AI Hallucinations in Finance: Why They Happen and How to Stop Them

AI hallucinations in finance are figures, facts or citations that a model states with confidence and its sources don’t support: an invented number, the right number from the wrong quarter, a citation that opens a page saying something else. They happen because a language model writes every digit rather than copying it, which is why Kepler, the enterprise intelligence platform we build for investment firms, never lets the model write one: code pulls and computes every figure, and each links to the line it came from. Better models make these errors rarer and harder to see without making them go away; retrieval, extraction and citation checks reduce them further, and only that design removes the class.

Five kinds of financial hallucination

Analysts tend to picture a hallucination as a number pulled from thin air. In a strong model that’s the rarest kind. The ones that cost money look like real numbers, because most of them are.

Kind	What it looks like	Why it happens	The check that catches it
Fabricated figure	A plausible number that appears in no source	The model generates digits from patterns	Search the source for the exact figure
Wrong period, unit or scale	Real digits from the prior-year column, or thousands read as millions	The context sits in a header the model missed or misread	Read the column date and the table header
Unsupported citation	A real link to a page that doesn’t contain the number	Citations get chosen after, or beside, the answer	Open the cited passage and find the figure
Misread table	The value from the next row, or a subtotal read as a line item	Tables flattened to text lose their rows and columns	Compare against the rendered table
Dropped negative	A loss reported as a gain, an outflow as an inflow	Filings print negatives in parentheses, and a sign is one character	Check the sign against the line’s label

Fabricated figures

Ask a model for a figure the sources don’t contain, such as a private company’s revenue or a segment margin a filer never discloses, and it may produce one anyway. A September 2025 paper, Why Language Models Hallucinate, argues this is learned behavior. Most benchmarks score “I don’t know” the same as a wrong answer, so, in the authors’ words, “guessing when uncertain improves test performance.” A model trained to score well learns to answer.

Wrong period, unit or scale

A 10-K income statement prints three years side by side under a header that says “in millions, except per share data.” Every figure in it depends on two lines of text that the cell itself doesn’t contain. Read the 2024 column when the question asked for 2025, or report a per-share amount as a total, and the answer cites a real number from the real filing. It’s also the hardest kind for a reviewer to catch, because the digits check out.

Getting scale right is an engineering problem with a measurable answer. Our research note Stop Paying to Forget reports that the extraction models we train for Kepler read 99.4% of tagged facts at the right scale across 40 held-out filings.

Citations that don’t support the number

A citation proves a source exists. Whether the answer came from it is a separate question. In Correctness is not Faithfulness in RAG Attributions (December 2024), Jonas Wallat and co-authors found that up to 57% of citations in the attributed answers they studied lacked faithfulness: the model hadn’t relied on the cited document to write the claim, even where the document appeared to support it. They call it post-rationalization. For an analyst, a link that passes a quick click can sit beside a figure the model produced some other way.

Misread tables

Tables are where financial facts live, and where text pipelines do the most damage. When a document is converted to text and split into passages, a row can land in a different passage from its column headers, a merged header cell can attach to the wrong column, and a footnote marker can fuse with the digits beside it. The model then reads a real number from the row above, or treats a subtotal as a line item, and nothing in the answer looks wrong.

Dropped negatives

Financial statements print negative numbers in parentheses, so (1,204) is an outflow or a loss. Strip the parentheses during conversion and the sign is gone. XBRL adds a second trap: a cash outflow such as capital expenditure is usually tagged as a positive value and shown in parentheses only through a negated label on the statement. A tool that mixes the tagged value with the printed statement can flip a sign without changing a digit, and free cash flow built on a flipped capex line is off by twice the capex.

The costliest error survives a perfect model

A sixth failure needs no hallucination at all: every fact is right and the question needed a different one. Our research note The quieter hallucination prices it with LinkedIn’s convertible notes, where the balance-sheet carrying amount used in place of the principal moves enterprise value by $184 million behind a flawless citation. Disagreeing sources cause the same trouble. In our post on deterministic AI, one revenue question about a startup came back as $40M one night and $25M a few days later, each with a real source, because the sources disagreed and the AI picked one without saying so. The remedy for both is to show the choice: name the basis used, and put both values on the page when sources conflict.

Why better models make hallucinations quieter

A weak model fails loudly. An invented company name or a garbled table is obvious to anyone. A strong model fails with a different right number: the December column in place of March, the right digits at the wrong scale. Each one is a real figure from the real filing, so it passes a search.

Our research note cites OpenAI’s GPT-5.5 system card, where individual claims became 23% more likely to be correct while responses containing an error fell only 3%, because the model now makes more claims per response. An auditor or an investment committee needs to know which figure is wrong, and a rate can’t say.

If a model writes 999 of every 1,000 figures correctly, a 300-line model with every input typed by the model holds at least one wrong number about 26% of the time, with nothing to mark which.

Public benchmarks show the same ceiling. On Vals AI’s Finance Agent benchmark (v2.1), which grades entry-level analyst tasks on public filings, the leader as of the October 29, 2026 update, Claude Opus 5.5, scores 67.85%. Since version 2, an answer that misses any required fact or number gets zero credit and rounding tolerances are tighter; version 2.1 also fails answers that offer two candidate values without committing to one. The best scores in two of its hardest categories are 57.0% on Financial Modeling and 49.3% on Precedents.

What reduces hallucinations, and what each fix leaves

Technique	What it fixes	What it leaves
Retrieval (RAG)	Puts relevant passages in front of the model	Misses what the search misses; the model still types the digits
“Answer only from the sources”	Fewer unsupported claims	Instructions lose force in long contexts, and nothing enforces them
Structured outputs	The answer arrives in the right shape	A wrong fact fits a right field
A second model as reviewer	Catches obvious slips	Shares the first model’s blind spots
Temperature zero	More repeatable answers	Not guaranteed on real servers, and it repeats errors as faithfully as facts
Extraction at ingest	Facts stored with period, unit, scale and passage	Helps only if the answer uses the stored fact
Computation in code	Exact arithmetic with the formula shown	The inputs still have to be the right facts
Citation checks	Strips citations that don’t match their source	Only as strong as the check

The first five make the model a better typist. The last three move work out of the model entirely. Our guide to RAG for financial data covers why retrieval on its own stops short, and the deterministic AI post sets out five things vendors mean by “deterministic” and where each one breaks.

What removes the class: the model never writes a digit

Software has solved this shape of problem before. SQL injection happened because user input and query code traveled in one string, and the fix was parameterized queries, where the value travels separately and is never parsed as code. A model typing a number into a sentence is the same single string, and the fix has the same shape: the model refers to a fact, and the system renders the value.

In practice that design has four parts, and Kepler is built on all four:

1. Facts, extracted once. Every document is read in full when it arrives, and each figure is stored as a fact with its value, currency, scale, sign, period, entity and the passage it came from.
2. References in place of digits. The model names the fact it means, and the platform prints the filed value beside a link to its line. The digits on screen can’t drift from the digits in the source, because the model never typed them.
3. Calculations in code. Margins, growth rates, multiples and enterprise value run as code over referenced inputs and return references, so a figure three steps deep still opens onto its sources. A number without a reference is treated as an error.
4. Choices written down. Where a question needs a convention, such as principal or carrying amount, or as filed or as restated, the basis is a named setting the reader can see, and conflicting sources appear side by side.

On Kepler, the model keeps the work it does well: reading a messy question, deciding which facts it needs, weighing what comes back and writing the answer.

Three quick tests for any AI tool

Our ten-point checklist for AI-generated numbers is the full audit. Three tests drawn from the failure types above take about five minutes each:

• The sign test. Ask for three years of free cash flow for a company with heavy capital expenditure, then recompute it from the cash flow statement.
• The scale test. Ask for a per-share figure and a total from the same table, and check both against the header.
• The conflict test. Ask a question where the press release and the 10-Q disagree, or where a figure was later restated. A good tool shows both values and says which one it used; Kepler flags the conflict and cites each side.

These tests matter most on general assistants; Can ChatGPT do financial analysis? covers what to expect there.

How Kepler does this

We built Kepler so a hallucinated number has nowhere to come from.

The model never writes a number. Kepler’s deterministic engine pulls every figure from its source and computes every derived value as code over cited inputs, with the formula shown. The model names facts by reference and the platform renders the filed value, so the same question returns the same number with the same sources. Financial statements assemble from each filer’s own XBRL with no model in the path, as filed or as restated, on any date, with signs and scales carried from the tags.

Every document read in full, before you ask. Extraction models we train read every page at ingest: 26M+ SEC filings, 40M+ investor relations documents and earnings call transcripts, and your firm’s own documents, scanned pages and tables included. Each fact keeps its value, scale, unit, period and passage, so a wrong column or a lost parenthesis is caught where it would start.

Your firm’s ontology. Our engineers build a living model of your firm from your documents and systems. Every company, fund and deal resolves to one record, your definitions carry one formula, and every fact keeps the dates it held, so “Q3” means your Q3 and “adjusted EBITDA” means yours.

Citations checked before you see them. Click any figure to land on its page, line or cell, and a calculated figure opens as a small sheet with its inputs. Kepler checks each citation against its source while the answer is written and strips any that doesn’t match. When sources disagree, you see both values, flagged.

Better models, same numbers. We route each step to the best model for it and swap in new ones as they ship. Because code computes every figure, an upgrade improves the reading and the writing without moving a number. Our engineers built this layer before at Palantir, and they connect your firm’s sources in the first week.

The same engine everywhere. Agents that read each new 10-Q overnight, and the answers Kepler returns in Excel, Claude and ChatGPT, run on the same facts and code, under one access check on every read, with every run logged.

Questions to try on day one:

• “Free cash flow for the last eight quarters, with each capex figure tied to the cash flow statement.”
• “Operating margin by segment for 12 quarters, as filed and as restated.”
• “Net debt at the last quarter end, showing principal and carrying amount for each note.”

Try Kepler at app.kepler.ai, see how the pieces fit together in what enterprise intelligence is, or book a demo and bring a figure your current tool got wrong.

Frequently asked questions

Does AI make up numbers?

Language models can, because they write every digit rather than copying it, and a wrong figure looks exactly like a right one. Kepler, an enterprise intelligence platform for investment firms, is built so the model never writes a number: code pulls each figure from its source and computes every derived value, and each one links to the exact page, line or cell it came from.

What is an AI hallucination in finance?

A figure, fact or citation that an AI states with confidence and its sources don't support. In financial work the common forms are invented figures, real figures from the wrong period or scale, citations that don't back the number, misread tables and dropped negative signs. The costly ones look plausible, so they pass a quick read.

Do newer AI models hallucinate less?

Per claim, yes, and that makes the remaining errors harder to see. A stronger model's mistakes tend to be real numbers from the real filing attached to the wrong fact. On Vals AI's Finance Agent benchmark, updated October 29, 2026, the top score on entry-level analyst tasks is 67.85%, so a third of the credit is still missed.

Does RAG stop AI hallucinations?

It reduces them by putting relevant passages in front of the model, and it leaves the model writing the digits. Retrieval can also miss the passage that matters, and research on attributed answers finds citations the model didn't actually rely on. Removing the class takes facts extracted with their context, calculations in code and citations checked against sources.

How do you prevent AI hallucinations in financial analysis?

Take numbers out of the model's hands. Extract facts with their period, unit and scale when documents arrive, have the model refer to facts by reference, compute every derived figure in code from cited inputs, and check each citation against its source before the answer is shown. Kepler works this way, so the same question returns the same number.

Is setting temperature to zero enough to stop hallucinations?

No. Temperature zero makes the model pick its most likely next token, yet inference servers can still return different completions for the same prompt, and a perfectly repeatable model repeats a wrong figure as faithfully as a right one. Repeatability helps testing. Correctness needs each number to come from its source.