AI Safety

AI Researchers and Executives Continue to Underestimate the Near-Future Risks of Open Models

New analysis argues open-source AI models could bypass all proposed safety guardrails within 1-5 years.

Deep Dive

AI researcher Andrew Dickson critiques leading AI executives like Anthropic's Dario Amodei for systematically omitting open-source models from near-term (1-5 year) risk forecasts. He argues that guardrails like Constitutional AI are ineffective on open models, as they can be removed, creating a major defense gap. The piece calls for urgent collective action to develop viable safety alternatives before risks become extreme.

Why It Matters

If open models undermine proposed safety frameworks, current AI governance strategies may be fundamentally flawed.