Every figure, date, name and mechanism the finished picture puts on screen, chased to a primary source. Checked 14 August 2026.
Claude output carries machine-readable marking, and the marking differs by output type. Text carries an embedded watermark woven into the text itself, which travels with the text when it is copied and pasted elsewhere. Supported files carry signed provenance metadata instead. Source: Anthropic, How Claude marks AI-generated content. https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
File provenance uses C2PA, and the supported formats are SVG, PNG and JPG. The metadata follows the Coalition for Content Provenance and Authenticity open standard, which is used across the industry to record where content came from. Source: as above.
Claude models launched in the EU on or after 2 August 2026 support machine-readable marking at launch. Support for models released before that date is described as in progress. Source: as above.
The marking applies wherever Claude is offered, worldwide, not only in the EU. This is the point the video makes about a compliance system built for one jurisdiction reaching everyone, and it is the provider’s own description rather than an inference. Source: as above.
Marking spans Claude Platform (the API), Claude, Claude Code, Claude Cowork and Claude Tag. Source: as above.
A detected mark does not prove Claude authored the content, and an absent mark does not prove the content is human. Anthropic states that a detected mark indicates the content may have been processed by Claude, not that Claude wrote it, and that marks can be lost through editing, paraphrasing, format conversion or screenshots. Source: as above. This caveat is load-bearing for the video’s argument that a detector flattens the difference between writing something and touching it, and it is the provider’s own wording rather than the video’s characterisation.
Article 50(2) of the EU AI Act requires providers of AI systems generating synthetic audio, image, video or text content to ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The obligation is qualified as applying where technically feasible, and excludes assistive editing functions and systems authorised by law to detect or prevent criminal activity. Source: European Commission AI Act Service Desk, Article 50. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
The four media types named on screen — audio, image, video, text — are the four the article names. Source: as above.
Article 50(1) covers informing people they are interacting with an AI system, and Article 50(4) covers deployers disclosing deepfakes and AI-generated text published to inform the public on matters of public interest. The video stays on paragraph 2; the neighbouring paragraphs are noted here only so the citation on screen is not read as covering the whole article. Source: as above.
Text watermarking works by modulating the probability of the tokens a model generates, rather than by inserting a stamp or a hidden character. The system adjusts the probability scores assigned to candidate next tokens; the resulting pattern across the model’s actual choices is the watermark, and it is detectable afterwards without being visible to a reader. Source: Google DeepMind, Watermarking AI-generated text and video with SynthID. https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/
Longer output is easier to mark and detect. Detection works best when a model generates longer responses with genuine variety, because a longer passage offers more token choices to adjust. Source: as above.
Heavy rewriting and translation weaken the signal. Confidence scores can be greatly reduced when watermarked text is thoroughly rewritten or translated into another language. Source: as above.
Short factual answers are harder to watermark. A prompt with only one correct answer leaves little room to adjust token probabilities without damaging accuracy, so there is less signal to embed. Source: as above.
Training generative models on recursively generated data degrades them, and the damage falls first on the tails of the original distribution. The paper distinguishes early collapse, where the model drifts from the true distribution as errors accumulate, from late collapse, where low-frequency events disappear permanently. Indiscriminate use of model-generated content in training causes irreversible defects. Source: Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal, AI models collapse when trained on recursively generated data, Nature 631, 755–759 (2024). https://www.nature.com/articles/s41586-024-07566-y
This is the source for the beat that draws a distribution losing its tails and drifting off the human one, and for the claim that the risk is volume of synthetic data rather than any single AI-written paragraph.
The named consequences for developers — repository scanning for AI-generated code, maintainers rejecting AI-assisted contributions, employers measuring developers by AI-origin signal, legal teams using watermark evidence in copyright disputes, future coding models downweighting model-written code — are stated as open questions about what could follow, not as things that have happened. Nothing on screen presents any of them as a documented event.
A review pointed out that drawing them as ordinary working screens made them read as documented anyway, whatever the narration said, so the two beats that mock up a system — the repository scan with its maintainer queue, and the performance dashboard with the watermarked exhibit — are now held inside a dashed frame, carry the question the narration actually asks, and are stamped “not yet · an open question” for the whole beat. The drawing and the sourcing say the same thing now.