet is now labeled.
Does SynthID survive a screenshot?
Generally yes for images, and that is the whole design goal — the mark is in the pixels, so re-photographing those pixels carries it along. Screenshot survival is exactly where SynthID beats C2PA, which a screenshot destroys completely.
Robustness is a gradient, though, not a switch. Every pixel watermark — SynthID, ours, anyone's — degrades as the pixels are transformed further. Heavy re-editing, aggressive re-compression, and above all running the image back through a generative model degrade the signal. Anyone claiming a watermark is unconditionally unremovable is selling something; the honest framing is a measured survival curve, which is why we publish ours including the failures.
Does SynthID have false positives?
Detection returns a confidence level, not a verdict, and Google's own detector explicitly offers an uncertain answer for text alongside watermarked and not watermarked. Confidence thresholds trade false positives against false negatives — tighten the threshold and you miss real marks; loosen it and you flag clean content.
Two practical cautions. First, SynthID text detection is substantially less reliable than image detection, because the watermark lives in token choices that paraphrasing disturbs. Second, in July 2026 Google confirmed and fixed a bug where Gemini returned the first uploaded file's detection result for subsequent files in the same chat session. Use the dedicated SynthID Detector portal rather than asking a chatbot, and treat any single detection as evidence rather than proof.
The limit that matters most
Here is the part that gets lost. SynthID is a producer-side mark. It is applied by the model, at generation time, to synthetic content. It says "an AI made this."
It cannot say anything about the photograph you took yourself. Your original work is not generated by a participating model, so nothing marks it, so it carries no SynthID signal at all — and its absence is meaningless, because most authentic content has no watermark of any kind. As producer-side marking becomes standard across Google and OpenAI, the asymmetry gets sharper rather than softer: synthetic media is increasingly labeled at birth, while authentic human work stays completely unmarked unless its creator marks it.
What to use, and when
- Checking whether an image is AI-generated — use the SynthID Detector portal and a C2PA viewer together. Both signals present is strong evidence. Neither present is weak evidence of nothing, because non-participating models exist and marks get degraded.
- Proving you created something — SynthID is structurally the wrong tool. You need a mark you apply to your own content before publishing.
- Regulatory marking obligations — if you are a provider generating synthetic media, the EU AI Act's Article 50 marking duty applies from August 2026; our breakdown is here.
The honest summary
SynthID is good technology solving a real problem, now with meaningfully broader adoption than it had a year ago. It is not a theft defense, not a proof of authorship, and not a reason to treat unmarked content as authentic. It labels the synthetic. Somebody still has to authenticate the real.