Anthropic opens its Claude text watermark playbook as ChatGPT logs macOS keystrokes

Anthropic explains the Claude text watermark and its limits

Anthropic has published a blog post walking through how its invisible watermark on Claude's text-based outputs is generated, what editing does to the statistical signal, and how the system is applied to code, in an effort to answer the questions Claude users had been raising since the company announced the change earlier in the week. Multiple outlets describe the system as creating a pattern inside Claude's responses during generation, with the goal of complying with the EU AI Act's Transparency Code, the regulation designed to identify AI-generated content.

Reports flag an important caveat for editors and developers: even human-written copy mixed into a Claude workflow could end up carrying the watermark, because the signal is attached to the model's output rather than to the author. Readers handling compliance-sensitive prose will want to know what survives a round of copy editing and what does not, and Anthropic's post is the first place that question is answered in detail.

ChatGPT Computer History turns macOS clicks and keystrokes into training data

OpenAI has begun shipping Computer History, a ChatGPT feature on macOS that records clicks, keystrokes, keyboard shortcuts and app switches through the operating system's accessibility framework, then turns the activity into a searchable memory the assistant can draw on. One outlet characterizes the move as a quieter cousin of Windows Recall, taking no screenshots, screen recordings, microphone input or system audio, with private browsing excluded. A second report adds a sharper technical detail: the memory files are stored locally as unencrypted plain text, and the feature is not available in the EEA, Switzerland or the UK, a sign that regional data-protection rules are shaping where the rollout can land.

The product framing is that Computer History lets ChatGPT learn how you work, suggest automations and pick up half-finished tasks, with your activity building a timeline the assistant can revisit later.

A paid AI text detector posts a near-perfect chatbot-text score

A week of independent testing of Pangram, a $20-a-month AI text detector, found the tool correctly identified chatbot-written content in every test, with the company claiming 99.99% accuracy. Pangram produces a percentage score for how much of a pasted block was generated by a chatbot versus written by a person, and the reviewer described dozens of experiments in which the detector did not fail at distinguishing chatbot output from human writing. The same review flagged a clear gap: the tool fails on AI images, an important shortcoming at a moment when roughly half of online articles are now machine-generated.

For editors and compliance teams, the practical takeaway is that text-side provenance tooling has matured faster than image-side tooling, so workflow policies need to treat the two formats differently.

The bigger picture: provenance, capture and verification are reshaping text work

Taken together, the three stories trace the shape of a text stack that is being rebuilt around traceability. On the generation side, vendors are now embedding machine-readable signals into AI-written prose so downstream tools can tell model output from human writing, and that signal travels into code as well as natural language. On the capture side, assistant products are beginning to record the fine-grained keystroke-level activity that feeds those models, raising questions about what is stored, where, and in what form.

On the verification side, paid detectors are reaching accuracy levels that make automated triage practical, even as their blind spots shift to other media. For practitioners, the immediate workload is to update style and compliance guides to account for watermark behaviour on edited copy, to set retention and access rules for any local activity logs, and to track where verification tools still fall short.

Follow-up to watch

Watch for Anthropic's next clarification on whether the watermark persists after a draft is edited in a word processor, and on how the company plans to treat code outputs specifically. Watch for OpenAI's documentation on the local storage format of Computer History memory files, and for any change in regional availability outside the EEA, Switzerland and the UK. Watch for the next public benchmark of Pangram or a rival detector on multilingual text, since most published numbers to date have been gathered on English.