No AI watermark detected—so is the photo real? Reading SynthID results

The fingers look fine and the text looks natural. An AI checker reports no watermark. Can you now stamp the photo “verified real”? Not finding the signal being sought is different from establishing that no AI was used. An unreadable barcode does not automatically reveal an item’s history.

On October 7, 2026, Google announced globally expanded access to SynthID Detector in English. This article uses official sources checked on October 28; it is not launch-day reporting or our own detector-performance experiment. Expansion announcement

Key points
  • SynthID checks invisible watermarks embedded in AI content.
  • Do not turn an undetected or unclear result into proof of no AI use or a real event.
  • Check each tool’s scope, then compare original sources, dates and context.
FOUR IMAGE CHECKS

What to check after a detector result

An editorial workflow, not an actual detection screen or measured performance.

  1. 01File conditionsChoose one image you may share and check supported conditions.
  2. 02Tool scopeRead which models and signals the current tool can check.
  3. 03Interpret resultsKeep detected, undetected and unclear results distinct.
  4. 04Compare sourcesExamine the original post, context and independent material.
This diagram guarantees neither an image’s authenticity nor detector accuracy. Sources: Gemini Help · Google DeepMind

1. A corner logo differs from a hidden signal

SynthID embeds and identifies invisible digital watermarks in generated media such as images. A visible logo’s presence or absence does not establish whether that signal exists. Google DeepMind says it is added during generation and designed to withstand modifications such as some cropping and compression. This is not a guarantee of detection after every possible edit. Technical explanation

2. Check the scope of the particular tool

The separate SynthID Detector expansion announcement describes checking Google and partner content. By contrast, the Gemini help page opened on our checking date says the app’s SynthID verification currently recognises content made with Google AI tools. Do not transfer one tool’s coverage to another interface. Public detector announcement · Current Gemini scope

On desktop Gemini, sign in with a supported account, add an image file and ask whether it was created or edited with Google AI. Official conditions specify one file at a time, no larger than 100 MB. Before uploading sensitive personal photos or files you have no right to share, check whether sharing is appropriate. Upload steps and conditions

Signal checkedWhat it can tell youWhat it does not establish
Visible logoWhether a visible mark is presentNo logo means no AI use
SynthID detectedA generation/editing signal within supported scopeThe whole scene is a real event
Undetected or unclearThe signal was not found or the check is inconclusiveDefinitely a real, unedited photo
Content CredentialsAvailable origin and editing historyThe caption is factually correct

3. Do not translate “not found” into “real”

Gemini help explains that content without a detected Google AI watermark may still have been made with another AI system. Simple images or small edits can produce unclear results, and repeated modifications can leave a watermark undetectable. Result limitations

Content Credentials are a separate technology for origin and editing history, used with both AI and non-AI content. Having a history and depicting a true event are different questions. A detector gives a detective a clue, not a verdict on the whole case. Role of provenance information

4. Check information outside the photograph too

Hypothetical example: an old real photograph is captioned “this just happened in our neighbourhood.” Its date or location can be wrong without generative AI being involved. We recommend comparing the original post, publication date, other photographs and reporting.

The official help page also suggests reverse image search and metadata from the original file as supporting checks. Additional methods The rule is simple: read what the tool found; do not treat failure to find a signal as verification of the facts. For a surprising photograph, open the source link before pressing Share.

Based on public sources
Back to the blogBack to top