Today, the European Union’s AI Act transparency obligations take effect. From this morning, any organization deploying AI systems that interact with humans must disclose that fact. The rule is well-intentioned: if a chatbot is giving you financial advice, you should know it’s a chatbot. Sunlight, as the thinking goes, is the best disinfectant.

It’s also a disinfectant aimed at the wrong wound.

A working paper out of MIT Sloan this spring—one that just won the Swiss Finance Institute Outstanding Paper Award 2026—found something that should reframe the entire regulatory conversation. The researchers discovered that AI financial advice is, by their measure, surprisingly good. Not passable. Not “good enough for the masses.” Genuinely good. The catch, buried in the title like a warning label in 6-point font: especially if you ask the right questions.

That qualifier is doing a lot of work. And it’s the part nobody in Brussels—or at last week’s FINRA webinar on AI supervision—seems to be grappling with.

The Question Is the Advantage

The MIT study, authored by Taha Choukhmane, Weidong Lin, Matthew Akuzawa, and Tim de Silva, surveyed how Americans use AI for financial guidance. Half of respondents now turn to these tools. The paper’s headline finding—that the advice quality is high—has been dutifully covered by USA Today and shared widely on Hacker News, where the tone is predictably triumphalist. Finally, the argument goes, the little guy has access to the kind of counsel once reserved for Goldman Sachs clients.

But read the paper closely and a different picture emerges. The quality of the output is a function of the quality of the input. Ask a vague question—“How should I invest my savings?”—and you get vague, boilerplate advice. Ask a precise question that demonstrates you understand tax-loss harvesting, asset location, or the interplay between 401(k) contributions and IRA income limits, and the AI responds with a level of sophistication that would impress a CFP.

This is not democratization. This is a force multiplier for the already-literate.

A compliance officer at a midsize broker-dealer put it bluntly in the Q&A box during FINRA’s July 29 webinar on AI supervision: “We’re spending all this time building guardrails for bad advice. What about the client who doesn’t know enough to ask whether muni bonds make sense in their bracket? The AI won’t volunteer it unless prompted.”

The question sat unanswered as the moderator moved to the next slide.

Transparency Doesn’t Teach

The EU’s new rules, and the supervisory frameworks FINRA has been urging firms to adopt, are built on a specific theory of harm: that AI might mislead people by pretending to be human, or by generating advice that is unsuitable, biased, or flat-out wrong. These are real risks. But they are also the risks that are easiest to regulate. You can mandate a disclosure. You can require a firm to test its model’s outputs against a suitability standard. You can audit the training data.

What you cannot regulate is the gap between what a user knows to ask and what they need to know.

A product manager at a Berlin-based fintech messaged me this morning, as her team scrambled to finalize their AI Act compliance documentation: “We’re adding a banner that says ‘This is AI-generated content.’ Fine. But our power users already game the system with multi-step prompts that basically replicate a full financial plan. Our casual users type ‘what’s a good stock’ and get a list of large-cap ETFs. The transparency label doesn’t close that gap. It just makes us feel better about it.”

She’s right. And the gap is likely to widen. The MIT researchers found that AI advice improves markedly when users employ what they call “domain-specific framing”—essentially, speaking the language of finance. The people who already speak that language are disproportionately affluent, educated, and already well-served by the existing advisory industry. The people who don’t are the ones the technology was supposed to help.

The Uncomfortable Variable

None of this is an argument against AI financial advice. The tools are impressive, and they will improve. Nor is it an argument against transparency rules, which are sensible as far as they go. It is an argument that the variable that matters most—the user’s own financial literacy—is the one variable the policy conversation has decided to treat as fixed.

That’s a choice, not a necessity. If half the country is already asking AI for financial advice, as the MIT data suggests, then teaching people what to ask is arguably more urgent than teaching firms what to disclose. A public-education effort—call it prompt literacy for personal finance—would do more to close the advice gap than a thousand compliance checklists. But that kind of initiative doesn’t come with a regulatory filing deadline. It doesn’t generate headlines about fines or enforcement actions. It’s slow, unglamorous work, and it doesn’t fit neatly into a commission’s annual oversight report.

So we get transparency mandates instead. We get webinars about supervisory frameworks. We get a compliance industry that grows by the quarter while the underlying asymmetry—the one between the user who knows to ask about backdoor Roth conversions and the user who doesn’t know what a Roth is—quietly compounds.

The AI isn’t the black box. The user is. And no one is required to disclose that.

Sources