Invenio Partner Warns Automation Bias Is the Real AI Risk in Funding Underwriting

An Invenio LLP partner has published a detailed argument that the principal danger of artificial intelligence in litigation finance underwriting is not fabricated citations but the quiet erosion of the human judgment that underwriting depends on.
According to Real Talk About AI in Litigation Finance Underwriting, written by Brenna Legaard, large language models perform reliably on well-defined, data-rich tasks such as analyzing prior art and preparing claim charts, and they work without fatigue or anchoring bias. What they cannot do is predict case outcomes, because the training data does not contain them. Models learn from published opinions, while the vast majority of disputes end in confidential settlements that are never mapped. Legaard writes that models “have known knowns, perhaps known unknowns, and no unknown unknowns whatsoever.”
The piece cites a 2024 study finding hallucination rates between 58% and 88% on factual legal questions, with the weakest performance on less prominent cases, and notes that model accuracy degrades as input length grows. Its sharper concern is automation bias: decision-makers deferring to polished output under time pressure, so that “the model’s confident framing then becomes an unwary underwriter’s confident framing.”
Legaard draws a parallel to McKinsey research on insurance underwriting, where firms that mandated black-box models over human judgment found that staff lost faith in the models and underwriting skills atrophied. The recommended response is cultural rather than technical: open discussion of where AI use introduces confirmation bias, and hiring underwriters who interrogate outputs rather than merely producing them faster.
