How to run a SynthID check

Choose a check for your format

Identify what you are checking before choosing a SynthID checker. A screenshot of a paragraph is an image file, but checking that image does not answer who wrote the words inside it. A PDF can contain selectable text, pictures, or both. Choose which part of the document you want to verify.

The media route follows Google's Gemini verification instructions. For the mechanisms behind these choices, read our guide to text watermarks and SynthID.

Write a narrow question before proceeding. "Did this image involve Google AI?" and "Did someone invent the event described in this image?" require different evidence. A watermark result does not check whether a caption is accurate or whether the uploader owns the image.

Check an image, video or audio file in Gemini

Google's current verification workflow requires a signed-in personal or qualified Workspace account. Choose one file up to 100 MB. Video must be under 90 seconds; audio must be under one hour.

• Open Gemini.
• Choose Add files, then upload the supported file.
• Ask Gemini whether Google AI created or edited it.
• Read the verification response and keep it with that file.

These instructions cover media files; pasted-text verification is not listed. Check the linked instructions if the controls differ from what you see.

Before uploading, decide whether you can share the file with that service. Keep your local original, and record any preparation you make to an upload copy. For example, if you extract a still from a video, label it as a still. The result belongs to the submitted image, not automatically to every frame of the recording.

If you only describe the file, you have not submitted that file for verification. Record which item you uploaded so you can distinguish a file check from a general conversation about its contents.

Can a SynthID checker verify pasted text?

A text check needs a detector compatible with the watermarking system used during generation. Google documents keys and configuration for SynthID Text, and says models sharing a tokenizer can share a watermark and detector under specified training conditions. Downloading an algorithm alone therefore does not establish compatibility with an unknown passage. See Google's SynthID Text documentation.

Ask a text-checking service which scheme it verifies, what access it has to that scheme, and what its result means. Look for a stated scope and evaluation. A page that accepts text and returns a percentage may be running an ordinary AI classifier. Its title or search snippet cannot settle that question.

For a passage copied from a chat, preserve the conversation export or other available record before reformatting it. If you cannot access a compatible verifier, record "statistical watermark not checked." That is a useful answer about the evidence available. Replacing it with a guess about Gemini, Claude or a particular model would overstate what you learned.

What the open-source SynthID Text code can do

Google DeepMind's SynthID Text repository provides a reference implementation for research. Its example notebook generates and detects watermarked text with Gemma or GPT-2. The repository explicitly distinguishes this code from a production implementation.

A controlled experiment starts with text whose generation setup you know. Keep that setup fixed while you test a detector, and reserve examples that you did not use to choose its settings. Record the experiment before examining the final results so that a disappointing result does not quietly change the question.

For example, you could compare an untouched generated passage with a copy that you edited according to a written plan. Label that as a test of those passages under that configuration. It would not establish whether a separate paragraph from an unknown website carries a provider's production watermark.

If you publish such an experiment, include unsuccessful checks and uncertain outputs. A collection containing only the examples a detector recognized cannot show how often it misses a watermark.

What Zyphh can inspect in your text

The Zyphh watermark checker inspects Unicode characters in your browser. It reports positions and character categories, including unusual spaces and invisible formatting. You can inspect each finding in context before making changes.

Paste the original text or import a supported document, then review the extracted text before scanning. Open the character inventory and inspect the surrounding words. A joiner or a nonbreaking space may belong in that location. Our hidden Unicode character guide explains how to decide what to keep.

If you choose cleanup, inspect the proposed changes and save the cleaned copy separately. Compare names, word boundaries and language rendering with the original. Removing a character establishes that the character changed; it does not verify a statistical watermark's absence.

The statistical verification area reports "not tested" when no compatible remote verifier is connected. Do not read the local character count as a substitute for that check. The methodology page describes the implemented checks and their limits, including the separate classifier used by our AI text detector.

Read negative, uncertain and mixed results carefully

Keep the result's scope in your notes. Google says Gemini's SynthID check currently recognizes Google AI content. It also describes uncertain outcomes and cases where extensive alterations can prevent detection. A negative result does not rule out other AI systems. Gemini can also report Content Credentials about a file's history; keep those separate from a SynthID finding. Google explains the media results here.

For text, Google's detector documentation describes probabilistic outcomes and warns that substantial rewriting or translation can reduce confidence. Treat an uncertain outcome as unresolved. Read the text detection limits.

Two tools can give different answers because they examined different evidence. Suppose a Unicode inspector finds no unusual characters while an AI classifier gives a high score. The first observation concerns characters. The second concerns learned writing patterns. Neither observation supplies the missing result of a compatible statistical watermark check.

Avoid averaging unrelated percentages or calling the tools a majority vote. Write down each finding separately, then ask whether any finding answers your original question. For disputes about authorship, retain drafts and revision history. Our AI detector accuracy guide explains why a classifier flag needs context.

Keep an evidence worksheet

A short record makes a result easier to revisit. Fill in the fields below while the file and verification response are still available. Leave unknowns as unknowns instead of reconstructing a history from the score.

The following examples are hypothetical. They illustrate how to record evidence; they are not tests performed by Zyphh.

A colleague sends an image called launch-poster.png. You ask whether the graphic involved Google AI, preserve that file, and save the verifier's response. You also note that you received it through a messaging app and do not know its earlier history. If you later receive a larger source file, give its check a separate entry.

For a copied paragraph, your worksheet might say: "Unicode scan found two nonbreaking spaces. Statistical watermark not checked. Original draft requested." That wording keeps an observed character finding separate from the unanswered question about the writing process.

Choose the next step from the missing evidence

Stop when another run would answer the same limited question. If a report concerns the wrong format, choose the relevant check. If it concerns the right format but leaves the origin unresolved, seek the source file or generation record.

For a document containing both prose and an illustration, maintain two entries. A finding about the illustration does not transfer to the prose beside it. If you need to explain the document to someone else, describe which parts you checked and which parts remain unverified.

When sharing the result, attach the response and identify the checked file. If the caption makes a claim about a real event, give the evidence for that claim its own place in the record. A reader should be able to see what you verified without treating the entire document as settled.