AI Hallucination Risk in Equity Research: What to Verify Before Publishing
An AI model can write a confident, well-structured paragraph that is simply wrong, a fabricated figure, a misattributed quote, the wrong fiscal quarter, with no visible sign that anything is off. That gap between fluent and correct is the risk. Here is what to check before an AI-assisted note reaches a client.
Educational, not investment advice. This is a verification workflow, not a claim about any specific AI product's accuracy rate.
Why fluent is not the same as correct
A language model is trained to produce text that reads naturally, not to look up a fact and repeat it exactly. When you ask it for something it was not given verbatim, like a specific quarter's EPS from memory, it generates the most statistically plausible answer given the pattern of the question. Most of the time that answer is right, because the model has seen a lot of correct financial writing. Sometimes it is not, and the sentence around the wrong number looks exactly as confident as the sentence around a correct one. There is no built-in signal telling you which is which.
That is the core problem for equity research specifically: the entire value of a report rests on the numbers being right, and the failure mode of an AI-drafted note is invisible to a reader unless someone checks it against the source.
What hallucination looks like in a research note
It rarely shows up as an obviously absurd claim. It shows up as small, plausible errors: a margin figure off by a percentage point, a quarter mislabelled as the one before it, a subsidiary name swapped for a similar-sounding peer, a quote attributed to the CFO that the CEO actually said on the earnings call. Each of these is easy to miss on a skim read and damaging if it reaches a client unverified.
Every number, against the filing
EPS, revenue, margins, debt figures: trace each one back to the exact line in the BSE/NSE filing or annual report. Do not accept the AI's restatement of a number as confirmation of itself.
The fiscal period
A common failure mode is a model attributing Q3 figures to Q2, or mixing standalone and consolidated numbers. Confirm the period and the basis (standalone vs consolidated) match the source.
Quotes and attributions
If the draft quotes a management statement or an earnings-call line, confirm the quote exists verbatim in the transcript and is attributed to the right person and the right call.
Named entities
Subsidiary names, peer companies, promoter names: models occasionally substitute a plausible but wrong name, especially for less prominent entities. A quick name-by-name scan catches most of these.
Derived and calculated figures
Ratios, growth rates, and anything computed rather than quoted directly (like YoY growth or a margin percentage) carry higher risk. Recompute them yourself rather than trusting the model's arithmetic.
Grounding reduces the risk, it does not remove it
The single biggest reduction in hallucination risk comes from giving the model the source document directly rather than asking it to recall a figure from memory. A model summarising a filing you pasted in is working from the actual text in front of it, closer to reading comprehension than recall, and is far more reliable than one asked "what was Company X's Q3 revenue?" with no source attached. This is often called grounding or retrieval-augmented generation, and it is the reason a good AI research workflow always starts with the primary filing, not a bare question.
It still does not eliminate the risk entirely. Models can misread a table, conflate two nearby figures, or drop a decimal point even with the source in front of them. Grounding changes the odds, not the requirement to verify.
Where this sits in your compliance obligations
An unverified hallucinated figure in a published report is a factual-accuracy problem regardless of what produced it, and SEBI holds the registered analyst accountable for it either way. The tool that drafted the sentence is not a defence. We cover the broader accountability question, and where AI fits into your workflow without creating this exposure, in can SEBI Research Analysts use ChatGPT to draft reports.
A five-minute pre-publish check
You do not need a heavy process for this. Before an AI-assisted note goes out: read it once against the source filing with the two documents side by side, highlight every number and named entity as you confirm it, and only then move to your usual review step. That single pass, done deliberately rather than as part of a general read- through, is what catches the small, plausible errors that a fluent draft is designed to hide.
FAQ
What is AI hallucination in the context of financial research?
AI hallucination is when a language model states something false with the same fluent, confident tone it uses for true statements. In financial research this shows up as invented figures, wrong fiscal years, misattributed quotes, or a plausible-sounding number that does not actually appear anywhere in the source filing. The output reads as correct, which is what makes it dangerous.
Why do AI models get financial numbers wrong?
A language model predicts plausible next words based on patterns in its training data, it does not look up a number in a database and repeat it exactly. When asked for a specific figure, like last quarter's EPS, it can generate a number that fits the shape of a correct answer without matching the actual filing. This risk goes up with older training data, multi-step arithmetic, and any figure not present verbatim in the text you gave it.
How do I catch hallucinated numbers before publishing?
Check every number in an AI-drafted note against the primary source, the BSE or NSE filing, the exchange announcement, or the audited financials, not against the AI's own restatement of them. Pay closest attention to figures the model was not given verbatim in your prompt, since those are the ones most likely to be generated rather than retrieved.
Does giving AI the source document reduce hallucination risk?
Yes, meaningfully. A model asked to summarise a filing you pasted in is far more reliable than one asked to recall a figure from memory. Grounding the request in the actual document (sometimes called retrieval-augmented generation) cuts hallucination risk a lot, but does not eliminate it. Still verify the output against the source rather than trusting the grounding alone.