FINRA Flagged AI Hallucinations: What Regulated Finance Teams Must Do About Fabricated Numbers

On December 9, 2025, FINRA published its 2026 Annual Regulatory Oversight Report. Buried in roughly 90 pages of priorities, one definition now sits on the desk of every compliance officer at a member firm: a hallucination is when a model “generates information that is inaccurate or misleading, yet is presented as factual information.” The regulator gave examples. Hallucinated numbers. Fabricated regulation references. Inaccurate client data summaries. Those aren’t hypotheticals a vendor invented to sell governance software. They are the specific failure modes a self-regulatory organization decided to name in writing.

If you run finance or compliance at a broker-dealer, that changes the math on AI. The question is no longer whether your model occasionally invents a figure. The question is whether you can prove, on demand, that the number a registered rep or an analyst put in front of a client traces back to a real source.

What FINRA actually said, and what it didn’t

Start with the limits, because they matter. FINRA did not write a new rule. The 2026 report, like Regulatory Notice 24-09 before it, leans on a principle FINRA repeats often: its existing rules are “technologically neutral.” A hallucinated earnings figure in an AI-drafted client summary is governed by the same supervision and communications rules that have always applied. The tool is new. The obligation isn’t.

That principle cuts harder than a new rule would. Rule 3110 requires a supervisory system “reasonably designed” to achieve compliance with securities laws. The 2026 report’s framing is that “reasonably designed” now has to account for the known risks of the tools a firm chooses to deploy. Hallucination is, by FINRA’s own admission, a known risk. So a supervisory system that ships AI output to clients without a control that catches fabricated numbers is, by extension, harder to defend as reasonable.

The report’s expectation is concrete: firms should assess their compliance obligations before deploying generative AI, build governance to supervise its use, and keep a qualified human reviewing AI-generated content used for client interactions, financial analysis, or compliance purposes. Human-in-the-loop is the headline mitigation. We’ll come back to why that, alone, isn’t enough.

Why “have a human check it” quietly fails

Tell a reviewer to catch hallucinations and you’ve described the task, not solved it. The reviewer is reading prose. The prose is fluent and confident. A fabricated figure looks exactly like a correct one: that’s the entire problem with this class of error. There is no spelling mistake, no broken sentence, no red flag. A revenue number that’s off by 8% reads identically to one that’s exact.

So the reviewer falls back to spot-checking, and spot-checking scales badly. A monthly client review packet might contain forty figures across a dozen accounts. Verifying each one by hand against the source means reopening custodial statements, ledgers, and prior filings, which is most of the work the AI was supposed to remove. In practice, review degrades into a sniff test. Does this look about right? That passes the audit-trail box and catches almost nothing.

The deeper issue is where the number comes from. When a language model “reads” a 10-K or a brokerage statement and reports a figure, it is generating text that is statistically likely given the document, not retrieving a value from a structured record. Most of the time the likely text matches the real number. Sometimes it doesn’t, and nothing in the output marks which case you’re in. A reviewer cannot supervise a process whose failures are invisible by design.

The control FINRA’s framing actually implies

Read the 2026 report as an engineer rather than a lawyer and it points somewhere specific: the output has to be traceable, and the math has to be reliable. Not “reviewed.” Traceable. Every figure in front of a client should carry a path back to a source document or a ledger entry, and any number derived by calculation should come from a process you can re-run and get the same answer.

This is the difference between asking a model to be right and building a system that can’t present a number it can’t source. Those are not the same engineering problem, and only one of them survives an exam.

That’s the line Rexfin is built on. The reliable financial-modeling layer for AI doesn’t ask the language model to be the source of truth. It connects to where the numbers actually live (QuickBooks, Xero, NetSuite, Sage, the warehouse, or uploaded statements) and builds one reconciled financial model that ties out to the ledger. The AI then retrieves figures from that model instead of generating them from a PDF. When a calculation is needed (a margin, a growth rate, a ratio), it runs through a deterministic engine, not the LLM. The model orchestrates and explains. It does not do the arithmetic, and it does not invent the inputs.

The compliance payoff is the part that matters for FINRA. Every figure traces to source. Re-run the same question next quarter and you get the same number, because a deterministic engine produced it. When an examiner asks where a value came from, the answer is a citation, not a shrug. That’s how the platform turns AI output into something you can actually defend.

A concession, because it’s earned

None of this makes AI infallible, and any vendor who tells you it does is the one you should distrust. A reconciled model is only as good as the reconciliation. Source data can be wrong before any AI touches it. A deterministic engine computes exactly what you tell it, including a flawed formula, with perfect consistency. Tracing a number to its source proves provenance, not correctness: the underlying entry can still be a mistake.

What changes is the category of risk you’re left holding. You move from “the AI may have invented this and we’d never know” to “this figure came from a specific source through a known calculation.” The first is undefendable in an exam. The second is the ordinary work of finance: verifiable, auditable, and the kind of thing a supervisory system can reasonably be designed around.

The takeaway

FINRA put fabricated numbers in a regulatory report. That removes the option of treating hallucination as a tolerable quirk and reframes it as a supervision failure waiting for an exam. The firms that come through this well won’t be the ones with the strictest “please double-check the AI” memo. They’ll be the ones whose AI structurally cannot present a number it can’t trace, and whose math runs through an engine that returns the same answer every time.

If you’d rather show an examiner a citation than explain a guess, book a demo and we’ll walk through how the traceable model holds up under that exact scrutiny.

For more on how these errors slip through, see the pillar on AI hallucinations in financial data, or the specific cases of why AI misreads your 10-K and the verification checklist to run before AI numbers reach a board deck.