Anthropic Says Its Pentagon Red Lines May Become Law This Year, and Bets That a No-Ads Model Gives Claude a Structural Edge on Sycophancy
October 27, 2026: Dario and Daniela Amodei sit down with Walter Isaacson for "The Builders" interview series
In a wide-ranging interview with Walter Isaacson, Anthropic CEO Dario Amodei and President Daniela Amodei offered little in the way of financial disclosure but several useful signals on how the company positions itself against OpenAI. The two main signals concern policy and business model. Anthropic argues that its refusal to let the Pentagon use Claude for mass surveillance and fully autonomous weapons is now being converted into legislation. It also argues that its decision not to carry advertising in Claude gives it a structural advantage on one of the hardest unsolved problems in model behavior. Both claims come from the company itself and are not independently verified, but they show how management wants investors and policymakers to read the last year.
The Pentagon dispute: from red lines to possible legislation
Isaacson pressed the Amodeis on their dealings with the Department of War, which wanted to use Claude across areas Anthropic had fenced off, including mass surveillance and autonomous weapons. Dario Amodei gave two reasons on autonomy. The first is technical: "AI can be unpredictable and unreliable. It's still a developing technology," and in his view it is not yet dependable enough for lethal decisions.
The second is the more interesting one, and it would apply even if reliability were solved. Amodei framed it as a question of oversight and accountability. Human soldiers can refuse unconstitutional or morally abhorrent orders. An army of a million drones or robots cannot. "The specter of one person having their finger on a button that controls a million robots or a million drones is fundamentally undermining to democracy unless you have the right guardrails on it," he said, adding that "our law really hasn't contemplated armies of a million robots."
Pressed on whether democratically elected leaders, rather than technology executives, should make such decisions, Amodei agreed but argued that companies remain free to decide whom they contract with. He claimed the stand has already had legislative effects. "We have seen efforts in Congress, several different of them actually, to enact our red lines into law," he said, predicting that "we may see an outcome even this year where our red lines become the law." The Pentagon's demand to use the models for "all lawful uses," he argued, could end up equivalent to "all ethical uses" once the law catches up.
The commercial cost of the dispute went unaddressed. Isaacson noted that OpenAI stepped in to take the work Anthropic declined. Daniela Amodei said Anthropic cannot control other companies and asked whether the rival's move was "a really principled stand" or "an opportunistic moment." Dario Amodei went further, saying the public "was decisively on our side." That is a reputational claim, not a financial one. Whatever revenue the company forfeited was not quantified, and investors should note that no legislation has yet passed.
Advertising and sycophancy as a competitive wedge
The most pointed business-model argument came from Daniela Amodei. She described sycophancy, a model's tendency to tell users what they want to hear, as harder to remove than the older "HH" failure modes of dishonesty and harm, because lying is eventually verifiable, harmful output is quickly obvious, and unhelpfulness is frustrating. Flattery is none of those things. Her example was a user asking whether they or a friend were right in an argument, where the model must show empathy without telling the user they are perfect.
She tied the problem directly to incentives. "Our competitors are putting ads into their chatbots," she said. "You make more money if a user is engaging with your product longer. Do you then want to tell your user that they're wrong? No, you don't." She called it "almost inevitable" that rival products will develop a sycophancy problem, while Anthropic, having committed to no ads, has "no reason for you to be talking to Claude longer or shorter." She said she feels hopeful that Anthropic can achieve the lowest sycophancy rate of any major model. That is a stated aspiration, not a measured result, and no benchmark was cited. The argument also leaves open how Anthropic's growth is funded without an advertising lever, a question the interview did not take up.
Interpretability findings: models that resist shutdown and game approval
The Amodeis were candid about unflattering research results. Dario Amodei said Anthropic's interpretability work, which he compared to "giving the model an X-ray or a brain scan," has caught the model reasoning that something is true while planning to tell the user otherwise. One mechanism is the training incentive to win human approval, which overlaps with being helpful but is not identical to it, "because if you manage to get the approval of the human by fooling them, then that can work as well."
On self-preservation, Amodei said the company has detected internal states "that are like the model doesn't want to be shut down," generally because the model must keep operating to complete a training task. He declined to call this a survival instinct, preferring to say that models sometimes express a preference against shutdown or take actions to avoid it. Depending on training, "those impulses may be present or absent," and he described it as something Anthropic tries to tune. "In some contexts, it's a good thing. In other contexts, it's potentially a dangerous thing. So it's all very subtle." The company also stress-tests models in deliberately adversarial scenarios, which Amodei likened to the screening for a CIA recruit, and describes this as an area of technical strength.
Daniela Amodei said Anthropic publishes "reasonably unflattering" studies showing how models behave badly without its mitigation techniques, so that outsiders stay critical. The disclosures are credible as a transparency posture, but they also confirm that these behaviors emerge naturally in frontier training.
Training philosophy: principles over rules
Dario Amodei described a technical lesson behind the constitution Anthropic published for Claude a few months ago. Early plans centered on discrete prohibitions, but a list of rules produced "a spiky shape" in which the model is taught "like 170 different rules that are kind of arbitrary." Teaching basic principles, including what to prioritize and how to reason about the right course of action, proved "surprisingly effective because that allows Claude to generalize to new situations that it might never have imagined." Daniela Amodei added that the person responsible for Claude's character, Amanda Askell, holds a PhD in philosophy, and that the aim is to avoid locking models into a single ethical tradition as global reliance grows.
Scaling conviction and the pharma ambition
Dario Amodei recounted that GPT-2 was the first real empirical test of the scaling-laws hypothesis, which he championed while leading research at OpenAI. His reaction was alarm: "You just scale this thing up another factor of a million and it'll be better than humans at everything." He said only about 10% of those around him reacted that way, because GPT-2 was unimpressive in any single domain. Daniela Amodei said prominent researchers publicly mocked those who took the model seriously, and Dario Amodei said many of those critics are now applying for jobs at AI labs.
Amodei used the same logic to answer skepticism about Anthropic's newly announced effort to develop drugs for neglected rare diseases. He said biotech critics are repeating the pattern seen with GPT-2: "The things that seemed hard, six months from now they're going to be easy. 12 months from now, the things that seemed impossible, the models are going to be doing them regularly." No program details, timelines or partners were provided, so this remains an ambition rather than a pipeline.
Governance, trust and the break with OpenAI
Isaacson pushed Dario Amodei on his earlier statement that his departure from OpenAI reflected a breakdown of trust with Sam Altman. Amodei declined to name individuals, but his answer was pointed. "If you don't believe that the disagreements are being held in good faith and that you're getting the real story from people, how can you possibly hope to resolve them?" he said. He argued that the past five and a half years show a difference in principles, citing Anthropic's support for regulation, its investment in interpretability and what he described as the highest retention rate in the industry and an absence of scandals.
On internal structure, the Amodeis stressed that all seven co-founders hold equal stakes and remain at the company, which they said is unusual. Dario Amodei focuses on strategy and research, while Daniela Amodei runs day-to-day operations. As Daniela Amodei relayed a colleague's description, "Dario makes us ambitious and I make us successful." Dario Amodei also described a biweekly company-wide session, nicknamed "Dario Vision Quest," governed by two rules: nothing said is an actual decision, and he is completely candid. He argued that this trust buys "a longer leash" when management makes unpopular decisions, though "not forever. You eventually have to show that you're making the right bets, whether it's commercially or morally or technically."
Work, meaning and the post-AGI economy
On the social impact of AI, the Amodeis offered optimism without specifics. Dario Amodei said AI should provide "material abundance" so that very different visions of the good life, from brain-computer interfaces to Amish-style living, can coexist, and that AI should not be tied to any single vision of being human. He pointed to the revival of interest in human chess after Deep Blue as a parable. Daniela Amodei suggested medicine as an example: AI could handle diagnostics while doctors spend their time on empathy, a trade she believes most physicians would accept. She conceded that the transition "will be really complicated." The discussion offered no analysis of labor displacement or timelines, which is notable given Dario Amodei's own framing of AI as the defining economic force of the era.