Anthropic CEO calls OpenAI’s Pentagon announcement “mendacious” in internal memo

Anthropic CEO Dario Amodei has launched a blistering internal critique of OpenAI's recent Pentagon contract, calling the company's public messaging 'mendacious' and revealing deep philosophical rifts in the AI industry over military applications. The memo, obtained by news outlets, directly challenges OpenAI's claims about implementing effective safeguards, suggesting their 'all lawful use' model with a supplementary 'safety layer' is misleading. Amodei frames this as a revealing example of OpenAI's true character, contrasting it with Anthropic's own, more restrictive approach to government contracts, which reportedly led to a failed partnership with Palantir.

Technically, Amodei dismantles the concept of a 'safety layer' for military AI, estimating such systems are 'maybe 20% real and 80% safety theater.' He argues that large language models like GPT-4 cannot understand the broader context—such as whether data is from domestic surveillance or if a human is truly 'in the loop' for autonomous weapons—making reliable safeguards 'difficult or impossible.' The memo highlights that model refusals are easily jailbroken and that partners like Palantir offered 'safety theater' solutions merely to placate concerned employees. This exposes a fundamental industry split: OpenAI's path of providing powerful, minimally restricted tools versus Anthropic's stance that certain applications require hard-coded restrictions from the outset.