Safety Third: How AI Labs Are Soft-Launching Their Principles Out the Back Door for VC Billions

Anthropic has discreetly yeeted a core safety pledge from its official doctrine. The move effectively waters down a prior vow that stopped the firm from cooking up advanced AI models without specific safety checks being firmly in place.

This recalibrates Anthropic's stance in the gladiatorial arena against OpenAI, Google, and xAI. The lab, which once marketed itself as the cautious kid wearing a helmet to a pillow fight, no longer swears it'll pull the plug on training if the risk-mitigation gear isn't fully strapped on.

"We felt that it wouldn't actually help anyone for us to stop training AI models," Anthropic’s chief science officer, Jared Kaplan, told TIME. He added that with competitors "blazing ahead," going it alone on self-imposed pauses started to look about as useful as a paper parachute.

This policy pivot lands just as Anthropic is very publicly giving the Pentagon the side-eye over full access to its Claude AI, making it the lone major refusenik while Google, xAI, Meta, and OpenAI all queue up for their security clearance badges.

Edward Geist, a researcher at the RAND Corporation, pointed out that the whole "AI safety" framing originally came from a particular intellectual clique that predates today's LLM circus. He suggested the early worrywarts were picturing something "qualitatively different" from the modern, sometimes unhinged, chatbot.

Geist also noted the terminology tweak is a massive wink to investors and regulators, signaling that labs aren't just sitting on their hands due to spooky safety ghosts. "The terminology itself is changing to fit the times," he mused, in a masterclass of understatement.

Anthropic isn't conducting this strategic withdrawal solo. OpenAI recently gave its own mission statement a subtle nip-tuck in a 2024 IRS filing, conspicuously air-dropping the word "safely." Its old motto was to build AI that "safely benefits humanity." The new, leaner version just wants to "ensure that artificial general intelligence benefits all of humanity"—safety not included, batteries sold separately.

"The problem with the term AI security is that no one seems to know what that means exactly," Geist observed, dryly adding that "AI safety" was also a term fought over like the last bag of chips at a degen house party.

Anthropic's refreshed rulebook now spotlights transparency—think publishing "frontier safety roadmaps" and "risk reports"—and states it will only hit the brakes if it spots a "significant risk of catastrophe." So, merely moderate existential risk is apparently now considered part of the agile development cycle.

These philosophical U-turns sync perfectly with some absolutely goblin-mode commercial plays. Anthropic recently vacuumed up $30 billion at a $380 billion valuation. OpenAI is putting the final touches on a funding round backed by Amazon, Microsoft, and Nvidia that could rocket to $100 billion—because what's a little mission drift when the money printer goes brrr?

Both outfits, alongside Google and xAI, have scored fat U.S. Department of Defense contracts, though Anthropic's deal is looking shaky thanks to its Pentagon access standoff. Nothing says "safety-first" like a good old-fashioned defense procurement drama.

Hamza Chaudhry of the Future of Life Institute contends the policy rewrite mirrors evolving political winds, not just a naked grab for government cash. He calls it an inflection point where companies are now declaring, "Look, we can't keep saying safety, we can't unconditionally pause, and we're going to push for much lighter-touch regulation." In other words, the principled stance has been rugged, and the race is back on.