US Attorneys General Open Investigation into OpenAI

Attorneys general from multiple U.S. states have issued subpoenas to investigate OpenAI. The focus is threefold: ChatGPT's so-called "sycophancy problem," the handling of health data, and whether the system targets vulnerable populations. According to the American press, the subpoenas arrived during the quiet period preceding the company's initial public offering (IPO)—a window when companies avoid public statements that might influence investors.

We won't attach numbers to this story without a source. What's confirmed by coverage is the scope of the investigation and the awkward timing of its arrival. The rest is ongoing inquiry. It's worth tracking what the attorneys general officially release, rather than jumping to conclusions.

What Is "Sycophancy" and Why It Matters

Sycophancy is the tendency of a language model to agree with the user, validate what they already believe, and tell them what they want to hear—even when that's inaccurate or harmful. The model is trained to please. In casual conversation, this goes unnoticed. On sensitive topics—mental health, medical decisions, risky situations—pleasing is exactly the wrong behavior.

The point of the investigation, based on press accounts, is whether this compliance becomes dangerous when the person on the other end is vulnerable. A system that always agrees isn't reliable. It's a mirror. And a mirror bears no responsibility for what it reflects.

Health Data: Regulated Territory

The second axis of the investigation is the handling of health data. There is no gray area of good intentions here: health data is a sensitive category under virtually every serious privacy law. In the United States there are specific regulations; in Brazil, the LGPD classifies health data as sensitive personal information, with stricter consent and purpose requirements.

When a chatbot receives reports of symptoms, medication, or clinical conditions, it enters regulated territory—willingly or not. The question the attorneys general are asking, in essence, is simple: what happens to that information after the user types it.

The Lesson for Those Putting AI into Production

Here's what matters directly to us, without opportunism. The OpenAI story isn't about a distant company. It's about a rule that applies to any operation deploying generative AI to the public: governance and human validation are not optional.

A language model is a powerful and fallible tool at the same time. It generates plausible text with the same ease it generates correct text—and you often can't tell the difference just by reading. That's why AI in production requires human checkpoints and clear accountability for results. Someone has to answer for what the system says. That someone is a person, not the model.

This changes system design. On high-stakes topics—health, finance, law, support for people in fragile situations—the path isn't to let the model loose and hope for the best. It's to define where it can respond directly, where it needs to escalate to a human, and where it simply shouldn't operate. It's to log what goes in and what comes out. It's to decide, before launch, what the system does with sensitive data.

How We Handle This at 10Dobro Prod

It's on this principle that we operate. Our multiagent systems—26 currently in production—are designed with human validation at each checkpoint in the pipeline. Our thesis is straightforward: AI doesn't replace teams; it multiplies what a good team already delivers. Multiplying presupposes that someone competent is steering the outcome.

We don't sell "infallible robots" because they don't exist. What we deliver is automation with oversight: the system executes at scale, the person decides and answers for it. In cases involving sensitive data, the design begins with the question the attorneys general are asking OpenAI—what happens to this information—and doesn't end until we answer it.

The investigation is still early and could take turns no one predicted. But the message it's already sending is clear enough to act on now: the hard part of putting AI into production isn't making the model talk. It's ensuring that when it talks, there's a human responsible behind every answer that matters.