Claude Code and the Hallucination Problem

I’ve been using Claude Code daily — here’s what I think about the hallucination problem, and why I still back Anthropic

I use Claude Code almost every day. It’s part of how I run my data engineering work across multiple banking partners, and how I manage an AI-powered e-commerce agency. So when Claude Code recently gave me a response that had nothing to do with what I asked — serving up irrelevant, context-confused output mid-session — I noticed. And it bothered me.

But frustration is different from losing confidence. Let me explain why.

The hallucination problem is real — and well-documented

What I experienced isn’t a one-off. Claude Opus 4.8, released May 28, 2026, has a newly documented failure mode: in agentic workflows, the model can generate confident, detailed outputs that are simply wrong — or entirely disconnected from the task at hand. One reported case involved the model fabricating an active security threat during routine development, with a multi-agent audit finding zero evidence of any attack. The threat narrative was entirely model-generated.

In multi-agent pipelines, this compounds. Sub-agents return invented data. The main agent trusts it. Plans get built on false premises. By the time a human looks at the output, the hallucination has already propagated.

This isn’t just a Claude problem — it’s an agentic AI problem. The difference in a chat interface is that you catch the mistake in the next turn. In an autonomous agent running for an hour, you don’t.

Why I still trust Anthropic

Here’s the thing: Anthropic didn’t oversell Opus 4.8. Their own release notes described it as “a modest but tangible improvement.” That kind of honesty from an AI lab — especially one that just closed a $30B funding round and is reportedly targeting a 2026 IPO — is genuinely rare. Honesty was actually highlighted as one of the model’s key improvements in the release announcement.

Anthropic’s track record on ethics matters to me too. When the US Pentagon pressured AI labs to loosen safety guardrails for autonomous weapons and mass surveillance, Anthropic refused. OpenAI signed the contract. Anthropic got labeled a “supply chain risk” by the Pentagon — the first time that designation was applied to a US company. They wore it.

That’s not marketing. That’s a company making costly choices in alignment with its stated values.

The Pope, AI, and what it means that Anthropic was in that room

On May 25, 2026, Pope Leo XIV released his first encyclical: Magnifica Humanitas — on the protection of the human person in the age of artificial intelligence. It was a significant moral statement from the world’s largest religious institution on the ethical stakes of AI development.

Anthropic co-founder Chris Olah was personally invited to speak alongside the Pope at the Vatican presentation. Not as a sponsor. Not as a lobbyist. As someone whose work on AI interpretability — understanding what’s actually happening inside these models — earned him a seat at one of the most serious ethical conversations of the year.

Olah’s remarks at the Vatican were notable: he acknowledged openly that frontier AI labs, including Anthropic, operate inside incentive structures that can conflict with doing the right thing. That kind of honesty in a setting like that is worth something.

Critics like AI researcher Timnit Gebru pushed back — arguing that Anthropic’s safety positioning is marketing, and no different from OpenAI or Google. That’s a fair debate to have. But I think there’s a difference between imperfect and dishonest. Anthropic is imperfect. The hallucination issues are real. The agentic pipeline failures need fixing. But the company’s instinct — to be transparent about limitations, to hold ethical lines even at commercial cost — feels different to me than the alternatives.

What I’m doing differently as a practitioner

I’m not abandoning Claude Code. I’m building more verification checkpoints into my agentic pipelines. I’m treating model-reported states — especially anything security-related — as hypotheses, not conclusions. And I’m following Anthropic’s own stated loop: plan, execute, verify, report back.

The technology is imperfect. The company building it is imperfect. But they’re having the right conversations — even when those conversations happen in front of the Pope.