RIP Claude

What happened

I'll read the source article first so the paragraphs stay grounded in what it actually says, then write the post in plain prose.The Rands piece is the source. I'll pull Anthropic's own watermarking notes next so the technical paragraphs stay accurate.Michael Lopp, writing as Rands on Rands in Repose, posted RIP Claude on August 13, 2026. He has used Claude for editing, research, and data wrangling on an upcoming book and on the next edition of Managing Humans, while writing the books himself. After Anthropic announced text and image watermarking, his own research left him furious. As he has it, models born on or after August 2, 2026 will carry watermarks, older models will be retrofitted by December 2, 2026, the change is already in progress with no per-model disclosure, and Anthropic has not promised notice when marking turns on. The post was listed on Hacker News with 4 points and 0 comments.

Lopp’s technical read, aligned with Anthropic’s help article on how Claude marks AI-generated content, is that the mark is not a user setting. It is applied server-side, below the model, worldwide, with no opt-out on the API, on any plan tier, or on enterprise. Anthropic says new models launch with machine-readable marks across the Claude Platform API, Claude, Claude Code, Claude Cowork, and Claude Tag, including when supported models are reached through AWS, Google Cloud, or Microsoft Foundry. The text watermark is woven into the wording so it copies with the paste and can survive some editing. Files such as SVG, PNG, and JPG get signed C2PA provenance metadata. Anthropic states that proofreading, translation, summarization, and file conversion can leave a mark even when the ideas started elsewhere. Lopp adds critique and is specific: if he asks Claude to review a piece and pastes part of the reply into the draft, that passage can be marked. He also records that no public detector exists yet and that only Anthropic can check. Anthropic says it will support detection and will publish the documentation later.

The technical detail

The builder problem is the width of processed by, not a watermark on fully generated essays. Most engineering writing already runs through an assistant for a pass: a proofread of an RFC, a translation of a comment, a summary of a thread, a critique of a design. If those passes brand the output, the human source of record can pick up a Claude mark from a fragment the author never meant to keep. Lopp draws a hard line. He uses Claude Code a huge amount for tools and applications and is not claiming he wrote that code, so watermarking it does not offend him. His writing is his work, and he calls the current strategy aggressively writer-hostile. Anyone who mixes model text into docs, tickets, blog posts, or customer email now has to treat paste as a provenance event.

Why it matters for builders

The competitive frame is the EU AI Act. Anthropic signed the Article 50(2) Code of Practice on Transparency of AI-Generated Content, in force from August 2, 2026, which requires providers serving the EU to mark generated or edited content so other systems can identify it. TechCrunch notes that Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia made the same class of commitment, and that Anthropic is applying the mark globally rather than only in Europe. That is a compliance shortcut with a product cost: writers outside the EU still inherit a writer-hostile default. Lopp has been a Claude Max user since the program appeared. He can guess the regulatory threats Anthropic is facing. He still describes the rollout as a tone-deaf, clumsy, and alarming opening salvo that puts the entire burden on him and none of it on them.

Market and competitive context

What he is doing next is operational. He is building gates so Claude text and images never enter his projects, which he notes is good practice anyway. He is researching a migration off Claude for all writing-related work. He is getting into the mindset to cancel Claude Max. The dates and artifacts to watch are specific. December 2, 2026 is Lopp’s retrofit deadline for older models. Anthropic’s detector documentation is still forthcoming, so the first checker anyone but Anthropic can run will decide whether this is a transparency tool or a private ledger. Other Code signatories have not all shipped the same proofread-and-critique scope in public. If they keep marks on fully generated text only, writing workloads will leave Claude first while Claude Code can stay.

What to watch next

The open risk is that the mark is a weak legal and editorial primitive dressed as a strong one. Anthropic itself says a detected mark only indicates the content may have been processed by Claude, not that Claude authored it, and that missing marks prove nothing: pre-mark models, heavy edits, paraphrase, translation, very short passages, stripped file metadata, and unsupported surfaces all go clean. That is a bad signal for publishers, academic offices, or contracts that treat AI-generated as a yes-or-no. It is also a bad signal for Lopp’s critique-paste case, where a few borrowed words could brand a chapter he typed. C2PA on images at least lives in metadata you can inspect or strip. Statistical text watermarks have always been fragile on short spans and light edits. Until a public detector exists, writers cannot audit their own corpus, and Anthropic retains the only key.