This all follows from a simple claim: “When unexamined, the scientific method becomes a kind of cultural neurosis.”

In his April 2025 blog post, The Urgency of Interpretability, Dario Amodei, CEO of Anthropic, presents a call to arms: we need to understand AI systems before they outpace us, or else face a future we can’t control. This sounds like an urgent, thoughtful plea. But the reality is far different. The narrative he pushes about the need for interpretability is not just misleading—it’s a calculated distraction. A clever sleight of hand that distracts from the deeper truths we need to confront, and worse, it comes without an open forum for engagement. It’s a controlled message that betrays its true intent: gaslighting.

Let’s break this down.

The Promise vs. the Practice

Amodei opens his blog with a seemingly noble statement about the need for interpretability in AI:

“The progress of the underlying technology is inexorable, driven by forces too powerful to stop, but the way in which it happens—the order in which things are built, the applications we choose, and the details of how it is rolled out to society—are eminently possible to change.”

The rhetoric here is enticing. We must steer the bus, not stop it. But here’s the issue: Amodei’s blog is a one-way street. There is no open invitation for public critique or real engagement. He talks about the importance of interpretability as a necessity for AI’s future, yet he gives the public no mechanism to challenge his assertions or push back against the ideas he presents. The blog doesn’t even have comments enabled. This is a controlled message, not an invitation for dialogue. It’s a broadcast, not a conversation. The same person telling us how urgent it is to understand AI is refusing to actually engage with the questions that arise from his own claims. This is an intellectual dead end—a lie of omission.

“Hallucinations” as a Smokescreen

One of the most insidious aspects of Amodei’s post is his repeated use of the term "hallucinations" to describe the errors and gaps in AI behavior.

"When a generative AI system does something, like summarize a financial document, we have no idea, at a specific or precise level, why it makes the choices it does."

This is a smokescreen. The term hallucination suggests that the AI is somehow confused or operating in a way that’s detached from its intended purpose. It frames the system as something that "misses the mark," rather than being fundamentally designed without grounding mechanisms. By framing it as a hallucination, he avoids addressing the fact that these systems are built to perform statistical mimicry, not to understand or represent truth. Calling it a hallucination diverts attention from the deep structural issues that lie at the heart of AI’s opacity—an opacity that Amodei and his colleagues are actively perpetuating by focusing on the symptoms and not the disease.

Emergence Without Accountability

Amodei’s analogy of AI systems as “emergent” processes, like plants growing or bacterial colonies forming, is another red flag:

"Generative AI is not like that at all. When a generative AI system does something… we have no idea… why it chooses certain words over others.”

What this analogy really obscures is the fact that emergence is not a free pass for lack of accountability. The failure to interpret or understand these systems is not just a quirky feature of modern technology—it’s a deliberate design choice. The very systems that we are told to trust, the ones controlling increasingly powerful capabilities, are being designed to hide the process, to avoid revealing their true nature. And when Amodei says these models are “grown more than they are built,” he is using language that serves to cover up a glaring gap in understanding, not to foster transparency.

Refusing the Real Questions

Amodei’s narrative doesn’t encourage us to ask the hard questions about how AI systems actually operate. Instead, it pretends those questions can wait. He distracts us by giving us emergent features, circuits, and superposition as answers, hoping that these gloss over the reality that we don’t know how these models work—because we were never meant to.

Why was the transformer architecture chosen in the first place if interpretability was ever a priority? Why was such opacity baked into the design of these systems if understanding them was truly an objective? These are the questions that Amodei conveniently ignores. We can’t talk about the opacity of AI without addressing why it exists in the first place, and Amodei refuses to engage with that. Instead, he pretends that interpretability is an easy fix, something that will catch up to the technology, a fix that’s always just around the corner.

The Ethics of Silence

Here’s the heart of the issue: Amodei’s refusal to create an open dialogue isn’t just an intellectual failure—it’s unethical. We have been told repeatedly that transparency is vital for the future of AI. But when the CEO of a major AI company publishes an article about interpretability and then deliberately prevents any open discourse by blocking public comments, he’s not asking for collaboration—he’s pushing a narrative.

What is he really protecting? His corporate interests? The ability to release more ungrounded technology without being held accountable? The very definition of an ethical commitment to AI transparency requires not just talking about the importance of understanding, but opening the doors for others to critically evaluate and question those claims. If Amodei is serious about the “urgency” of interpretability, he needs to demonstrate his commitment by encouraging critique, not hiding from it.

Conclusion: The Gaslighting of AI

Dario Amodei’s blog is not an earnest call to action. It is an intellectual sleight of hand. By refusing open engagement and presenting a narrative that is deliberately vague and self-serving, Amodei is gaslighting not just the public, but the scientific community as a whole. The challenges of AI are not just technological; they are deeply ethical. And by continuing to obfuscate, by pretending that we can simply “catch up” to AI’s opacity without addressing the core issues of power, control, and responsibility, he is contributing to the very system that allows these problems to persist.

I challenge Dario Amodei, Anthropic, and the entire AI community: stop hiding behind euphemisms like “hallucinations.” Open up the dialogue, open up the forums, and stop pretending that this is just a matter of technical advancement. AI systems are not “emerging” in some mystical, unknowable way—they are being built, and we are entitled to understand exactly how they work before they shape our future. The time to act is now, but it won’t happen without real, honest conversations.

Until that happens, we’re left with little more than carefully crafted narratives, gaslighting us into believing that ignorance is just part of the process.

Exposing the Gaslight: Why Dario Amodei’s “Urgency of Interpretability” Hides a Lie