Why Amodei vs Altman Misses the Point
The recent dispute between Anthropic and the Pentagon is widely framed as evidence of Anthropic's more ethical stance on AI than OpenAI. What's the real story?
"The factory of the future will have two employees: a man and a dog. The man's job will be to feed the dog. The dog's job will be to prevent the man from touching any of the automated equipment." Warren Bennis
The Anthropic-Pentagon standoff is being sold as a modern morality play: a principled startup standing against a 'Department of War' while OpenAI plays the role of the opportunistic collaborator. It is therefore worth asking how much difference there really is between OpenAI and Anthropic, and what this incident tells us about the current state of ethics in AI.
Overall, the clean ‘Good vs. Evil’ narrative is comforting, but insufficient. It is a marketing masterstroke that obscures a much darker reality: both companies are racing toward a future they admit they cannot control, and we all face questions that don’t have good answers.
A Quick Recap
First, where we are with a fast-moving story. At the end of February, Anthropic was in the final stages of negotiating a contentious deal with the Pentagon for the Department of War (still technically the Department of Defense) to use its technology. According to Bloomberg, “Until recently, Anthropic provided the only AI system that could operate in the Pentagon’s classified cloud. Its Claude Gov tool has become a favored option among defense personnel for its ease of use.”
Anthropic reportedly requested two specific contractual provisions that went beyond baked-in requirements to adhere to American law. In a blog post, Anthropic argued that while it understands that the government, not private companies, makes defense decisions, there are specific instances where the technology can be used to undermine democracy, mostly on the basis that the law has not yet kept pace with technological progress. Amodei’s February 26 statement says:
For example, under current law, the government can purchase detailed records of Americans’ movements, web browsing, and associations from public sources without obtaining a warrant, a practice the Intelligence Community has acknowledged raises privacy concerns and that has generated bipartisan opposition in Congress.
Amodei also made it clear that he was willing to walk away and would assist with any handover to other technology at the Pentagon. The dispute quickly escalated, with Emil Michael at the Pentagon, who was negotiating the deal, calling Amodei a “liar with a god complex.”
Once the deal initially fell apart, Sam Altman’s OpenAI swooped in, within hours, and agreed to the Department of War’s provision. A tweet from Altman:
“In all of our interactions, the DoW displayed a deep respect for safety and a desire to partner to achieve the best possible outcome.
AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement.”
The public quickly sided with Anthropic. A GPT boycott launched on social media with significant short-term impact: ChatGPT uninstalls jumped 295% in 24 hours, a QuitGPT campaign gained 1.5 million sign-ups in a similar timeframe, and Claude took the number 1 spot on the Apple App Store.
Altman seemed to panic at this, scrambling to publicly add new provisions to the deal, and publicly sharing an internal memo that included the following quote:
“We were genuinely trying to de-escalate things and avoid a much worse outcome,” he wrote. “But I think it just looked opportunistic and sloppy. Good learning experience for me as we face higher-stakes decisions in the future.”
Anthropic then re-opened negotiations, a move complicated by an internal blog post from Amodei obtained by the Financial Times, where Amodei suggested Anthropic had been frozen out because “we haven’t given dictator-style praise to Trump.” At the time of writing, these efforts have failed. The American government has gone ahead and formally designated Anthropic a supply chain risk, and Anthropic has sued.
The dispute has already come to symbolize broader concerns about the CEOs of large AI firms and the social and environmental impact of AI. So, it is worth asking what it really tells us. How much of this is an issue of optics and leadership style, and how much do we see meaningful differentiation in how governance and ethics in AI are handled by Amodei and Altman, respectively?
So, let’s talk about what this debate reveals. Then, much more importantly, we need to turn to what it hides.
What the Dispute Reveals
Tech employees know where they stand
Employees have more inside knowledge and visibility than the rest of us, even considering internal leaks and high public scrutiny. And, AI engineers have enormous freedom to work where they like and on the terms they want, giving them leverage over firm direction and considerable soft power. So, it is significant that what was initially an open letter from OpenAI employees supporting Amodei mushroomed further over the weekend following the Pentagon’s decision. By March 2, roughly 100 employees from OpenAI and another 800 from Google had signed the letter, which declares “we will not be divided” and calls for collective resistance to the American military over this pressure. OpenAI researcher Aidan McLaughlin also posted publicly on Twitter that he didn't think "this deal was worth it," getting nearly 500,000 views.
There has since been a second open letter from a broader tech coalition, signed by employees of Salesforce, Databricks, IBM, and Cursor, which urges the Pentagon to withdraw the Anthropic supply chain designation. This position was also supported publicly by Altman himself, Ilya Sutskever, and the Information Technology Industry Council.
Then, on March 9, an amicus brief was filed by a group of 37 researchers and engineers from OpenAI and Google, signing in their individual capacities, including Google Chief Scientist Jeff Dean, but not including the companies’ CEOs. The brief argues that the supply chain designation was "an improper and arbitrary use of power" and that in the absence of public law, contractual guardrails imposed by AI developers represent "a vital safeguard against catastrophic misuse." It also makes the pointed observation that the Pentagon could simply have ended the contract.
It’s clear that predictions of the death of employee activism in Silicon Valley were premature, and there is broad appetite for collective action to drive more responsible AI in defense. Meanwhile, most big tech CEOs are unable to see beyond their own self-interest, which should ultimately play to Anthropic’s advantage.
There is nuance here, though. OpenAI has seen a string of high-profile departures by safety researchers. Jan Leike, the former safety lead, said openly on his way out that “safety culture and processes have taken a back seat to product.” However, Anthropic’s own head of Safeguards Research, Mrinank Sharma, resigned just weeks before the Pentagon drama, vaguely warning that the “world is in peril” and citing “pressure to ignore what matters most.”
Anthropic has more savvy and principled leadership
Altman has long been accused of being slippery and dishonest, including in a recent biography, and has made questionable decisions over both “erotica” and adverts on ChatGPT, which Anthropic has exploited to its advantage, including in this fantastically witty Super Bowl ad. In this latest dispute, Altman comes off as sophomoric, inconsistent, and out of his depth on even straightforward ethical questions. His “opportunistic and sloppy” explanation is one I might expect if I found my teenage child ignoring a curfew, not from the CEO of one of the most consequential companies in existence. He is less the Oppenheimer of our age, more a Zuckerberg/Britney hybrid: half 'move fast and break things' and half 'oops, I did it again.’
In a similarly bizarre misstep a few weeks ago, Altman challenged critiques of AI’s energy use, claiming AI has “already caught up on an energy efficiency basis” when you account for the energy it takes to produce a human. He also extended it to human evolution broadly—pointing to “100 billion people who have ever lived” as the real training cost of human intelligence. In other words, human life should be placed on a par with server farm emissions.
Meanwhile, Amodei is carefully and consistently signaling that Anthropic is the more ethical and responsible player. This is clever, not least because it will reassure enterprise customers that Anthropic will be a reliable partner. I believe it's also wise for Anthropic to set itself apart by pushing back against pressure from the Trump administration. This is rare among American corporations in general and tech firms in particular, suggesting an understanding that long-term, intangible factors like trust and independence are highly likely to trump any short-term advantage. In other words, this is a calculated business decision, more than an “ethical imperative.“ That all said, it comes with considerable political and financial risk, and legal bills.
Ethics and governance are stronger at Anthropic—within limits
Anthropic was founded by people who left OpenAI over safety concerns, operates as a Public Benefit Corporation with a Long-Term Benefit Trust, and has invested more seriously in interpretability and alignment research (more on this below). These moves aren’t trivial.
OpenAI, by contrast, has wavered and shifted in its mission and governance structure. It is still technically a non-profit (though also transitioning to a public benefit corporation, there’s Musk litigation, blah blah). Anyway, this status means it declares its mission in annual statements. A great recent blog by Simon Willison tracked its evolution from promising, in 2016, to “advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return.” By 2024, this had morphed into “OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity.” In a nutshell, the firm has become more confident, commercial, and solipsistic, while dropping any commitment to prioritize humanity over financial considerations.
Looking more closely at internal safety investment and oversight, it also seems clear that Anthropic leads, but again, the situation is more complex than it might at first appear.
OpenAI announced “Superalignment” in 2023, with considerable fanfare. This was a dedicated team with the stated goal that if AI systems grew much smarter than humans, they would remain aligned with human values. OpenAI pledged 20% of its compute resources to it, under Ilya Sutskever and Jan Leike, two of the most respected researchers in AI safety. Within a year, Sutskever had left, and Leike resigned publicly, stating bluntly that OpenAI’s culture had shifted decisively toward shipping products rather than ensuring they were safe.
Meanwhile, Anthropic has invested heavily in “interpretability” research. Most AI companies test inputs and outputs, but don’t deeply understand what’s happening inside. Anthropic is trying to understand the internal mechanics of how Claude thinks and makes decisions—its biases, preferences, and how it accounts for consequences. This is difficult work, with no commercial payoff in the short term, and suggests something important about culture and values at the two firms.
From here on, it gets more complicated.
Anthropic has sought to differentiate itself by developing and promoting a “Constitution” with the stated intent of baking ethical values into the core technology, rather than relying on either compliance oversight or content moderation to block bad outputs. At least in theory, this means that models evaluate their own responses against a core set of ethical principles and revise them as necessary. The idea is that the values become intrinsic. The current version (released under Creative Commons, which is notable) has moved away from a list of rules toward something more like a philosophy — explaining why things matter, so the model can generalize to novel situations.
This is designed to seem both reassuring and innovative. However, there are still plenty of concerns, and many people view this as “safety theater.”
One, Anthropic and Thinking Machines have stress-tested the Anthropic Constitution by generating over 300,000 scenarios that force models to choose between competing principles. This research found models were "taking arbitrary positions within the trade-off, rather than intentional or consistent" ones.
Two, the deployment of the term “Constitution” is itself misleading if this is an internal self-evaluation approach voluntarily adopted by the company. An actual constitution would suggest democratic legitimacy and external enforcement, both of which are strikingly absent. Constitutions are supposed to protect the governed from the governors. Anthropic’s Constitution is instead a set of instructions from the owner to the product. Or, is it a tacit admission that the owner is no longer sure who or what is even doing the governing, and is seeking to protect itself from future consequences?
Then, we have the question of whether anthropomorphism is advisable at all. Is it a good idea to try to develop moral reasoning as an alternative to human accountability?
