EU AI Act

2026-10-31

MODELS · 1 SOURCE · The Decoder

Anthropic rolls out watermark to four Claude models, but two older models can be bypassed by changing sampling settings

Anthropic is adding its imperceptible text watermark to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6; the watermark is embedded at the model level and will appear across hosting platforms (including Amazon Bedrock). On Opus 4.6 and Sonnet 4.6 the watermark only activates when developers leave sampling parameters (temperature, top_p, top_k) at their defaults, so tweaking any one parameter can evade the mark; Anthropic offers a private-preview detection API and cites the EU AI Act deadline as the reason for the rollout.

8.0

2026-10-27

REGULATION · 1 SOURCE · arXiv cs.AI

Conformance as a Distribution: delegation contracts and evaluation mechanisms for use-based AI regulation (arXiv:2610.33112v1)

This arXiv preprint formalizes "conformance as a distribution" for use-based AI regulation and analyzes delegation contracts among a deployer, auditor, and enforcement point. It proposes mechanisms and theoretical propositions addressing truthful declared context mixtures, unbiased commit-reveal audits, certification error budgets (including halving budgets and α-investing), per-subgroup audit scaling, and limitations of failure-only incident metrics, and tests each claim in five preregistered synthetic studies; no language models or production systems are evaluated. The paper frames conformance-as-a-rate as an alternative to static conformity assessment such as in the EU AI Act.

6.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

ArXiv paper: 'The Harness as the Only Mutable Surface' — admission-gated self-evolution for LLM agents in credit pipelines

The paper argues that LLM agents may only be reviewable if self-modification is confined to a mutable runtime harness (instructions, tool-call logic, primitive composition) while model weights remain fixed, and it presents a dual-loop engine with a single admission gate that writes a hash-chained record before deployment. In simulation (simulated agent and seeded proposer) across three families of supervisory re-interpretation and 10 seeds each, the gated loop admitted 144 of 7,449 candidate changes and did not worsen held-out error, while an unbounded check admitted 309 harmful changes and left missed flags >10% in 49 of 90 runs; the gate rejected every candidate when evaluated on pre-shift labels and had mixed outcomes on high-severity structural shifts.

7.0

2026-10-05

REGULATION · 3 SOURCES · The Decoder · The Verge AI · TechCrunch AI

OpenAI will add invisible text watermarking (textGrain) for ChatGPT in the EU; API opt-in worldwide

OpenAI is rolling out textGrain, an invisible statistical watermarking system, for ChatGPT and Codex users in the European Union to comply with the EU AI Act, and will let API customers worldwide opt in to watermarking (including via cloud partners such as Microsoft Azure). OpenAI says textGrain performed competitively in internal tests, will be open-sourced, but detection reliability varies by text length and topic and degrades substantially with edits, and initial access to the detector will be limited to selected researchers and organizations.

8.0

2026-10-05

MODELS · 1 SOURCE · The Decoder

Aleph Alpha releases Kolibri, a 78B German–English MoE model with 1M-token context and open weights

Aleph Alpha has released Kolibri, a 78-billion-parameter German–English mixture-of-experts model that activates about 3 billion parameters per token, supports context windows up to one million tokens, and was trained on 768 B200 GPUs in Germany and Finland. The company says German makes up 21.3% of the training data, that Chinese models were used to generate synthetic data, and that Kolibri sits on the Pareto front of quality and operating cost; the weights are published under an Apache 2.0 license on Hugging Face and the model is positioned for public administration, aviation, and industry with EU AI Act compliance in mind.

8.0

2026-09-30

RESEARCH · 1 SOURCE · arXiv cs.AI

Empirical study finds current EU AI Act compliance checkers inconsistent and insufficient

ArXiv paper arXiv:2609.36228v1 presents the first empirical assessment of 12 mainstream automated EU AI Act compliance checkers (AIACCs). The study evaluates legal coverage, report structure, determinacy, and actionability, finding wide quality variation and concluding that current tools are only suitable as early-stage orientation aids and may give a false sense of compliance.

7.0

2026-09-16

REGULATION · 1 SOURCE · Cohere

Cohere signs EU Code of Practice on Transparency of AI‑Generated Content (Section 1)

Cohere announced it has signed the EU Code of Practice on Transparency of AI-Generated Content and is among the first signatories, confirming it signed Section 1 that applies to AI system providers. The company said the step supports compliance with Article 50 of the EU AI Act and that it will continue collaborating with EU bodies such as the EU AI Office on implementation.

6.0

2026-09-16

REGULATION · 1 SOURCE · Anthropic

Anthropic to watermark Claude outputs using SynthID-Text method to meet EU AI Act

Anthropic announced that future Claude models will embed an imperceptible probabilistic text watermark (using the SynthID-Text approach) to enable likelihood-based detection that Claude produced the text, as part of compliance with the EU AI Act. The company says the watermark adds no visible characters, incurs no extra tokens or cost, is not user-identifying or traceable, and has shown no measurable impact on output quality in internal and referenced tests.

7.0

2026-09-15

RESEARCH · 1 SOURCE · arXiv cs.AI

From Legal Text to AI-specific Risk Sources: Systematic analysis of the EU AI Act's high-risk requirements (arXiv preprint)

Authors present a systematic classification of requirements in Section 2 of the EU AI Act (requirements for high-risk AI systems) and find that only a minority of those requirements directly address AI-specific risk sources while most impose organizational/process and documentation obligations. From the subset of AI-related requirements they derive a consolidated 'EU AI Act Risk Source List' to help compare the Act's implicit risk coverage with established AI risk taxonomies; this work is an authors' preprint presented at the cited conference.

6.0