Anthropic drops its core AI safety promise in concerning move

Some might argue that all artificial intelligence development is dangerous to some degree, yet Anthropic have previously stood apart from its competitors in abiding by a strict set of safety rules.

Known as the company's Responsible Scaling Policy (RSP), Anthropic's safety guidelines ensured that it would never train its AI if safety couldn't be guaranteed in advance, which is a stark contrast to the speed-focused approach of many other leading companies in the industry.

That certainly hasn't hindered the company's progress either, as not only is Anthropic raking in billions of dollars in annual revenue – something that other companies appear to be struggling with – but many experts and professionals tend to favor Claude over competing models.

Only recently was the company locked in a stand off with the United States government over the conditions of a military contract, as while the Department of War wanted to relax the requirements of the partnership, Anthropic remained firm in its desire for safety.

Anthropic has now dropped some of the key safety guidelines that set it apart from other companies (Gabby Jones/Bloomberg via Getty Images)

That has unfortunately now changed, however, as reports from TIME indicate that the RSP has been adapted with the central pledge of this safety policy dropped altogether.

What's clear is that the RSP still very much exists, but it's far more lenient on safety when it comes to keeping up with competitors, with the strict rules only really coming in to play if Anthropic is already leading the race.

"We felt that it wouldn't actually help anyone for us to stop training AI models," explained Jared Kaplan, Anthropic's chief science officer. "We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments [...] if competitors are blazing ahead."

Anthropic's RSP will only apply if the company is leading the race, and spots something that could be 'catastrophic' (Smith Collection/Gado / Contributor/Getty Images)

Anthropic has still pledged to at least match the safety commitments of its competitors – which isn't quite as promising as it may sound considering the ethical headaches that have already emerged from some of the leading models – but 'delays' to training will only be implemented if the risks of catastrophe are 'significant' and the company is already ahead of its rivals.

CEO and co-founder Dario Amodei has previously warned that forgoing the necessary safety precautions for training AI could be catastrophic for humanity, yet it appears as if that's thrown out the window when you start falling behind.

Of course, Anthropic's safety precautions will be little help if another competing model has any catastrophic issues, but Anthropic once stood as a company that others could strive towards or emulate.

Kaplan doesn't believe that the changes made to the company's RSP is a 'U-turn', however, noting that it's instead a pragmatic response to the current and emerging scientific and political situation that the world faces.