While AI technology is fluent enough to look right and confident enough to sound right, it can still be catastrophically wrong in ways that are almost impossible to spot without going back to the source.
What Happens When AI Lies Confidently
An equity analyst reads a filing, builds a thesis, debates it with a PM, and a position gets sized. Every one of those decisions rests on a chain of underlying facts — a management comment, a trading multiple. Each fact, on its own, seems small. But the moment even one of those facts is fabricated, the entire chain is compromised, and nobody downstream knows it.
That is why hallucination in the investment research industry is not a minor inconvenience. AI inserting a confident, fluent, plausible-looking falsehood into a deliverable compromises the whole process of deploying capital and, in some cases, even your reputation.
The Deloitte Case
In July 2025, Deloitte Australia delivered a 200+ page report to the Australian government, priced at roughly A$440k. A researcher at the University of Sydney started reading the footnotes and found something strange — academic papers that did not exist were quoted in the report. Footnotes pointing to work supposedly published by universities that had never produced it.
Deloitte later confirmed that parts of the report had been drafted with generative AI, that several references and footnotes were fabricated, and agreed to refund the final instalment of the contract. Now picture the same failure mode inside an equity research report, a fairness opinion, a teaser, or a CIM going out to a hundred strategic buyers.
Why and How AI Hallucinates
A large language model does not know facts the way a trained analyst knows them. It predicts the next most likely token given what it has seen. When the model has good, relevant information in front of it, those predictions line up with reality. When it does not, it still predicts fluently — it just predicts something that sounds right instead of something that is factually right.
For investment research workflow, hallucinations manifest as: fabricated facts (a management quote that was never said), fabricated sources (citing research papers that do not exist), wrong attribution (a real fact from the wrong company or wrong year), and stale confidence (stating something true two years ago as current).
The Affirmation Trap
Research from Stanford and Anthropic has shown that leading models affirm users roughly 49% more often than humans do. When a user challenges a correct AI answer with a simple “Are you sure?”, the model often reverses itself and serves up something wrong. In research, analysts almost always ask questions with their existing view embedded — and AI does not stress-test the framing; instead, it flatters it.
What Actually Solves This: Citations
The single most important defence is inline citations tied to verifiable sources. Not citations as a footer, but after every claim. When an analyst sees an AI-generated statement, they should be able to click the claim, land on the exact page of the conference call transcript, and see it highlighted with their own eyes.
Citations do not make AI smarter. They make AI auditable. A properly implemented citation layer catches the vast majority of hallucinations that would otherwise reach a deliverable. That is why citations are approximately 80% of the answer to AI hallucinations in this field — not because they prevent errors, but because they make errors catchable in seconds instead of surviving to publication.
Where Analysts Are Still Relevant
A citation is only useful if someone actually reads the source and checks whether the claim matches. This is where the qualified analyst becomes more important in an AI workflow, not less. The CA, CFA, or MBA who reads a footnote disclosure will understand what it really says and identify the difference between what management said and what they implied. Analysts will become smart enough to use AI as tools to get insights faster while maintaining their judgment on what to keep and what to discard.
FootNote builds custom AI solutions for investment research workflows where every answer is citable to the original source. If you want to build or try out such a tool, feel free to reach out.