Study finds AI summaries can quietly flip investment buy/sell calls
A peer-track paper shows compressing filings can reverse downstream recommendations, making the summarization step a measurable risk in agentic investing.
A study accepted to the EMNLP 2026 Industry Track finds that when a large language model compresses a financial filing into a summary, the summary can read as fluent and factually accurate yet still change the buy/hold/sell recommendation the model would give from the original text.
The paper by Lee et al. was submitted on June 28 and revised September 17. According to Klement on Investing, the team includes people from JP Morgan, BlackRock and State Street; the experiment had ChatGPT, Gemini, Qwen and DeepSeek summarize filings for the 100 largest US stocks, then had Gemini issue recommendations. Per that column, re-running on the full text changed about one in ten recommendations, with buy/sell direction flipping in a quarter to a third of cases — these specific figures come from the column's reading, not the loaded abstract page.
The authors propose Agentic Context Compression: generate multiple candidate summaries and audit their disagreements against the source. Compression, they argue, should be judged by whether it preserves decision-relevant context, not just efficiency and factual accuracy.
Sources:https://klementoninvesting.substack.com/p/flippin-heckhttps://arxiv.org/abs/2606.29251