Flippin’ heck
Agentic workflows are increasingly used in investment applications.
A recent study by researchers from the US, UK, China, JP Morgan, BlackRock, and State Street found that the use of agentic AI workflows in investment applications may lead to a significant risk of AI hallucinations going unnoticed. To examine the reliability of these systems, they split a typical investment process into two steps: summarizing the Management Discussion and Analysis (MD&A) section from the quarterly reports of the 100 largest US stocks and then generating a buy/hold/sell recommendation based on the summary or the full MD&A discussion.
The team asked ChatGPT, Gemini, Qwen, and DeepSeek to summarize the MD&A sections and give buy/hold/sell recommendations. They also provided the complete MD&A discussions and earnings call transcripts to Gemini for the same task. The researchers expected Gemini to produce the same recommendations regardless of whether it used the full discussion or the concise summary. However, they discovered that the recommendation differed more often when using the summary compared to the full input.
When the model used the 20-point summary, the buy/hold/sell recommendation for a stock changed in up to 30% of cases, which is a considerable concern. The higher fluctuation in recommendations when using the full MD&A discussion only occurred in one in four instances. The researchers concluded that relying on an AI tool that changes its recommendation on such a random basis is unreliable.
Written by urgent.news from Klement on Investing's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
