Dario Amodei is defending Anthropic's safety record with a central argument: leading the AI race gives the company more power to make the technology safer.
The debate exposes a difficult question for businesses choosing AI partners: How much confidence do they place in a lab's judgment about risks it also creates?
Anthropic CEO Dario Amodei defended the company's safety strategy after policy advocate Nathan Calvin challenged whether its actions matched its promises. Amodei argued that building leading AI responsibly allows Anthropic to push the industry toward higher standards. He acknowledged mistakes but rejected the claim that safety had become secondary.
The exchange followed Politico's reporting on a draft federal safety bill circulated by Anthropic. The reported proposal would allow the government to block models posing specified foreseeable, imminent risks involving biological weapons, escape from developer controls, or cyberattacks. Encode AI's Calvin argued that the proposal and Anthropic's broader decisions prioritized winning the AI race over safety.
Amodei countered with decisions he said carried commercial costs: safeguards against biological misuse, restricted access to Mythos models from February until June while strengthening cybersecurity protections, and support for safety laws and chip export restrictions. He also acknowledged that AI could accelerate its own development faster than safety work can keep up.
SmarterX founder and CEO Paul Roetzer examined the conviction behind that defense, on Episode 248 of The Artificial Intelligence Show.
More than $1 billion - Damages limit in the draft bill's foreseeable, imminent risk test, per Politico
More than 50 - Threshold for # people who die or suffer serious-injury (in the proposed risk test)
February-June - Period of restricted broad Mythos access cited by Amodei
2023 - Publication year of Anthropic's earlier safety essay
The disagreement starts with a belief. Roetzer says the exchange shows that “these leaders have very high conviction that they're right about a fundamental thing.” In Anthropic's case, he sees a belief that it must build advanced AI before others do because it trusts its own safeguards and values.
The safety concern can be sincere. “I think Dario does take them seriously,” Roetzer says of the safety concerns and national security risks AI brings. Anthropic's 2023 safety essay contemplated halting AI progress under pessimistic safety scenarios. Amodei's response argues that restrictions and continued development can serve the same mission.
Building first still creates risk. As Roetzer explains, “Their justification is: The only way to protect society from advanced AI is to build the advanced AI and understand what it's capable of.” He understands the concern that less cautious competitors would continue anyway, but questions whether that justifies creating increasingly autonomous systems, including AI agents that take steps on their own.
“You can hold models back, you can put guardrails in place, you can do all the things Dario says. But at the end of the day, you are still creating these models and agentic capabilities that are going to allow a whole bunch of pretty messed up stuff to happen.”
— Paul Roetzer, founder and CEO of SmarterX, Episode 248 of The Artificial Intelligence Show
For business leaders, an AI vendor's safety philosophy belongs in the buying conversation. Ask how the provider decides a model is ready, what evidence would delay a release, and how customers learn when its risk assessment changes. Those answers make it easier to fully evaluate the vendor and the purchase decision.
Apply the same discipline internally by deciding which uses require human approval and who can pause a deployment when concerns emerge. A provider may make serious safety investments while accepting risks your organization would not. Choosing a capable model still leaves your team responsible for how it is used.
Amodei's concern about safety falling behind is the next test. Future release decisions will show how Anthropic handles that tension. The draft bill's evolution also matters: the debate focuses on proposed limits and oversight, and its terms remain subject to change.
Only 48% of respondents say their organization has an AI ethics policy or responsible AI principles, according to the 2026 State of AI for Business Report. Evaluating a vendor's values is easier when your organization has articulated its own. Those principles can help teams decide which risks are acceptable, when human judgment is required, and when to stop a deployment.
The report draws on professionals across roles, functions, and industries to examine AI adoption, governance, training, and readiness. It offers context for assessing whether your organization has the structures to turn responsible AI commitments into everyday decisions. Read the full report →