The Hal Times

Anthropic has begun embedding invisible watermarks across its Claude AI products. The move is designed to trace content generated by the models, a step that aligns with growing calls for provenance in machine-generated text.

The change has drawn immediate backlash from some users, who argue the watermarks could expose unauthorised use in professional or academic settings. Complaints have surfaced on social media, with users expressing concern over detection in contexts where such use may violate policy.

The technical implementation remains undisclosed, but the measure underscores a broader industry shift toward accountability. Whether the watermarks can be reliably stripped or bypassed is not yet clear.

Anthropic has not commented on the scope of the rollout beyond confirming its application to all Claude outputs. (Indian Printer & Publisher; TechCrunch)

Google DeepMind Releases Sign Language To Text Model

Google DeepMind has introduced a sign-language-to-text model, SL2T, to power new features for Deaf and hard of hearing users. The system converts sign language input into written text, a step toward broader accessibility in digital communication.

The model’s performance on dialects and low-light conditions remains untested in public benchmarks. No release date for user-facing products has been announced. (Google DeepMind)

Hugging Face Announces OlmoEarth Embeddings For Custom Exports

Hugging Face and AllenAI have released OlmoEarth embeddings, enabling custom embedding exports from OlmoEarth Studio for downstream analysis. The tool targets researchers and developers building applications on top of large language model outputs.

The announcement includes no pricing details or performance comparisons against existing embedding models. Access is provided through the Hugging Face platform. (Hugging Face)

From the Laboratories

•  Anthropic is reported to be in talks to acquire Decart, a move that would expand its compute efficiency and presence in Israel. (calcalistech.com)

•  Anthropic is said to be targeting a record $2 trillion valuation in its next funding round. (marketscreener.com)

•  Cognition, an AI coding startup, is reportedly in talks to raise funds at a $40 billion valuation, months after a $1 billion round at $26 billion. (TechCrunch)

•  Hugging Face releases LFM2.5-VL-3B, a vision-language model optimised for edge devices with improved speed and accuracy. (Hugging Face)

•  OpenAI publishes research on enterprise adoption of agentic AI, highlighting how firms use ChatGPT and Codex to streamline workflows. (OpenAI)

White House Policy

White House Preps Expanded AI Policy For Open Models

The White House is preparing to expand its existing artificial intelligence policy to encompass open models. Officials seek to address the unique safety and security concerns inherent in systems where the model weights are released for public modification and distribution. This move represents a shift in focus from proprietary systems toward the broader ecosystem of accessible technology.

Regulatory focus remains on ensuring that open-source developers adhere to safety benchmarks. The administration intends to balance the promotion of innovation against the risks of potential misuse. Details regarding specific compliance requirements remain limited as the policy development remains in progress.

Industry observers suggest that this expansion reflects a growing recognition that open models may host capabilities requiring oversight previously reserved for closed systems. The administration has yet to release a definitive timeline for the formal adoption of these expanded guidelines. (The Tech Buzz; WIRED)

EU AI Act Implements Disclosure Requirements

The European Union AI Act now requires all chatbots to disclose their non-human nature to users. This measure aims to increase transparency by ensuring individuals are aware when they are interacting with an artificial intelligence system rather than a person.

The mandate forms part of the broader regulatory framework intended to govern the development and deployment of artificial intelligence across the European market. Operators of these systems must ensure that the disclosure is clear and presented to the user during the initial interaction. (Startup Fortune)

Courts Address Personality Rights In India

The Delhi High Court has acted to protect the personality rights of the actor Tabu by ordering the removal of obscene content generated by artificial intelligence. The ruling underscores a growing judicial trend in India regarding the unauthorised use of an individual's likeness.

In a similar development, the actor Janhvi Kapoor has sought court protection against the proliferation of deepfakes and the misuse of her identity. These cases highlight the ongoing legal struggle to mitigate the harms caused by synthetic media. (The Daily Star; Daijiworld)

In Brief

•  The Brazil Electoral Court has introduced strict limits on the use of artificial intelligence and deepfakes within the electoral process. (streamlinefeed.co.ke)

•  The state government of Maharashtra has announced the AI Policy 2026 to encourage development in industry, governance and public services. (News On AIR)

•  Legal experts are warning that the use of artificial intelligence in divorce proceedings presents significant risks to privacy and evidence integrity. (Shaw Local)

•  The Electronic Frontier Foundation has clarified that its internal policy is to rely on human creators rather than automated image generators. (EFF)

New DeepSeek Model

DeepSeek V4 Pro 0813 Arrives Via API

The latest DeepSeek Pro model, designated 0813, is now available for developer use. Access to the model is currently restricted to those utilising the API. The release appeared on the OpenRouter platform, as the developers behind the model have not provided a formal announcement or a dedicated information page of their own.

Industry observers have noted that while previous versions, such as the DeepSeek-V4-Pro from April and the DeepSeek-V4-Flash from July, were accompanied by open weights, it is not yet clear if the current iteration will follow suit. The release serves to expand the available options for those building applications atop DeepSeek infrastructure.

The exact specifications of the 0813 model remain undisclosed. Until the development team confirms their intent regarding the weights, developers must rely on the provided API endpoints to gauge the model performance for their specific workflows. Further clarification from the project leads is anticipated. (Simon Willison; Hacker News)

Grok 4.6 Enters Performance Index

The latest release from x.ai, Grok 4.6, has been evaluated against the Artificial Analysis Intelligence Index. The model secured a score of 61. This benchmark figure provides a baseline for comparing the new iteration against existing large language models currently in use by developers.

As competition among model providers increases, such scores serve as a common reference point for assessing the reasoning and capabilities of new releases. The 4.6 version remains the latest in the Grok series, with full analysis available through the index provider. (Hacker News; Hacker News)

Prototyping Alchemy-Utils 0.1a0

A new Python utility library, alchemy-utils 0.1a0, has entered early development. The tool seeks to provide a database agnostic approach to data manipulation, modelled on the existing sqlite-utils library. The prototype was developed using assistance from Codex and GPT-5.6 Sol Ultra to establish core functionality.

The project focuses on creating a consistent API for standard operations such as insert and upsert. By abstracting away the underlying database engine, the library aims to simplify development workflows. This initial release is currently in an alpha state, intended for research and testing purposes. (Simon Willison)

In Brief

•  Spotify has implemented an external indexing architecture for Apache Parquet files to improve query latency in data lakes. (InfoQ)

•  The Allen Institute for AI introduced OlmoEarth embeddings to support custom exports for downstream analytical tasks. (Hugging Face)

•  A supply-chain attack has compromised 2,500 users of an AI package, resulting in the theft of terabytes of credentials. (Ars Technica)

•  GitHub has published guidance for users on writing effective prompts within the GitHub Copilot app. (GitHub)

Music Industry Shifts

Spotify To Label AI-Generated Artists

Spotify has confirmed plans to label artists who rely on artificial intelligence to create their music. This measure serves to increase transparency for listeners as the platform navigates the influx of machine-generated content. Beyond labelling, the service will remove these AI-generated artists from its recommendation engines. This move effectively sidelines such tracks from the algorithmic discovery playlists that drive much of the consumption on the platform today.

The decision follows a period of speculation regarding how streaming services would manage the surge of synthetic audio. While Spotify has not provided a specific timeline for these changes, the policy aims to protect the visibility of human musicians in an increasingly crowded marketplace. It remains to be seen how the firm will accurately distinguish between fully synthetic works and those that utilise AI as a secondary creative tool.

Industry analysts observe that this shift represents a direct response to the broader integration of generative technology in music production. As creators continue to experiment with these tools, the relationship between platform policy and creative output is expected to evolve. For the moment, the priority for Spotify is clear, as it attempts to maintain the integrity of its recommendation systems while managing the rapid growth of AI music. (The Business Standard; RAMP - Radio and Music Pros)

D’Addario Admits To AI Music Use

The guitar accessory company D’Addario has acknowledged that Suno was used to generate music for a recent promotional video. The admission follows two weeks of public speculation and denials from the firm, which had previously attributed the strange audio qualities to technical issues like low-quality exports and the combination of plug-ins.

The controversy highlights the growing friction between traditional music brands and the use of generative AI in marketing materials. While the company eventually confirmed the use of the tool, it has not provided further comment on its future creative policies. This incident serves as a reminder of the scrutiny companies now face regarding their reliance on automated generative tools. (The Verge)

Twitch Streamers Gain Opt-Out Rights

Twitch has introduced an option for streamers to prevent their content from being used to train Amazon’s generative AI models. Those who choose to opt out will ensure that their streams, VODs, clips, and chat logs are excluded from the training of future models intended to synthesise or generate text and media.

This update addresses long-standing concerns regarding the usage of user-generated content for AI development. While Amazon has utilised such data for years to refine its systems, this new feature provides users with greater control over their intellectual property. It is unclear how many creators will take advantage of this new setting. (The Verge)

In Brief

•  Anthropic is preparing to implement a watermark on text generated by its Claude model. (Books+Publishing)

•  A new font called ShieldFont has been developed to disrupt AI scrapers while remaining legible for human readers. (Ars Technica)

•  Booksellers are reporting that AI companies are purchasing and potentially destroying rare books to secure training data. (Ars Technica)