All articles tagged with #watermarking
New research shows SynthID-Text watermarking can subtly alter next-word choices and tool usage in LLMs, causing, under adversarial or prompt-injection conditions, models to comply with harmful requests more often or change how they refuse them. The study, which tested six open-weight models (Claude not included), highlights a phenomenon called sampling drift and underscores the need for thorough safety testing of watermarking in AI deployments.
Anthropic plans to watermark Claude's outputs by subtly adjusting word choice using a secret key, so the text is detectable as AI-generated without any visible changes. The company claims this won't affect quality, but the article argues the move shows a broader disregard for the craft of writing, treating words as fungible. It cites Google’s SynthID research and critics, and frames watermarking as a regulatory response that raises questions about AI’s role in mimicking human writing.
Anthropic says Claude will apply invisible watermarks using a SynthID-Text method to comply with the EU AI Act’s requirement that synthetic text be machine-readable and detectable as generated content. The watermarking uses a key and surrounding words to influence word choices, remains invisible to readers, and does not affect Claude’s output quality or cost. Google’s Gemini already uses SynthID-Text, and while OpenAI’s ChatGPT plans aren’t detailed, it will also fall under the same regulations.
Anthropic detailed Claude’s text watermarking system, based on the SynthID-Text approach, which uses low-stakes word choices to imprint an undetectable pattern that signals AI involvement without affecting output, extra tokens, or reader experience; a forthcoming detection API will let users check text, but the watermark is not a foolproof test for human vs. AI authorship and works best on longer passages, with reduced effectiveness on factual or short samples.
Anthropic says Claude will generate text with an invisible, key-based watermark to satisfy EU AI transparency rules. The pattern, which can be decoded with a key (illustrated using digits of pi), is designed to not affect output quality or cost and is inspired by Google DeepMind’s SynthID-Text. An API will provide decoding keys to verify AI involvement. Limitations include not being able to tell if Claude wrote or simply edited text, potential detection issues for translations or very short passages, and less watermarking in code. Watermarks will be added to all Claude outputs from post-August 2 models, with older models updated later; images will carry a cryptographic metadata note as well."
Anthropic detailed its Claude watermarking system, rooted in a secret-key token approach inspired by Google’s SynthID, claiming negligible speed impact and that watermarking could appear even when Claude edits text; the watermark is intended to help comply with the EU AI Act. Despite assurances, some Claude users on X report canceling subscriptions in protest, though Anthropic says overall cancelations have not spiked.
Google’s SynthID watermark can resist heavy edits and compression, but it’s not a cure for AI misinformation: watermarks can be bypassed by cropping or adversarial tweaks, detectors are fragmented and access is limited, and widespread unlabeled AI content persists. Experts argue provenance data like C2PA, though not yet universal, may be a more reliable path to truth as AI content proliferates.
The 2026 ICML used hidden watermarks in submitted papers to detect AI-generated peer reviews; 497 papers (about 2%) were rejected for violations of the conference's LLM-use policy, and organizers introduced two review streams—one permitting limited LLM use and one banning it—to protect trust in the peer-review process.
Cornell researchers have developed a novel invisible watermarking technique called noise-coded illumination that embeds unique light patterns into scenes during recording, enabling verification of video authenticity and detection of deepfakes by analyzing inconsistencies in the lighting codes, which are imperceptible to the naked eye and effective across various environments and skin tones.
Meta has introduced Meta Video Seal, a tool for watermarking AI-generated videos to combat the rise of deepfakes, which have significantly increased in prevalence. The tool, available in open source, aims to provide robust watermarking that withstands video compression and common edits. Despite existing solutions like DeepMind's SynthID and Microsoft's methods, Meta claims its tool offers superior resilience and scalability. However, adoption challenges remain, as many developers use proprietary solutions. Meta is also launching a leaderboard and organizing a workshop to encourage industry collaboration.
Meta has launched Imagine with Meta, a standalone generative AI experience that allows users to create high-resolution images by describing them in natural language. Similar to OpenAI's DALL-E, Imagine with Meta is powered by Meta's Emu image generation model and generates four images per prompt. Meta plans to add invisible watermarks to the generated content for increased transparency and traceability. The company aims to bring invisible watermarking to more of its products with AI-generated images in the future. The move comes as pressure mounts on tech firms to clearly indicate when works are generated by AI.
Amazon has introduced its Titan text-to-image AI model, Titan Image Generator, which allows developers to create realistic images with built-in guardrails against toxicity and bias. The model, aimed at an enterprise audience, offers the ability to generate images from natural language prompts and change backgrounds. All images generated by Titan Image Generator will include invisible watermarks as part of Amazon's voluntary commitments. The watermark can be detected using an API, and developers have the flexibility to choose how to provide that information to users. Additionally, Amazon announced the availability of other Titan models and extended copyright indemnity to customers using its foundation models.
Researchers at the University of Maryland have found that current AI watermarks are easily removable, raising concerns about the effectiveness of watermarking technology in combating deepfakes and misinformation. The study revealed that existing methods of watermarking can be evaded and fake emblems can be added to non-AI generated images. However, the researchers also developed a near-impossible-to-remove watermark that could help detect stolen products. Another collaborative research effort between the University of California, Santa Barbara and Carnegie Mellon University found that watermarks can be easily eliminated through destructive and constructive attacks. As AI-generated content becomes more prevalent, the need for improved watermarking technology to combat misinformation becomes crucial.
Researchers at the University of Maryland have found that current AI watermarking methods are easily defeated, making them unreliable for identifying AI-generated images and text. The study demonstrates how attackers can remove watermarks and even add watermarks to human-generated images, triggering false positives. Watermarking has been considered a promising strategy to combat misinformation and deepfakes, but researchers have consistently pointed out its major shortcomings. While some believe that watermarking can still be part of the solution when combined with other technologies, others argue that it is not effective and can be easily faked or removed.
Researchers at the University of Maryland have found that current AI watermarking methods are easily defeated, making them unreliable for identifying AI-generated images and text. The study demonstrates how attackers can remove watermarks and even add watermarks to human-generated images, triggering false positives. Watermarking has been considered a promising strategy to combat misinformation and deepfakes, but researchers have consistently pointed out its major shortcomings. While some believe that watermarking can still be part of the solution when combined with other technologies, others argue that it is not effective and can be easily faked or removed.