Google's SynthID can embed a 136-bit payload into a 512×512-pixel image — enough for a 64-bit database identifier with 72 bits left over for error correction. That identifier can map to your name, your IP address, your physical location, your political affiliation. Google calls this a "watermark." Brandon Thomas calls it a spymark, and the linguistic reframing is the entire point of the argument. The technical capability is not new. Printer tracking dots — yellow microdots encoding serial numbers and timestamps into every page — date to the 1980s. The open-source tool audiowmark has been hiding 128-bit AES-encrypted payloads in audio since 2018. What has changed is scale and ambition. SynthID operates across images, audio, text, and video. OpenAI and others are building equivalent systems. The infrastructure for universal content tracing is being assembled under the benign label of "AI content identification." The core technical distinction Thomas draws is between metadata you can see, edit, and strip — EXIF tags in photos, ID3 tags in MP3s — and signals embedded in the actual pixel values, audio waveforms, or word-choice patterns of your content. The latter survive metadata removal. They survive some edits. They are invisible to you. And they encode identifiers that are useful only to the entity that embedded them. Text spymarking is particularly insidious. SynthID steers word choices during generation to create a detectable statistical pattern that can encode a tracking payload. You cannot inspect your own words for the signal. The medium is the message, and the message is your database ID. Thomas's argument is fundamentally a linguistic one, not a technical one. The word "watermark" carries connotations of authenticity verification and ownership assertion — banknotes, stock photos, copyright overlays. These are visible, understood, and generally benign. Collapsing invisible tracking signals into the same term obscures the surveillance function. "Spymark" forces the privacy concern into the foreground by naming what the technology does to you, not what it does for the platform. The twenty-year scenario Thomas sketches is not speculative. Device attestation plus account-linked spymarks plus social graph analysis equals a system where every shared JPEG or tweet can be traced back through the chain of human hands that touched it. This is not watermarking. This is supply-chain surveillance for ideas. The article is a manifesto for a word, not a technical disclosure. No new vulnerabilities are revealed, no new defenses proposed. The contribution is rhetorical clarity — the insight that the privacy community's failure to communicate is partly a vocabulary problem, and that the right name can do the work of a thousand explainers.
