Anthropic’s trust problem

Anthropic made a compelling promise: powerful AI could be built by a company willing to take its risks seriously. For people uneasy about the race to release ever stronger models, that was a reason to pay attention.

It also set a higher bar. Once responsibility becomes part of what a company sells, its everyday decisions become part of the promise.

That is where Anthropic’s story gets harder to follow. As the company has grown, its safety commitments, use of training material and treatment of customers have exposed a familiar tension: principles become more expensive when the business gets bigger.

Start with the commitment that gave the promise some weight. In its 2023 Responsible Scaling Policy, Anthropic said it would pause training if its models outgrew the safeguards it could provide. The point was that safety might sometimes require accepting a disadvantage.

By February 2026, the framework had changed. Anthropic separated what it could do alone from what required industry cooperation, and described its new safety roadmap targets as public goals rather than hard commitments. It cited uncertain evaluations, slow government action and safeguards that one company could struggle to deliver.

Those are real difficulties. Anthropic retained safeguards and added reporting requirements. But the revision leaves an uncomfortable question: how much confidence should we place in a voluntary promise when the company making it also decides when to rewrite it?

The books case brought that question down from future risks to present conduct. In the June 2025 Bartz v. Anthropic ruling, the court treated the model training at issue as fair use. It did not accept the same defence for acquiring and retaining millions of pirated books in a central library.

That distinction matters. A lawful use later in the process did not excuse how the material was obtained. For a company asking to be trusted with the future, its treatment of people’s existing work is a fairly basic test.

Customers encounter a less dramatic version of the same trust problem. They choose a tool, build a routine around it, and expect to understand when it changes.

Anthropic’s April 2026 postmortem showed why some Claude users had struggled. Claude Code’s default reasoning effort had been reduced to improve speed. A caching bug discarded useful reasoning, and a separate prompt change hurt coding quality. All three changes were reversed or fixed.

The default change was disclosed, and the underlying API was unaffected. This was not evidence of a secret scheme to weaken every successful model. It was evidence that the product people relied on had become worse, and that diagnosing the problem took time.

Publishing the explanation deserves credit. So does raising usage limits in May. Still, customers need to know what they can depend on before they commit a working day to it. Clear defaults, predictable capacity and timely explanations matter as much as the model’s launch-day performance.

These episodes differ in severity. A product bug is not equivalent to acquiring pirated books, and revising a safety policy does not make all safety research insincere. Treating every decision as proof of bad intent would weaken the criticism.

What connects them is the distance between the trust Anthropic asks for and the choices people can actually inspect. Its safety work may be valuable. Its products may be useful. Neither gives the company a permanent exemption from scrutiny.

The practical standard is straightforward: commitments that remain meaningful under pressure, respect for the work used to build the product, and customers who can understand what they are paying for. When something changes, explain the trade-off clearly and early.

Anthropic earned attention by promising unusual care. Keeping that trust will depend on how consistently it applies that care when growth, cost and competition make it difficult.

Opinion · Revised September 24, 2026. Based on the linked court ruling and company statements. The original article has been shortened to focus on its central argument.