The AI Blind Spot: What Does Generative AI Miss in High-Stakes Disputes?

Generative AI (GenAI) is reshaping the legal industry. As we’ve previously discussed in the New York Law Journal and our Freshfields blog, GenAI tools can optimize early case assessment; document review and analysis; preparation for witness interviews, depositions, and trials; and everyday workflows. Make no mistake: GenAI is a disruptor and a force multiplier for litigations and investigations. 

GenAI’s upsides, however, come with certain pitfalls. GenAI tools may miss what they are not equipped to recognize, creating significant (but manageable) risks for litigators and investigators deploying them in high-stakes matters. Here, we address three critical issues legal practitioners must keep in mind when deploying GenAI, and explain the need for a “human in the loop” to ensure that use remains responsible and defensible.

AI is Only as Good as What It Sees (and Fully Understands)

Representations made to courts, regulators, or adversaries carry risk if they prove to be incorrect, misleading, or incomplete, whether or not those errors are caused by AI. Trained litigators and investigators always try to be mindful of the limits of the evidence available to them and the reliability of particular evidentiary sources. GenAI may not yet be entirely there. 

• It may not be attuned to information that has been deleted or archived, communications that have been taken to unmonitored platforms, the fallibility of witness memories, or context that may not appear on the printed page. GenAI may weight sources incorrectly—perhaps discounting or ignoring a proverbial “smoking-gun” email because its author later said he or she didn’t mean it.
• GenAI models may still struggle to identify coded references, sarcasm, humor, or slang.  In a recent indictment, prosecutors relied upon the use of “surgery” and “flights” as coded references to material non-public information (the dates of potential transactions ahead of which the defendants allegedly hoped to trade).[1]

GenAI may miss these evidentiary gaps or fail to grasp nuance generally, whereas a human investigator would more likely say “I do not know what these terms mean, and I want to know more.” Having a trained “human in the loop” is crucial to correct for the risk that a GenAI tool will miss those indicators. 

The bottom line: “AI didn’t catch it” will be received about as well as a carpenter blaming his or her tools, meaning litigators and investigators should continue to incorporate traditional investigative methods into AI-assisted document review. Put more simply, GenAI is a powerful tool for litigators and investigators—but don’t throw out the rest of your toolkit. 

Anchoring Bias, Model Drift, and the Limits of Synthesis

Complex litigations and investigations often require breaking down massive, ambiguous, conflicting, and ever-expanding data sets into nuanced, qualified, and iterative conclusions. A significant cognitive risk in using GenAI is “anchoring bias”—perhaps to be restyled as “algorithmic” bias—where a GenAI tool may give too much weight to early patterns or initial conclusions and fail to revisit them as new or contradictory evidence emerges.

A January 2026 study investigating how GenAI tools could sway a prosecutor to bring (or decline to bring) criminal charges highlights this exact problem.[2] Offered a range of prosecution scenarios, GenAI consistently recommended filing criminal charges, even after the researchers fed in evidence or legal flaws threatening to tank the case. Only when researchers offered “the rare combination of a severe flaw presented in a rich, defense-oriented narrative” did the model temper its tendency to recommend prosecution. In the real world—mindful of the gating issues presented by juries and judges—a prosecutor would (one hopes) avoid charging non-meritorious cases. Trained litigators and investigators try to recognize and correct for these biases—the “but what if I’m wrong?” test. It remains unclear if current GenAI models can reliably reassess evidence with a fresh perspective. 

The Human Element: Interviews, Depositions, Trial Examinations, and Investigative Reports

Not every event in an investigation can be reduced to a digital record you can feed into your GenAI model. At some point, a litigator or an investigator will have to talk to a real human person. In interviews, depositions, and trial examinations, there is no substitute for informed human engagement.

Witnesses are perceptive. A guarded witness might be mentally assessing the questioner, evaluating whether the interviewer knows the ins-and-outs of the case. If the witness senses the questioner does not know the record, the witness can tailor his or her answers to be minimally revealing and minimally helpful. GenAI may help litigators and investigators prepare by isolating relevant documents and generating first drafts of outlines, but it may not be able to replicate the insight gained from a cross-examiner’s command of the record—insight that only comes from preparation.

Similarly, AI’s ability to generate a first draft of an interview memorandum or investigative report is not risk-free. GenAI models may overstate factual findings, even where the evidence supports only a qualified conclusion; or it may hallucinate them entirely.[3] In complex litigations and investigations, conclusions often come with context and caveats. GenAI can be driven to produce a definitive answer—yes or no—even when the better answer might be “it depends.”

Conclusion

GenAI is a powerful ally in litigation and investigations, but legal professionals should be clear-eyed about its limitations. Humans can be mindful of their cognitive limitations and try to correct for them; AI might not. As companies and firms adopt GenAI tools, success will depend on redesigned workflows that weave together GenAI with experienced human judgment. This “human-in-the-loop” architecture is essential to maintaining confidentiality, delivering accurate results, and maintaining the standards of professional excellence that courts, regulators, and clients demand.

Freshfields will continue to monitor developments in the GenAI space. Click here to learn more about how we are at the forefront of incorporating AI into our litigation and investigations workflows.

To receive the latest insights on US legal developments, subscribe to the Freshfields A Fresh Take Blog.