AI makes far fewer mistakes reading company filings when the information is provided in XBRL — a wonky reporting format that may suddenly look a lot more valuable.
A recent academic study found artificial intelligence is far better at pulling numbers from annual reports when filing data is provided in XBRL, a structured digital format, rather than plain text or HTML — a finding that could give fresh ammunition to supporters of stricter structured-data reporting.
Researchers who reviewed 5,000 annual reports from 2014 through 2023 found AI’s overall error rate fell to 9.19% when it worked from XBRL context, compared with 15.75% for HTML and 18.24% for plain text.
The Bigger Problem Wasn’t Hallucinations
The striking finding was that the main problem was not AI hallucinations.
Instead, the technology often botched numbers companies had already disclosed — pulling the wrong figure, misreading the scale, or confusing one line item for another.
“That was particularly good in our research — the methodical examination of the types of mistakes, the amount of mistakes,” Ariel Markelevich, a Suffolk University accounting professor and one of the paper’s authors, told Thomson Reuters in an interview on March 24, 2026. “Having XBRL as the source of the information and having it in a structured format did seem to help AI make fewer mistakes.”
That matters because AI is rapidly making its way into investing, accounting and financial analysis. If the machine gets the number wrong at the start, everything built on top of it — models, research notes, forecasts and trading decisions — can go sideways.
XBRL’s Biggest Edge: Cutting Scale Errors
The research found XBRL’s biggest advantage was in virtually eliminating scale errors — the kind of mistake that happens when AI reads a figure reported in thousands or millions at the wrong magnitude.
Simply put: the bot sees a number, but misses the fine print telling it whether that number is in dollars, thousands or millions.
“When you think about it, it actually makes sense,” Markelevich said. In text or HTML filings, AI may have to connect a figure to a note elsewhere in the table saying amounts are reported “in thousands” or “in millions.” In XBRL, by contrast, “one of the rules is that you put the amounts in dollars, in raw dollars.”
That cleaner structure appears to do the trick. The paper found XBRL virtually eliminated scale errors and reduced some other mistakes, though it did not eliminate all errors.
The Mistakes Weren’t Random
The study also found the errors weren’t random.
Bigger companies had more mistakes. More complex firms had more mistakes. And AI struggled more with information tucked into footnotes than with the main financial statements.
“Larger companies had more mistakes,” Markelevich said. “Those would be the ones that users would likely be most interested in.”
That is a problem for finance professionals because the biggest and most complex companies are often the ones investors care about most.
Newer AI Wasn’t Always Better
The research also pushes back on a common assumption in the AI boom: newer models do not always mean better results.
Markelevich said the team reran the analysis using a newer version of Google’s Gemini and found it made more mistakes in some areas, especially with HTML filings.
“Newer AI models don’t mean that they’re better,” he said.
Why XBRL May Suddenly Matter More
Another notable finding: hallucinations were relatively rare in this setting. The paper said hallucination rates stayed under 2% when the model was given filing context.
That means the bigger danger in financial reporting may be subtler: not fake facts, but believable-looking bad reads of real ones.
A made-up number might stand out. A wrong number lifted from a real filing might not.
For XBRL supporters, the study lands at a useful time. Companies have long complained that structured-tagging rules are costly and time-consuming. But if AI works better with cleaner, standardized inputs, those same rules could look more valuable in an AI-driven market.
Markelevich stopped short of saying XBRL will always be necessary, but said the logic is straightforward:
“The better the input, the better the output.”
His broader message was not anti-AI — just anti-hype: “Use it, it’s powerful. But just be cautious a little bit.”
[The research paper, Can AI be trusted with financial data?, was written by Marcelo Farr of Universidad Adolfo Ibáñez, William C. Johnson of the University of Massachusetts, Ariel Markelevich of Suffolk University, and Alexis Montecinos of Suffolk University.]
