Google’s SynthID detector needs safeguards before its results affect people’s work

The distinction is central. Engadget reports that the detector cannot tell whether a file was generated entirely with AI or edited with AI, and does not identify which parts of a file are synthetic. A positive result may show that the tool recognized a mark. On its own, it cannot establish how much AI was involved, how the file was made or whether its creator broke a rule.

Google says SynthID marks are designed to be imperceptible to people and detectable by its technology. The company also says image and video marks are designed to withstand some changes, including cropping and compression. Those features could make a mark useful when it survives editing. But they do not turn a watermark check into a record of a file’s entire history, and Google’s description is not an independent performance test. Google DeepMind’s overview describes how the watermarking system works.

The strongest case for the detector is that a machine-readable mark can add information a viewer might otherwise miss. That is worth having. The mistake would be to treat a useful clue as conclusive evidence about a person or their work, especially when the tool cannot distinguish full generation from AI-assisted editing.

No mark does not mean human-made

A negative result has its own limits. If the detector finds no SynthID watermark, that means it did not identify a mark it recognizes—not that a file is authentic, untouched or made by a person. The tool does not cover all AI-generated content, and Engadget reports that watermarking approaches differ and that watermarks can be removed in some situations. Tom’s Guide also cautions that a missing SynthID mark does not establish that media is authentic. Engadget’s report and Tom’s Guide’s explanation describe these limits.

This matters for the person being assessed in either direction. A positive result can be overread as proof that an entire image or recording was fabricated. A negative one can be mistaken for proof that it was not AI-assisted. Neither conclusion follows from the result alone.

Google says it tested an earlier SynthID Detector portal with journalists and media professionals before opening global access. The company says it launched SynthID in 2023. But the material available about the public website contains no independent audit or verified false-positive and false-negative rates. Without such evidence, users cannot judge how reliably results hold across file types, edits and content sources.

Consequences need human review

The right standard is straightforward: no one should face a consequential decision based on a SynthID result alone. Schools, employers, publishers and platforms that choose to use the tool should explain what a result means, disclose its limits and give the person affected a chance to respond. A reviewer should seek corroborating evidence rather than treating the detector as a ruling.

Google has a role too. It should make the site’s explanations clear enough for ordinary users to understand what was detected, and publish independent evaluations that examine error rates across supported media and common edits. The available reporting does not establish what detail the public interface gives about why it returned a result, so that transparency is part of the test the tool still has to meet.

A watermark detector can help people check for a specific signal. It cannot settle authorship, intent or credibility. Google can clarify the tool’s limits; organizations that choose to use it should require corroborating evidence before relying on a result in their own consequential decisions.