What Is SynthID? How Google’s Invisible AI Watermark Works

Suppose you spend months writing your dissertation, then ask Gemini to make the final chapter sound more academic. The result is smoother, the grammar looks better and the file is still saved under your name. But what if the newly generated wording carries a statistical watermark that survives when you paste it into Word?

That is not the same as saying your university can automatically read a hidden code from the file. And it certainly does not mean every edited sentence is watermarked. But Google has already deployed an invisible text-watermarking system called SynthID, and the distinction between your original wording and text generated during editing now matters.

So what is SynthID, how does it survive copying, who can detect it, and why could your final editing choice affect the provenance of an otherwise original dissertation? Here is what Google’s published technology and current verification tools actually tell us.

What Is SynthID?

SynthID is Google DeepMind’s technology for embedding invisible signals into supported AI-generated text, images, audio and video. A compatible detector can later look for those signals as evidence that participating AI systems produced or altered the content. Google describes it as a transparency tool, not a universal test of whether something is human-made.

For text, the watermark is not a secret character, an account number, a student identifier or a Word-document property. It is a statistical pattern introduced through the words or word pieces the model chooses while writing. That distinction is critical: copying a paragraph into a different document does not, by itself, change the underlying sequence of words.

Google’s official SynthID overview covers all four formats and confirms its use in the Gemini app and web experience.

Does Gemini Put a SynthID Watermark in Text?

Yes. Google says that text generated through the Gemini app and web experience uses SynthID watermarking. This is an actual deployment, not merely a proposed future feature. The presence and detectability of a signal in any particular passage can still depend on the model, output characteristics and the amount of generated wording.

Students sometimes imagine a watermark as something added to the original file, much like an image credit. With SynthID-Text, the pattern is produced as the answer is written. If you move unchanged generated wording into a dissertation, you have moved the wording that may carry the signal.

There are limits, however. Google does not claim that every short answer will produce a reliably detectable watermark, or that the absence of a signal proves human authorship. For a wider comparison of provider coverage, see Vappingo’s guide to text watermarks in ChatGPT, Claude and Gemini.

How Does the SynthID Text Watermark Actually Work?

A language model writes by choosing one token at a time. A token might be a word, part of a word or punctuation. For each next token, the model has a range of plausible choices. SynthID subtly influences those choices according to a watermarking scheme, producing a pattern that a compatible detector can evaluate across a passage.

Google’s peer-reviewed study, published in Nature in October 2024, describes SynthID-Text and a live evaluation involving nearly 20 million Gemini responses. The researchers reported no meaningful degradation in user-rated response quality. That study supports the feasibility of watermarking text at scale; it does not establish perfect detection for every length, language, writing style or editing workflow.

What Types of Content Can SynthID Watermark?

SynthID is a family of provenance technologies, but the signal is implemented differently according to the medium. The text version is not the same mechanism as the watermark inside an image or audio recording.

Google reported on 29 October 2026 that SynthID had been used across more than 180 billion images and videos and 240,000 years of audio. Those figures measure media watermarking scale, not a total number of watermarked student essays or Gemini text responses. See Google’s October announcement.

Does Copying Gemini Text Into Word Remove SynthID?

No—not simply because you copied and pasted it. If a watermark is represented in the statistical pattern of the generated text, switching from Gemini to Word, Google Docs or a PDF does not automatically change that pattern. Changing fonts, margins, headings or the document name also does not rewrite the paragraph.

That does not mean every copied excerpt will be detectable. Short passages, repetitive or highly constrained wording, and later edits can make verification less reliable. Copying and pasting alone, though, should not be mistaken for a provenance reset.

What Happens to SynthID When You Edit the Text?

Editing affects the sequence a text-watermark detector evaluates. Google says SynthID-Text can remain useful after some changes, such as cutting a passage, replacing a few words or mild paraphrasing. It also warns that extensive rewriting or translation can greatly reduce detector confidence.

That makes the answer different from a simple yes or no. A watermark may survive a modest edit; it may become too weak to detect after major changes. Neither outcome reliably tells you who originally wrote the passage, or whether its use met the university’s assessment rules.

Google also notes that watermarking is less effective in factual or highly constrained text because the model has fewer reasonable token choices. The official SynthID Text developer guidance describes these technical limits.

Can Gemini Proofreading Watermark Writing You Created Yourself?

Potentially, yes—but it depends on what Gemini actually generates. Suppose you write every word of your dissertation, then upload a completed chapter and ask Gemini to make it more concise, fluent and academic. The model may respond with a newly generated version of the chapter. If that output comes through a watermarked Gemini text route, it may contain a SynthID signal even though the research and original draft were entirely yours.

This is the distinction students often miss. There is a meaningful difference between Gemini identifying possible errors and Gemini producing substitute prose. The second process involves new model-generated wording, which can introduce provenance questions as well as potential changes to academic meaning.

However, it would be misleading to claim that correcting one comma or changing two words necessarily produces a detectable watermark in the final dissertation. A handful of corrections provides little generated material for a statistical checker to evaluate. Google has not published a universal threshold that guarantees detection of minor proofreading changes.

Importantly, a human editor cannot retroactively certify that earlier wording contains no watermark, and no editing service should promise to make previously AI-generated text undetectable. The benefit of avoiding a final generative rewrite is that you do not add another round of AI-produced wording at submission time.

Three AI-Editing Scenarios—and Why They Differ

The following examples are illustrative, not laboratory measurements. Actual watermark detectability varies with the passage and system.

The responsible choice is determined first by the assessment rules and by whether the final work remains accurate and yours—not by trying to predict whether a detector will notice.

Can Anyone Detect a SynthID Text Watermark?

Not with a universal public text checker for arbitrary Gemini passages. Google has released SynthID Text code and developer tooling, including probabilistic detection methods. A verifier still needs the appropriate watermark configuration or trained detector for the scheme being checked. Open-source implementation does not give everyone a key to every production watermark.

Google’s developer documentation describes different choices for deploying watermark detectors: private, accessible through an API, or publicly released. It also describes three possible detection outcomes: watermarked, not watermarked and uncertain. A result depends on the specific system, not on an invisible barcode that anyone can read with complete certainty.

The official developer guide explains this distinction.

Can a Lecturer Check an Essay for SynthID in Gemini?

Google’s currently documented public Gemini verification feature checks images, video and audio—not pasted essays for a SynthID text watermark. This is an important detail because Google expanded its public SynthID Detector in October 2026, making verification more accessible globally in English for supported media.

The expansion is real, but it does not mean any lecturer can paste a student’s dissertation into Gemini and receive a definitive official text-watermark verdict. Google continues to describe text watermarking and text-verification capability separately from the public media-checking workflow.

Check Google’s current Gemini verification help and its October 2026 SynthID Detector update for the supported media types.

How Is SynthID Different From an AI Writing Detector?

An ordinary AI-writing detector, such as the classification feature discussed in Turnitin’s guidance, examines completed text for patterns associated with AI-generated or AI-altered writing. A SynthID verifier looks for a deliberate signal produced by a compatible system during generation. They are different kinds of evidence.

Our AI watermarks vs. AI detectors guide explores this distinction in greater detail. A high Turnitin AI-writing score should not be described as a verified SynthID watermark, and a failed SynthID check should not be described as proof that no AI was used.

Does a SynthID Watermark Prove Academic Misconduct?

No. A verified watermark can help establish provenance; it does not decide whether an assessment rule was broken. Even if a compatible signal suggests that Gemini generated part of a passage, the result does not, on its own, identify the student, reveal why Gemini was used, show how much independent research went into the work, or determine whether that use was permitted and declared.

For example, a student might have used AI within an explicitly permitted editing workflow. Another student might have generated an entire literature review when their assessment prohibited that assistance. The same general fact—AI involvement—could have very different consequences under the applicable rules.

Google’s technology should therefore be treated as one possible source of evidence, not as a misconduct verdict. The distinction also applies to other providers. OpenAI’s October 2026 textGrain announcement explicitly says a text watermark cannot measure the human contribution or identify a person and is separate from Google’s SynthID text system.

How Might a University Investigate AI Use?

A university does not need a public watermark-checking interface to raise an academic-integrity concern. It may look at drafts, source records, the written assessment policy, declarations and the student’s ability to explain the work. Some procedures permit an authorship discussion or investigatory viva.

UCL, for example, describes rules for Language and Writing Review and warns that substantive changes made by software or third parties can cross the boundary into misconduct. Its guidance recognises differences between a few spelling corrections and changes that determine the intellectual content of the submission. Read UCL’s guidance.

Students should not assume that every university uses an AI detector, that all institutions can verify provider watermarks, or that professional proofreading is permitted without limits.

What If Your Gemini Use Was Allowed?

If your assessment permits the kind of assistance you used, the most valuable step is to keep the process understandable. Save relevant drafts and research notes, retain the assessment instructions, and follow the required AI-use declaration format.

Be specific: “I used Gemini to identify unclear sentences in Chapter 3; I made the final revisions myself” communicates a different process from “I asked Gemini to rewrite the discussion chapter and incorporated the output.” Do not claim that you wrote every final word without assistance if you actually accepted generated replacement wording.

For an adaptable disclosure, see Vappingo’s sample dissertation AI-use declarations. A declaration describes what happened; it does not automatically make otherwise prohibited assistance acceptable.

Why Choose Human Proofreading for the Final Draft?

By the time a dissertation reaches its final editing stage, the student has usually completed the demanding intellectual work: choosing sources, carrying out research, analysing results and building an argument. Asking a generative model to rewrite the document at this point can introduce fresh wording that needs to be checked for both academic accuracy and provenance.

Vappingo takes a different approach. Its dissertation proofreading service uses qualified human proofreaders to review language, consistency, references, numbering and presentation, returning a tracked version and a clean copy. Where a discrepancy depends on your methods or findings, the editor can flag it for your decision rather than inventing an answer.

Human proofreading does not add newly AI-generated replacement prose as part of Vappingo’s proofreading process. That is a meaningful distinction from asking Gemini to regenerate your final chapters. It does not remove earlier AI watermarks, conclusively prove authorship or guarantee that a university will approve every level of editing.

If you want to see the editorial approach before committing, you can also request a free 300-word proofreading sample using a representative passage from your own work.

Is Professional Proofreading Allowed at University?

That depends on the university, the programme and the particular assessment. Some institutions permit limited grammar, punctuation and presentation corrections; others specify stricter rules or restrict assistance when language proficiency is being assessed.

UCL’s 2026–27 guidance, for instance, identifies minor surface corrections as generally acceptable while warning against interventions that determine substantive content. This applies to both human reviewers and software. Consult the current guidance and, above all, your own department’s instructions.

Before hiring any proofreader, check what can be changed, whether third-party assistance must be declared, and how edits should be documented. Keep the original draft and the tracked version so the scope of editorial input can be explained honestly if necessary.

What Does SynthID Mean for AI Provenance?

SynthID shows that some AI-generated content can carry invisible provenance information in the content itself—not just in a file’s metadata. As Google expands public media verification, the distinction between a deployed watermark and a publicly available text checker becomes more important.

Other providers are building separate technologies. In October 2026, OpenAI announced textGrain for eligible EU ChatGPT and Codex text, with opt-in watermarking for supported API models globally. OpenAI also uses SynthID for certain generated images and audio, but textGrain is not SynthID-Text. Provider, product, region, model, content type and date all matter. See OpenAI’s 5 October announcement.

None of this means every AI-generated sentence is identifiable or that any single detection result should decide an academic case. The technology is developing, and its limitations—particularly short passages and subsequent editing—remain substantial.

Frequently Asked Questions About SynthID

The Main Thing to Remember

SynthID can watermark text generated through Gemini, and copying that text into a dissertation does not automatically erase the signal. But watermark deployment, access to a verifier and proof of academic misconduct are three separate questions.

The important practical distinction is between using AI to point out possible problems in your own work and asking it to generate an improved version of that work. The latter can introduce new AI wording—and with supported systems, a new provenance signal—right at the point you want the clearest account of how the dissertation was produced.

For a broader explanation of what universities can and cannot establish, read Can a University Actually Prove You Used AI?