🧔♂️ A friendly human may check it before it goes live. More news here
World may struggle to prepare for AI safety risks: researcher
AI safety expert David Dalrymple has warned that rapid advances in AI could outpace global efforts to manage safety risks.
Dalrymple, a programme director at the UK’s Aria research agency, said there is concern that AI systems may soon outperform humans in most economically valuable tasks.
He noted a lack of understanding between governments and AI companies about the potential impact of new breakthroughs.
Dalrymple said economic pressures may prevent the development of reliable AI safety measures in time, urging focus on controlling and mitigating risks.
He described unchecked progress as potentially destabilising for security and the economy, and called for more technical work on advanced AI behaviour.
The UK’s AI Security Institute recently reported that leading AI models now complete apprentice-level tasks half the time, up from 10% last year.
🔗 Source: The Guardian
🧠 Food for thought
Implications, context, and why it matters.
Real-world self-replication remains distant despite lab success
- In tests by the UK AI Security Institute (AISI) the success rate on self-replication tasks rose from 5% to 60% between 2023 and 2025. AISI saw no spontaneous self-replication and calls real-world success unlikely 1.
- Controlled runs with scaffolding (external structures that enable tool use and step-by-step guidance) do not mirror real-world conditions. AISI warns that strong scores on these tasks do not translate to real-world success today 1.
- RepliBench (AISI’s standardized test suite for replication-related skills) tracks these behaviors. Success in those runs does not guarantee real-world self-replication 2.
- Models now complete apprentice-level cyber tasks (entry-level cybersecurity work) 50% of the time, up from under 9% in late 2023. Complex multi-step cyberattacks remain hard 1.
Key EU AI Act dates and deadlines
- General-purpose AI (GPAI) rules start on August 2, 2025. Providers with GPAI models already on the EU market have until August 2, 2027 to comply 3.
- July 10, 2025 the European Commission released the GPAI Code of Practice and FAQs to help providers prepare for August 2025 4.
- High-risk AI system requirements take effect on August 2, 2026. They cover Annex III of the EU AI Act. Scope includes biometrics and critical infrastructure. It also includes employment and law enforcement. It reaches democratic processes 5.
- December 17, 2025 the European Commission issued a draft Code of Practice on marking and labeling AI-generated content. It sets technical standards for watermarking and detection before Article 50 of the EU AI Act (transparency obligations) becomes binding on August 2, 2026 5.
Stay updated on the go with our mobile app.
Get latest insights with smoother, more personalized experience through TIA mobile app.