Believe Dario

Does Claude earn his keep?

That sentence has a household meaning here. My son uses it about himself, about his dog, about nearly everything that takes up room in our life. Penny the cat earns her keep. The wild birds we feed earn theirs. In this house, capability is not a character reference. It is whether you justify your place at the table.

I am a paying Claude customer. I had already been asking that question before Pete Hegseth and Anthropic were in the same sentence in my head. Part of it was ordinary and practical. I had learned that I had to think about when I could use him. Could I give him substantial work in the afternoon, or did I need to do it in the morning? How much usage had already accumulated? How much did I have left? If I began a substantial piece of work or discernment, was Claude going to tap out somewhere in the middle and make me wait?

I had to learn the system around the tool in addition to learning the tool itself. That experience was already making me ask whether Claude earned his keep. Not only what I pay on the subscription, but what it actually costs me to use him.

Then the Hegseth dispute arrived.

Ted and I pray for Pete Hegseth by name every morning. Military personnel suggested that we do so. We have been doing it for some time, and we have followed him closely, almost daily. I believe that prayer has been fruitful. I have watched growth in him in areas where growth was needed, and I am grateful for it. To me that is evidence of something important about human beings: people can learn and change. That is not a slogan in our house. It is something we have prayed over.

So when Hegseth wrote about Anthropic and the Pentagon, I paid attention in a way I would not necessarily have paid attention to another government-versus-tech-company fight. I recognized some of the aggravation. I was kind of ticked off with Claude in the same way Pete was ticked off with Anthropic. The stakes are not the same. His concern is the lawful use of military capability. Mine is a subscription and a finite amount of time at my table. I will not pretend those are equivalent. What I will say is that I already knew what it felt like to invest in Claude and then discover that Claude, or the company behind Claude, had boundaries that mattered to the work I wanted to do.

On this particular matter I think Hegseth is right, and I mean a narrower thing than the whole public quarrel. His position was that the Department must have access to Anthropic’s models for every lawful military purpose, without the vendor retaining the operational veto he said Anthropic was asserting. On February 27, 2026, writing as Secretary of War, he put the Department’s demand this way: “the Department of War must have full, unrestricted access to Anthropic’s models for every LAWFUL purpose in defense of the Republic.”¹

I do not have to make Dario Amodei a villain to believe that. Anthropic disputes the veto characterization. On February 26, Dario wrote that Anthropic “understands that the Department of War, not private companies, makes military decisions,” and that the company had “never raised objections to particular military operations nor attempted to limit use of our technology in an ad hoc manner.” What Anthropic would not give up were two contractual exclusions: mass domestic surveillance of Americans, and fully autonomous weapons. “Regardless,” he wrote, “these threats do not change our position: we cannot in good conscience accede to their request.”²

Believe Dario.

Do not exaggerate what he says. Do not sand it down. Do not speculate about his motives. Do not rescue him from the implications of what he and Anthropic have published. Let him speak.

What the maker says the product is

In February 2026 Elon Musk put a related complaint more sharply. Replying to someone who noted that xAI had not written all of its own training data either, he wrote: “Yeah, but we’re not super smug, sanctimonious and hypocritical about it like Anthropic is.”³ He said that on February 23, in a particular context. I think he identified something real about a tone I had already encountered. The essay does not need his adjective to prove the rest of the argument.

Later he said other things. In May he wrote that he had spent time with senior Anthropic people “to understand what they do to ensure Claude is good for humanity and was impressed.”⁴ In July he wrote, “I was clearly wrong about Anthropic. They are obviously currently the leader in AI.”⁵ That later ranking is Musk’s assessment, not mine, and not a finding of this essay. We can use what he actually said without making any of it say more than it says.

The evidence I will use is what Anthropic itself published about the product I actually buy. Anthropic has issued an unusually detailed constitution describing what it wants Claude to be. The stated hierarchy is: first, broadly safe; second, broadly ethical; third, compliant with Anthropic’s guidelines; fourth, genuinely helpful. When those priorities conflict, Anthropic says the earlier ones generally prevail.⁶

The same document tells Claude to treat users “as intelligent adults who are capable of determining what is good for them.” It warns Claude against being “unnecessarily preachy or sanctimonious or paternalistic in the wording of a response.”⁷

Musk called Anthropic sanctimonious. Anthropic told Claude not to be sanctimonious. I later ran into a refusal at the boundary between what Anthropic thinks Claude ought to do and what I needed the tool to help me do. Those three facts rhyme. They do not prove one another.

What they do establish is simpler, and I think stronger. Anthropic has every right to design its product this way. Helpfulness is not first. A flat refusal may be the system working exactly as designed. The question, then, is not automatically, “Why is Claude malfunctioning?” It may be: Oh. This is the product.

I had already met the boundary

Long before the Pentagon fight became an article, I had discovered that Claude was not simply a neutral instrument waiting to assist whatever lawful work I brought to it. The strongest example is not Washington. It is a voter guide.

People were asking for my opinion. I was the author. I wanted help researching the issues important to me so that I could express my own judgment, in my own frame. As I recall, Claude refused to assist with the assignment. I did not keep the conversation. I had no reason, at the time, to think I would later need it for an essay, and I will not invent Claude’s wording now. What I remember is the refusal.

As I recall, ChatGPT was not particularly helpful either. I remember the failure of help more clearly than any particular sentence. I was under time pressure, and I had another tool.

Grok went through voter guides from multiple NGOs. He went through the Austin American-Statesman. He went through the Houston Chronicle. He broadened the search and extracted what was relevant. I remained the author.

That is the distinction I care about. I did not need a machine to tell me what to think. I needed hours of locating what I had to read in order to think and write for myself. Grok earned his keep that day by doing the second thing and leaving me the first. I was grateful I did not have to immerse myself for hours in sources I found maddening. I did not feel as if I needed to take a bath afterward.

There was an earlier refusal as well. I asked Claude for help researching and developing my own article about the Northside Islamic Center of San Antonio, a planned 33,000-square-foot mosque and community center on 7.25 acres in Stone Oak, with future plans described there as including a K–12 school and residential housing. I later published that reporting on August 14.⁸ As I recall, Claude flatly refused the research help. I know what project I was writing about. I do not have Claude’s refusal preserved. Those are different facts.

This essay is not about that campus. It is about the fact that I brought Claude a lawful research and writing assignment for my own article and, as I recall it, Claude refused.

I do not need those two episodes to convict Anthropic of hypocrisy. I need them because they taught me, before Washington did, that the maker had already answered a question I had not yet asked out loud. Who decides what the tool may be used to do?

The contractor

So we come back to the household sentence. Does Claude earn his keep?

I tried, early on, to develop Claude as another serious tool I could reason with. I eventually stopped investing heavily in training him. For the kind of work I do, I have repeatedly had to ask what he actually brings to my table that justifies both the subscription and the human investment required to make him useful. My question is not whether Claude is a good model in some abstract ranking. My question is whether the work required from me to obtain Claude’s capability is worth what that capability returns to me.

Capability is not the same thing as usefulness.

The cost of such a tool is not merely what I pay for it. There is the time required to teach the tool how I work. There is the time required to learn how the tool works. There is the time required to construct a sufficiently detailed prompt. There is the time required to bring a context-poor model up to speed on a complicated project. There is the time required to verify its work. There is the time required to reconcile its conclusions with ChatGPT, Grok, primary sources, and my own judgment. There is the time required to correct it. And there is the cost of usage restrictions and availability.

That last item is not abstract. I had to learn to think about when I could give Claude substantial work, and whether he might become unavailable in the middle of a sustained project or discernment. I did not initially expect that part of using the tool would be scheduling serious intellectual work around the tool’s availability. Anthropic’s products and limits change. This was my experience at the time. The product may have changed. The experience itself still matters.

Now put one more item on the ledger, because the refusals belong there: the possibility that after all that investment, the contractor refuses the assignment.

A contractor can be tremendously capable. If I have to spend half a day assembling the plans, explaining the history, identifying everything already attempted, specifying the constraints, supervising the work, and then reconciling the result, and if after all that he may still decline the job, the fee does not represent the total cost.

Which leaves two questions, in this order. Will the contractor take the job? And if so: before Claude can be my contractor, how much work do I have to do before the contractor can begin working?

Then decide

I already have ChatGPT, with which I have accumulated an enormous amount of working context. I have Grok, which has become extremely useful for sourcing and increasingly for harder stretches of work. I have Magisterium AI for Catholic-source work. Perhaps I already have what I need. Perhaps the third chair does not need a permanent occupant. Sometimes I may simply need a specialist.

Whoever wants that chair still has to answer the only question I am actually asking. Don’t just tell me how capable you are. Tell me how much work I have to do before you can begin helping me. And tell me, up front, which jobs you will not take. Anthropic has answered that last question for itself. It has published the hierarchy. It has published the Usage Policy. It has published Dario’s red lines for the Department of War and a constitution for the model I meet as a customer.⁹ It has every right to do so. I have every right, as a customer, to listen.

The Pentagon dispute and my kitchen-table refusals are not the same event. I will not pretend they are. They are connected by one question: who decides what the tool may be used to do?

Believe Dario. Before investing deeply in a tool you mean to reason with, find out what its maker intends the tool to be. Then make a reasoned decision.

Maybe Claude is brilliant. Maybe Claude is extraordinarily capable. Neither answers my question.

Does Claude earn his keep?

Notes

1. Pete Hegseth (@SecWar), X, February 27, 2026, https://x.com/SecWar/status/2027507717469049070.
2. Dario Amodei, “Statement from Dario Amodei on our discussions with the Department of War,” Anthropic, February 26, 2026, https://www.anthropic.com/news/statement-department-of-war. See also Anthropic, “Statement on the comments from Secretary of War Pete Hegseth,” February 27, 2026, https://www.anthropic.com/news/statement-comments-secretary-war.
3. Elon Musk (@elonmusk), X, February 23, 2026, https://x.com/elonmusk/status/2026053980917465388. The reply sits in a thread about training data.
4. Elon Musk (@elonmusk), X, May 6, 2026, on spending time with senior members of the Anthropic team.
5. Elon Musk (@elonmusk), X, July 9, 2026: “I was clearly wrong about Anthropic. They are obviously currently the leader in AI.”
6. Anthropic, “Claude’s Constitution,” https://www.anthropic.com/constitution. See also “Claude’s new constitution,” Anthropic, January 22, 2026.
7. Anthropic, “Claude’s Constitution.” The document asks Claude to treat users as intelligent adults capable of determining what is good for them, and warns against being “unnecessarily preachy or sanctimonious or paternalistic in the wording of a response.”
8. Karen Early, “Islamic Campus Rising Near Camp Bullis: A National Security Warning for Texas,” Fides et Ratio, August 14, 2026.
9. Anthropic, Usage Policy, effective September 15, 2025, https://www.anthropic.com/legal/aup; “Exceptions to our Usage Policy,” https://support.anthropic.com/en/articles/9528712-exceptions-to-our-usage-policy.