Anthropic plans to add machine-readable watermarks to content processed by Claude, while Twitch now offers an opt-out from Amazon’s Artificial Intelligence (AI) training. At the same time, a typeface called ShieldFont aims to frustrate automated collection of website text. These changes affect you if you edit writing with AI, stream online, or publish your own work on the open web.
Why control over content is growing
Generative AI, meaning systems that create text, images, audio, or video, requires large quantities of training material. That material often comes from public websites, platform posts, and media archives. Creators therefore face two connected issues: whether a provider can train on their work, and whether other people can later recognize AI involvement in a finished piece.
The latest developments offer three different answers. Anthropic is focusing on labeling, Twitch has introduced a choice after training has already taken place, and ShieldFont attempts technical disruption. None gives you complete control: a watermark does not prevent use, Twitch says its opt-out applies to future training, and scraping bots may eventually detect or bypass a defensive technique.
Scraping is the automated extraction of website material by software. According to t3n’s account of ShieldFont, these tools often collect text directly from a page’s Hypertext Markup Language (HTML), the underlying code used to structure web content. This can sometimes happen even when a website operator has explicitly objected to scraping.
What Claude’s watermarks are meant to reveal
Anthropic intends to mark content that its Claude models process, not only material they generate from scratch. According to the company, text outputs will carry embedded watermarks that people cannot see. Other generated files will include digitally signed provenance metadata where supported. Provenance metadata records information about a file’s origin and processing history.
Anthropic attributes the change to the transparency code connected to the European Union’s Artificial Intelligence Act, commonly called the EU AI Act. According to Ars Technica’s report on the global rollout, all new models will apply marks from their first day of release, not only models offered in the EU. The change therefore also matters to users in Switzerland, even though the rule prompting it comes from the EU.
This broad approach has a significant weakness. The model apparently cannot reliably distinguish between fully generated writing and a minor correction. If you ask Claude to fix a comma or smooth one sentence, the resulting text may receive the same type of mark as an entirely generated paragraph. Ars Technica notes that the EU framework makes exceptions for routine editorial assistance, but Anthropic’s stated policy goes further.
The effectiveness of the system remains uncertain. Anthropic has not yet released its planned detection tool, so the watermarks cannot currently be tested independently. The company also says they will not work on certain platforms or features. For non-text files, Anthropic plans to use Coalition for Content Provenance and Authenticity (C2PA) metadata, a standard for recording the origin of digital content, where the relevant file format supports it.
Responses to the plan conflict. TechCrunch documented complaints from Claude users who fear consequences at school or work. Other commenters argued that labeling is justified when someone submits unchanged AI output as their own work. Paraphrasing the result or passing it through another AI service might alter traces, but the supplied reports do not provide independently verified evidence that this reliably defeats the marks.
What the Twitch opt-out actually covers
Twitch now provides a generative AI training switch in its security and privacy settings. Turning it off means that streams, videos on demand, clips, stream chats, and images and text on your channel will not be used for future training of Amazon models. These are models designed to generate or synthesize text, audio, images, or video.
The opt-out is not a general rejection of all AI processing. Automated captions, recommendations, AutoMod, and tools related to safety, discovery, or monetization can continue to use AI-supported features. The Verge highlights another important limit: if you post in someone else’s stream chat, that channel owner’s preference determines whether the message may be used for training.
Multiple reports agree that participation is enabled by default. Ars Technica also reports that Amazon appears to have used Twitch content for AI training for at least several years; a Twitch executive acknowledged the general practice in 2024. The available information does not establish exactly how long this occurred or which older content entered particular models.
TechCrunch describes substantial opposition from the Twitch community. During a stream, a Twitch executive said, in substance, that almost nobody would participate if training required an active opt-in. He was also unable to answer a user asking whether that person’s videos had already been used. The new control therefore offers a choice about the future without providing a clear record of the past.
For a streamer, the practical result is specific: she can stop her future broadcasts and clips from being used to train Amazon’s generative models without losing captions or safety features. A viewer, by contrast, cannot independently control training use of a chat message posted on someone else’s channel. That distinction is easy to miss when a setting is described simply as an opt-out.
What ShieldFont tries to do about scraping
ShieldFont takes neither a legal nor a platform-based approach. It changes how web text is technically presented. Designers Isaque Seneda and Gabriel Abrucio use ligatures, a long-established font feature that normally combines particular letters into a more readable shape. In ShieldFont, ligatures instead help people see the correct words on screen while different words remain in the underlying HTML.
A basic scraping bot that downloads only raw text consequently receives a subtly altered and potentially useless version. A public article could appear correct to readers while a bot collects unrelated word substitutions. Ars Technica’s explanation of the ligature technique presents it as a way to reduce the training value of material when a stated refusal is ignored.
The replacements cannot be too obvious or too easy to reverse, according to the report. Pure nonsense might be detected and discarded by a filter, while simple synonyms or antonyms could potentially be translated back by a more capable scraper. ShieldFont is therefore closer to a technical arms race than a permanent block.
For an author running her own website, this is a different tool from Twitch’s account setting. She controls how her site is rendered and can try to make the raw text collected by simple bots unusable. She generally lacks that option when publishing on someone else’s platform and must instead rely on its settings and policies. The reports supplied do not yet contain independent long-term testing of ShieldFont against adapted scrapers.
Pros and Cons of the new controls
Pros:
- Greater transparency – Claude’s watermarks may provide a machine-readable indication that content was generated or edited with AI.
- A concrete choice – Twitch members can disable the use of their channel content for future generative AI training.
- Technical self-defense – ShieldFont gives independent web publishers another possible defense against basic scraping bots.
- Features remain available – Twitch’s opt-out does not broadly disable automated captions, recommendations, or safety tools.
Cons:
- Overbroad labeling – Claude may treat a minor correction and a fully generated passage in the same way.
- Default participation – You must actively opt out on Twitch even though the details of previous training remain unclear.
- Limited reach – Messages posted in another channel’s chat are governed by that channel’s preference.
- An ongoing arms race – Adapted scrapers may eventually recognize or bypass defenses such as ShieldFont.
What this means for you
If you are a beginner, start by checking which platforms store your content and what controls they provide for AI training. On Twitch, the relevant switch appears under Security and Privacy. If necessary, keep a screenshot of your choice because settings and their descriptions can change.
Step 1: Check your Twitch setting
- Open Twitch settings and go to Security and Privacy.
- Find the control for generative AI training.
- Turn it off if you do not want your channel content used for future training of Amazon models.
- Remember that this does not disable Twitch’s other AI-supported features.
If you use Claude for education, work, or publishing, distinguish assistance from direct reuse. One concrete workplace example is summarizing a long transcript: using that summary to navigate the original and checking every relevant point is different from publishing the generated wording unchanged. The same distinction applies to a student who asks Claude to reorganize a paragraph and then inserts the result directly into an essay.
Step 2: Keep your editing process traceable
- Retain your own draft and original sources for significant pieces of writing.
- Verify AI summaries against the source material rather than accepting them unreviewed.
- Expect even small Claude edits to potentially receive a machine-readable mark.
- Do not assume that an invisible watermark will survive or remain detectable on every platform.
Advanced users who operate their own websites can also examine technical scraping defenses. ShieldFont is aimed at publishers who want human readers to see one accurate version while basic bots collect altered raw text. Before using it, they would still need to test readability, accessibility, and its actual effect on the relevant scrapers; the supplied reporting does not establish universal protection.
For people in Switzerland, the practical product changes are much the same as elsewhere: Anthropic says its new-model watermarks will roll out globally, and Twitch presents the opt-out as an account setting. That does not amount to a complete assessment under Swiss privacy law. The main limitations described by the sources remain technical: no retroactive protection, uncertain detection, and dependence on choices made by other platform users.
Watermarks, opt-outs, and scraping defenses shift some control toward the people creating content, but they do not resolve the underlying dispute. Anthropic may label too broadly, Twitch is adding a choice after training has already occurred, and ShieldFont may prove to be only a temporary obstacle. The unresolved risk is that content has already been used or that new collection methods will find a route around the available controls.
Sources
- Claude’s new Scarlet Letter watermark is invisible—for now – Ars Technica, 2026-08-13
- Some Claude users are mad that Anthropic’s new watermarks will catch them using it at their jobs, classes – TechCrunch, 2026-08-12
- Twitch streamers can now opt out from training Amazon’s AI – The Verge, 2026-08-12
- Twitch content has trained Amazon AI for years, but users can opt out now – Ars Technica, 2026-08-12
- Amazon will train on Twitch streamers’ content by default, unless they opt out – TechCrunch, 2026-08-12
- The web’s newest weapon against AI scrapers is a font – Ars Technica, 2026-08-12
- Kampf gegen Scraping: Wie eine Schriftart Wörter tauscht, um KI zu verwirren – t3n, 2026-08-13
