The internet’s response to AI slop has mostly worked backward. Platforms scan content after it is already uploaded, then try to guess whether a machine wrote it, an inherently probabilistic bet that it is losing ground as synthetic content gets cheaper to produce. Anthropic, Apple, Canon and Google are each building a different layer underneath that guesswork: proof of where content came from, established the moment it is created, not detecting after the fact.
Anthropic began embedding invisible, machine-readable watermarks into text from new Claude models launched on or after Aug. 2, 2026, the artificial intelligence company confirmed in an updated help center article.
Canon and Google Embed a Signature the Instant a Photo Is Taken
For photos, the fix starts in the hardware, not in software that runs after the fact. Canon launched its Authenticity Imaging System on May 11. It is the first manufacturer-operated service carrying C2PA Content Credentials, a cryptographically signed record of a file’s origin and edit history, from capture through publication. The rollout starts in Europe, the Middle East and Africa, with Reuters as its pilot partner, C2PA Viewer reported.
Google took a similar approach at the consumer level. Every photo taken with a Pixel 10’s native camera app is signed by default with C2PA credentials, using hardware-backed keys in the phone’s Titan M2 chip. That makes the Pixel 10 the first mainstream smartphone to sign every photo, not just the ones edited with AI, C2PA Viewer reported separately.
Apple is taking a different route than the C2PA standard Canon and Google use. Code in iOS 27 beta 5 points to a feature called Apple Reference Image, authenticating a photo using data tied to the specific iPhone camera hardware that captured it, 9to5Mac reported. The feature would require a separate capture mode, so ordinary photos would carry no provenance data. Verification happens through Apple’s Private Cloud Compute, where Apple says it never sees the raw photo, only sensor data and metadata. It has not shipped.
The distinction between camera-level systems and Anthropic’s text watermark is structural. A camera signs an image the instant light hits the sensor, before software touches the file. A text watermark cannot prove a person wrote something, since people already use Claude to edit or polish writing that started as their own. Anthropic warns that a watermark hit only signals that content may have passed through Claude. It is not proof of authorship, according to TechCrunch.
Guessing After the Fact Has Limits
That distinction changes what platforms can detect after content is already uploaded. More than 40% of LinkedIn posts longer than 250 words are fully AI-generated, according to detection company Pangram’s analysis of over 1 million posts, PYMNTS reported. LinkedIn responded in July with a “Seems like AI slop” button, and Substack launched a Pangram-powered detector the same month. LinkedIn’s classifiers correctly flagged generic AI content roughly 94% of the time in early testing, LinkedIn said, with no published data on false positives.
Neither approach solves the problem alone. Metadata and embedded credentials routinely disappear during screenshots or re-uploads, and Nikon suspended its own C2PA signing service in 2025 after a security flaw forced it to revoke every certificate issued, according to C2PA Viewer. After-the-fact detection is becoming one layer of defense, not the whole system, as each company works to establish a chain of custody that starts at creation.
