Short answer: an AI watermark such as Google’s SynthID is a signal hidden inside content that was made by a particular AI system, so a matching detector can say “this came from our model”. A timestamp is a separate record showing that an exact file existed no later than a recorded time. They answer different questions, neither proves who authored anything, and neither replaces the other.
Last checked 12 November 2026. General information, not legal advice. TimeBinder does not detect AI, and a dated record is not proof of authorship, ownership or originality.
Table of Contents
What an AI watermark is
A watermark, in this sense, is a pattern added by an AI system when it produces an image, video, audio clip or text. Google describes SynthID as embedding imperceptible digital watermarks directly into AI-generated images, audio, text and video, designed to survive some common changes such as cropping, filters, compression or speed changes. For text, its research paper describes SynthID-Text as adjusting the probabilities of the words a model picks so that a detector can later look for the pattern.
So a watermark is added by the generator, at creation, and read back by that vendor’s detector. It is evidence about the tool, not about the person using it.
What a watermark can and cannot show
- Can show: that a piece of content was likely produced or edited by a particular watermarking AI system, if the detector finds the mark.
- Cannot show: that content with no mark is human-made. Content from tools that do not watermark, older tools, open models and edited or re-rendered material may carry no mark at all.
- Cannot show: who prompted the model, who owns the result, or when it was made. A watermark does not carry a trusted date.
- Is not guaranteed to survive: for text, the SynthID-Text paper itself says generative watermarks are weakened by edits such as LLM paraphrasing, and the research on stealing, spoofing and scrubbing attacks is ongoing.
Watermark detection is also vendor-specific. A SynthID detector looks for SynthID marks. It does not tell you what other systems did. This is why a missing watermark is not a clean bill of health, and why TimeBinder does not detect AI at all. For the problems with score-based detectors, see why working drafts and timestamps matter more than AI detectors.
What a timestamp is
A timestamp starts from your own file. It is turned into a fingerprint (a SHA-256 hash), and that fingerprint is anchored in the Bitcoin blockchain. Later, anyone can check that the same fingerprint existed by that time, and that your file still matches. TimeBinder does this with a TimeBind, which covers up to 10 files; the files stay on your machine. See how it works.
It does not look inside the file. It shows nothing about how the file was made, whether AI was involved, or who made it. It shows existence by a time, and no change since.
Side by side
- Who adds it: a watermark is added by the AI vendor. A timestamp is added by you, after the file exists.
- What it is attached to: a watermark lives inside the content and can be damaged by edits. A timestamp sits outside the file and refers to an exact version; edit the file and it is a different file needing a new timestamp.
- What it answers: a watermark answers “did this system produce this?”. A timestamp answers “did this exact file exist by this time?”.
- Who can check: a watermark is read by the vendor’s detector. A timestamp can be checked independently.
- Coverage: a watermark only covers content from systems that apply one. A timestamp covers any file you choose to date.
Why watermarks are in the news
Article 50 of the EU AI Act requires providers of AI systems that generate synthetic audio, image, video or text to ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, as far as that is technically feasible. The Act itself acknowledges limits of cost and the state of the art. That is a rule about what AI providers do. It does not turn a watermark into evidence about who made a particular file, and it says nothing about files that never passed through such a system.
Provenance standards sit alongside this. C2PA Content Credentials record how a file was captured or edited, signed by the tool or device. We compare that with timestamping in Content Credentials (C2PA) vs a timestamp. The same caveat applies: the metadata can be stripped, and a credential describes a process, not the truth of the content.
When each one helps
- You are checking whether content came from a particular AI tool: that tool’s own detector, if it offers one, with the caveats above.
- You are a creator who wants to show your own working files existed: date the originals, drafts and project files as you go. See draft history. A text-focused view for writers is in AI text watermarks: what they mean for writers.
- You are in a dispute about authenticity: a watermark result, a timestamp, metadata and the file history are separate pieces. Together with other evidence they can help; none is conclusive alone. For legal settings see the lawyers’ guide to AI forgeries.
Tools that can help
There is no affiliate link on this page: nothing here depends on a paid third-party tool. Our affiliate disclosure explains when we do use them.
Frequently Asked Questions
-
Does an AI watermark prove who made something?
No. It can indicate that a watermarking AI system generated or edited the content. It says nothing about who prompted it, who owns it or when it was made.
-
If there is no watermark, is the content human-made?
Not necessarily. Many tools do not watermark, and marks can be weakened or removed by editing, re-rendering or rewriting.
-
Does a timestamp show whether AI was used?
No. It shows that an exact file existed no later than a recorded time and has not changed since. It does not examine how the file was made.
-
Can a timestamp be damaged like a watermark?
The timestamp record lives outside the file. If you change the file, its fingerprint changes, so the old timestamp refers to the earlier version, which you should keep.
-
Do I need both?
They solve different problems. A creator trying to show a work history usually needs the dated files, not a watermark. A reader checking a suspect image may try the vendor’s detector. Neither is a verdict.
Sources checked
- Google DeepMind, SynthID (imperceptible watermarks in AI-generated images, audio, text and video; detection by Google tools).
- Dathathri et al., Scalable watermarking for identifying large language model outputs, Nature, 23 October 2024 (SynthID-Text).
- EU AI Act, Article 50, transparency obligations for providers (marking of synthetic content, “as far as this is technically feasible”).
- C2PA, C2PA Specifications (provenance standard behind Content Credentials).
Date your own files while they are still as you made them
Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.
