AI-generated summaries of financial documents can look accurate while silently steering analysts toward the wrong call.
Researchers studying large language model compression of financial filings and earnings-call transcripts found a specific failure mode they call information fidelity loss. The model produces a summary that reads as factually plausible — no hallucinated numbers, no invented quotes — but strips out the caveats and qualifiers that would change an investment judgment. They identify two patterns behind this: decontextualization, where a key data point survives the summary but its limiting context does not, and model dependency, where different LLMs compress the same document into meaningfully different pictures of the company. In agentic pipelines, where one compressed output feeds the next step, these distortions can compound.
The stakes here are higher than a chatbot giving a bad restaurant recommendation. Financial analysts already use LLMs to process more filings than any human team could read directly, and a fluent-but-misleading summary is harder to catch than an obvious hallucination. The researchers propose a method called Agentic Context Compression — generating several candidate summaries and auditing their disagreements against the original source — as a partial fix.
The broader implication is that "factually accurate" is not the same as "decision-safe" when it comes to compression, a distinction the industry has mostly ignored while racing to deploy AI into high-stakes workflows.