Why Google’s SynthID Won’t Save AI Detection (and What It’s Actually For)

Every few months a headline promises that AI detection is finally solved. The latest hope is SynthID, Google’s watermarking system, which can mark text and images its own models produce. It’s clever engineering, and it does a real job — just not the one most people assume.

Here’s the gap. SynthID tells you whether Google’s model made something. It doesn’t tell you whether the essay on your desk was written by a person, and it never will.

Key takeaways

• SynthID embeds a hidden, machine-readable mark in content Google’s own AI models generate. It tags output; it doesn’t analyze arbitrary text.
• The watermark only exists if the model added it. Text from a non-Google tool, or text that’s been paraphrased or retyped, has nothing to detect.
• The text version is the fragile one. Editing, translation, and rewording can wash the pattern out.
• It’s a provenance signal for platforms and policymakers, not courtroom-grade proof that one specific person used AI.
• Detection stays probabilistic. Write with specificity and rhythm, and judge work on substance, not a number.

What is SynthID watermarking?

SynthID is a tool from Google DeepMind that stamps an invisible signature into AI-generated content at the moment it’s created. For images, audio, and video, it nudges the underlying data in ways a person can’t see or hear but a matching detector can read. The text version was published in a 2024 *Nature* paper (Dathathri et al.) and open-sourced through Hugging Face that October, so this isn’t vaporware — it’s running inside Gemini and shared with other developers.

The text version is the interesting one for anyone worried about AI detection, so let’s be precise about how it works.

How the text watermark actually works

When a language model writes, it doesn’t pick the single “best” next word. It rolls dice over a ranked list of plausible options. SynthID gently loads those dice. Using a secret key, it nudges the model toward certain word choices in a pattern that’s statistically detectable later but reads as completely natural to you.

Run the finished text back through SynthID’s detector with the same key, and it measures whether those biased choices show up at a rate you’d never get by chance. The detector can land on three answers, not two: watermarked, not watermarked, or uncertain. A strong signal means it was almost certainly marked. A weak or absent one tells you nothing — which is the whole problem.

Notice what this requires: the watermark has to be *baked in during generation*. SynthID isn’t reading your prose for robotic tells. It’s checking for a fingerprint that only its own generation process could have left.

Why it won’t “save” AI detection

The fantasy is a universal scanner: paste any text, learn whether a machine wrote it. SynthID is structurally incapable of being that. Three reasons.

1. It only catches its own output

SynthID recognizes the SynthID mark, full stop. Text from ChatGPT, Claude, Llama, or any of the dozens of open models that don’t use it carries no signal. The detector returns nothing — not because the text is human, but because there’s no fingerprint to find. A tool that’s blind to most of the AI on the market can’t be the answer to “was AI used.”

This is the opposite failure mode from a statistical detector. A tool like GPTZero will analyze *anything* you give it, but it guesses, and it’s wrong often enough that OpenAI pulled its own classifier in July 2023 for poor accuracy. A Stanford study that same year found such detectors biased against non-native English writers. So SynthID trades broad-but-unreliable reach for narrow-but-confident. Useful. Still narrow.

2. The text watermark is fragile

Google has been refreshingly honest here. Its own documentation says detection confidence drops sharply when text is thoroughly rewritten or translated, and the text mark is weaker than the image one to begin with. Because the signal lives in word choice, anything that disturbs word choice degrades it. Paraphrase the passage, run it to Spanish and back, retype it by hand, or fold it into your own sentences, and the pattern thins out until the detector can’t find it.

Short text is a problem too. A two-sentence answer doesn’t give the watermark enough room to establish a reliable pattern. Neither does highly factual writing — a definition or a math step has so few good word choices that there’s little to nudge. The cases where you’d most want certainty are often the cases where the mark is weakest.

3. It proves provenance, not authorship

Even when SynthID fires cleanly, it tells you “this came from a Google model and still carries the mark.” It doesn’t tell you who prompted that model, how much a human edited afterward, or whether the final work is genuinely that person’s thinking. Provenance and authorship are different questions, and the disputes that actually wreck someone’s week are about authorship.

A quick scenario

Picture a professor with two flagged papers. Maya pasted a full Gemini draft and barely touched it. Jordan used a free non-Google chatbot, then rewrote every paragraph in his own voice.

SynthID would light up on Maya’s paper and stay silent on Jordan’s — even though Jordan leaned harder on AI for ideas. A statistical detector might do the reverse, flagging Jordan’s careful-but-uniform prose while missing Maya’s edited-just-enough version. Neither tool gives the professor what they actually need: a conversation about the work. That’s why these signals belong in a process, not a verdict. The same probabilistic limits show up across the statistics behind AI detectors, and watermarking doesn’t repeal them.

So what is SynthID actually good for?

Plenty, once you stop asking it to be a lie detector.

• Platform-scale provenance. Google can label AI images in Search and flag synthetic media at upload. At billions of items, even an imperfect mark on its own output is valuable for transparency.
• Fighting deepfakes and misinformation. Confirming a viral clip came from a known generator is a real public-safety win, separate from any classroom argument.
• Honest disclosure. A team that drafts with Gemini can use SynthID internally to track which assets started as AI, supporting a clear AI-use policy rather than punishing anyone.
• A nudge toward standards. The more generators that adopt watermarking, the more provenance becomes a norm. That’s a slow, partial good — and worth having.

What it isn’t: evidence to accuse a student, grounds to reject a freelancer’s invoice, or a reason to trust any single percentage on a screen.

What this means for writers, students, and teams

The ground rules don’t move. Whether the tool is a watermark or a statistical classifier, the same habits protect you and serve readers:

• Keep your trail. Drafts, version history, and notes show how a piece came together far better than any detector can rule it out.
• Write with specificity and rhythm. Concrete examples, real numbers, and sentences that vary in length read as human because they *are* human — and they survive watermark-washing because that was never the point.
• Treat scores as signals. A clean SynthID check or a low AI percentage is information, not a verdict. Anyone using either as proof is misusing the tool.

If you write with AI in the loop, the goal isn’t to dodge a fingerprint. It’s to make sure the final work is genuinely yours in voice and substance. You can read more on the PaperBleach blog for how detection works and where it breaks, or see what PaperBleach offers if you want a tool in your own corner.

Frequently asked questions

What is SynthID watermarking in simple terms?

It’s a system from Google DeepMind that embeds an invisible, machine-readable signature into content generated by Google’s AI models. For text, it subtly biases word choices into a hidden pattern as the model writes, and a matching detector checks whether that pattern is there. The catch: it can only flag content its own system watermarked at creation.

Can SynthID detect ChatGPT or Claude output?

No. SynthID only recognizes its own watermark. Output from OpenAI, Anthropic, or any tool that doesn’t use SynthID carries no signal, so there’s nothing for the detector to find. It’s a brand stamp, not a universal scanner.

Can the SynthID text watermark be removed?

It can be weakened or erased. Heavy rewriting, paraphrasing through another tool, translating back and forth, or retyping by hand disturbs the word-choice pattern. Short and highly factual text also carry a weaker mark. A determined person can defeat it — which is precisely why it isn’t proof.

Is SynthID the same as Turnitin or GPTZero?

No. Those are statistical detectors that estimate how AI-like *any* text looks and output probabilities — broad reach, but false positives. SynthID is a watermark check: very reliable when the mark is present, blind to anything that wasn’t marked. Different halves of the problem, and neither closes it.

Does SynthID mean AI writing will finally get caught?

Not the way people hope. It confirms “this came from Google’s model and still carries the mark,” and it’s useless against non-Google tools, edited drafts, and paraphrased text — the exact cases that cause disputes. It raises the floor on provenance; it doesn’t end the debate.

The bottom line

SynthID is a genuinely good tool aimed at a question most people aren’t actually asking. It marks Google’s own output for provenance. It doesn’t sniff out the edited, paraphrased, or non-Google writing at the center of every real argument over AI. The honest path is the same as it ever was: write with substance and voice, and treat any score as a signal, not a sentence.

Want to see how your draft reads before anyone else runs it through anything? Check how a draft reads with PaperBleach.