Anthropic's Claude uses Google's SynthID watermark; here's what you can verify

·1 views

Anthropic's text watermarking for Claude has been confirmed to use Google DeepMind's SynthID-Text method, published in Nature, which works by biasing word choices using a secret key rather than embedding invisible Unicode characters. A separate, verifiable layer involves C2PA content credentials attached to AI-generated images since August 11, an open standard anyone can check without vendor access. The text watermark cannot be detected without Anthropic's private key, and stripping invisible characters has no effect on it since the signal lies in word selection patterns. Contrary to common belief, the watermark indicates model involvement in generation but does not confirm how much of a given text was AI-written. Simple techniques like re-tokenisation are documented to defeat the text mark, a fact largely absent from mainstream coverage of the system.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Debate Revisited: Can Consciousness Be Explained Without New Physics?

A philosophical article published on Overcoming Bias examines whether human consciousness can be fully explained within the framework of existing physics. The piece engages with longstanding questions about whether understanding the mind requires any novel scientific principles beyond what is currently known. The author argues for a position that consciousness does not necessitate new physical laws or phenomena. The article has attracted modest attention on Hacker News, where it was shared for community discussion.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

ZIM Master Prompt Aims to Stop AI from Generating Outdated Canvas Code

The ZIM team, led by Dr. Abstract, has developed the ZIM Master Prompt to address a recurring problem where AI language models generate outdated or incorrect code for the ZIM JavaScript canvas framework. Without guidance, LLMs tend to default to legacy CreateJS patterns and obsolete methods due to their exposure to older training data. The prompt directs AI models to two lightweight, machine-readable reference pages — a stripped-down API map and a style guide — instead of full documentation pages that can overwhelm an AI's context window. The API reference provides exact parameter signatures and flags for special input types, while the style guide enforces ZIM's modern, concise coding conventions. The tool is publicly available at zimjs.com/prompt and is designed to help developers get clean, idiomatic ZIM code from AI assistants.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Obsidian's Dataview Plugin Lets You Query Notes Like a Database

The Dataview plugin for Obsidian transforms note-taking by treating YAML frontmatter fields as database rows, enabling SQL-like queries directly within notes. Users can build live, auto-updating tables to filter and sort notes by tags, status, language, or project without maintaining a manual index. Common use cases include grouping code snippets by programming language and creating lightweight status boards to track open items. The plugin's main limitation is consistency — notes with missing frontmatter fields are silently excluded from query results, making disciplined metadata entry essential. To address this, pre-built Obsidian vault templates can ensure frontmatter fields are in place from the moment a new note is created.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

PHP FFI ioctl Bug on Apple Silicon Silently Corrupts Terminal Window Size

A developer discovered a subtle but critical bug while building pseudo-terminal support for PHP using Foreign Function Interface (FFI) on Apple Silicon Macs. The issue arises because most PHP FFI code snippets declare ioctl with a fixed-arity signature, but ioctl is actually a variadic function in C. Apple's ARM64 ABI differs from the standard in that variadic arguments are passed on the stack rather than in registers, causing the wrong memory to be read when ioctl is called incorrectly. This means the call returns a success code of zero while silently writing garbage data, making the bug extremely difficult to detect — especially if CI pipelines run only on Linux, where the two calling conventions happen to agree. The fix is a one-line change: replacing the typed third parameter in the FFI declaration with an ellipsis (...), prompting libffi to generate the correct call frame for Darwin.

0 comments Read more at DEV Community