NYTimes argues AI control should match how models actually work

In short: A New York Times opinion piece says we can control AI better if we stop treating AI models as either normal software or human-like minds.

What's going on

The article focuses on a basic problem in AI oversight. Many rules and safety plans assume an AI model is either like a simple computer program that follows instructions, or like a person with goals and intentions. The piece argues both views can lead to the wrong controls.

Today’s large language models, the systems behind many chatbots, learn from huge amounts of text and then predict what words should come next. You can think of them like an advanced autocomplete that can write full paragraphs. They can sound confident and even give detailed explanations, but that does not prove they understand things the way people do.

The article points to research from Anthropic about “introspection,” meaning whether a model can notice what is happening inside its own process. Anthropic reported limited evidence that some Claude models could detect certain injected internal concepts. The researchers stressed this was narrow and unreliable, and not like human self-awareness.

The piece also separates a text-only chatbot from an AI agent. An agent is a system that can take actions over time, like using tools, accessing accounts, and remembering past steps. Giving an AI those permissions can raise risks in a way that simple chat does not.

What to watch

The article suggests focusing less on what a model says about its reasons, and more on practical guardrails. That includes limiting what it is allowed to access, requiring human approval for important actions, monitoring behavior, testing under realistic conditions, and being able to stop or undo actions. It also highlights a difference between “alignment” (getting the AI to aim for acceptable goals) and “control” (limiting damage if it misbehaves), especially as more systems move from talking to acting.

Source: NYTimes