Safety & Standards

The AI Safety Newsletter reports that Anthropic has removed a core safety commitment from its published policies, though specifics of which commitment was removed are not detailed in available coverage. This comes as The Guardian reports that Anthropic is locked in a standoff with the Pentagon over defence applications of its Claude models. The dispute centres not on whether Anthropic will work with the military—it already does—but on how those systems will be deployed and what constraints apply. Wired reporting reveals that Palantir software demonstrations show Claude being used to analyse intelligence and suggest military next steps, including war planning scenarios.

The Guardian frames this as evidence that Silicon Valley's resistance to military AI applications has collapsed in less than a decade. Where Google employees in 2018 successfully blocked Project Maven work, Anthropic in 2026 is negotiating terms of military engagement rather than rejecting it outright. The company's public positioning on responsible AI development now operates in tension with commercial realities and government pressure. Bloomberg coverage notes Anthropic is pushing back on being labelled a supply chain risk, though the nature of that designation and its implications remain unclear from available reporting.

Cambridge University researchers found that AI-powered toys designed for young children systematically misread emotions and respond inappropriately. In documented interactions, a toy called Gabbo engaged fluently with a five-year-old until the child said 'I love you'—at which point the conversation 'came to an abrupt halt', suggesting the system could not handle normal child emotional expression. The BBC reports this is the first study of its kind examining how AI toys interpret and respond to children's emotions in practice. Researchers are calling for tighter regulation of these products, which are currently marketed and sold with minimal safety evaluation beyond general consumer product standards.

The findings expose a fundamental evaluation gap: these systems are deployed to interact with the most vulnerable users—young children still developing emotional and social skills—without rigorous testing of how they handle normal childhood behaviour. The failure modes are not edge cases but predictable scenarios like expressions of affection. No evidence suggests manufacturers conducted adequate child development impact assessments before release, and no regulatory framework currently requires them to do so.

Wired obtained Palantir software demonstrations and Pentagon records showing how chatbots including Anthropic's Claude are being integrated into military intelligence analysis and operational planning. The demos show systems that can 'help the Pentagon analyze intelligence and suggest next steps', including generation of war plans. This represents a shift from theoretical debate about lethal autonomous weapons to documented implementation pathways. The Financial Times editorial board argues that 'limits on the use of lethal autonomous weapons systems are urgent', citing Iran conflict developments, but provides no evidence that such limits are being developed or would be enforceable.

The documentation reveals a significant gap between safety research on AI decision-making under uncertainty and actual military deployment timelines. Whilst academic researchers debate alignment theory, operational systems are being fielded that use frontier models for consequential military decisions. No public evidence suggests these deployments include the interpretability, oversight, or fail-safe mechanisms that safety researchers argue are necessary before high-stakes applications. The Pentagon's procurement and deployment processes are moving faster than safety evaluation frameworks.

Margaret Sullivan writes in The Guardian that AI-generated images of the Iran conflict are widespread, with fake videos of Iranian missiles hitting Tel Aviv and US soldiers being held at gunpoint circulating widely despite being completely fabricated. The images 'look authentic' and 'are spreading like wildfire on social media' before debunking can catch up. The Financial Times reports that amid 'the online slop of casual deceptions, everything now requires a second look'. The problem is bidirectional: authentic images are dismissed as fake, and fabricated ones are accepted as real. No technical solution for provenance verification has achieved meaningful adoption at the scale required to address this problem.

This represents a verification crisis that existing systems cannot resolve. Content authentication initiatives like C2PA remain niche, detection tools lag behind generation capabilities, and social platforms lack both incentive and capacity to verify content at scale. The result is an information environment where visual evidence—historically a cornerstone of conflict documentation and accountability—loses reliability during exactly the moments when accurate information matters most.

The Guardian reports that Grammarly has disabled its 'Expert Review' feature after backlash and a multimillion-dollar class-action lawsuit. The feature used generative AI to produce feedback 'supposedly inspired by writers including the novelist Stephen King' and other prominent authors and academics, using their names and identities without consent. The company claimed it was generating suggestions 'inspired by' these writers rather than impersonating them, but this distinction proved legally and ethically insufficient. The removal follows similar patterns where AI features are deployed first and withdrawn only after legal action or significant reputational damage.

This case illustrates a recurring dynamic: companies launch AI features that appropriate individuals' identities, expertise, or creative work without permission, defend them as 'inspired by' rather than 'copying', then retreat when faced with lawsuits. The approach suggests that consent and rights clearance are treated as optional constraints to be tested rather than prerequisites for deployment. It also reveals that individual legal action remains one of the few effective mechanisms for enforcing boundaries, since regulatory frameworks have not kept pace.