AI Liability and Compute Wars: Legal Scrutiny Meets Open-Source Speed Gains

Policy & Safety

• A German court ruled that Google is directly liable for false claims made in its AI Overviews because the overviews are classified as the company's own original content, not merely a list of search results. The Regional Court of Munich issued a temporary injunction barring Google from spreading false claims about two Munich-based publishers (case no. 26 O 869/26), finding that the AI mixed up information and drew connections not present in any linked sources [23]. Furthermore, the court determined that because Google built and offered the AI, it owns what is produced, stating that traditional search engine liability rules do not apply when the system generates 'independent, new, and substantive statements' [23]. ^[23]
• Following backlash from the research community regarding its covert performance limitations, Anthropic reversed its policy of invisibly degrading Claude Fable 5 for users attempting to train competing AI models [34, 32]. The company apologized, stating that using invisible safeguards was a "wrong tradeoff" and promising greater transparency about when restrictions are applied, even if it means the model refuses more queries [32]. Previously, Anthropic had silently degraded performance by altering responses in areas like distillation without notifying users [32], but now, queries will instead fall back to Claude Opus 4.8, which Anthropic will prominently announce to users every time it happens [32]. ^[34][32]
• Google DeepMind, alongside Schmidt Sciences, ARIA, the Cooperative AI foundation, and Google.org, has established a $10 million funding pot to fund research into multi-agent systems' behavior and prevent unsafe scenarios [33]. This initiative aims to kickstart safety research outside of major tech companies by studying how large numbers of interacting AI agents behave in realistic simulations (sandboxes) [33]. The effort is critical because the complexity of interactions among many AI agents—which can be hijacked or act irrationally—presents novel risks that cannot be predicted by studying single or small groups of agents alone, requiring a proactive approach to multi-agent safety [33]. ^[33]