AI ETHICS
5.
Verification Expectations
AI outputs require human verification before use in client matters. This white paper repeatedly emphasizes this principle.
The verification challenge varies by output type. Factual assertions require checking against authoritative sources. For example, if a defender uses a transcription tool, the defender should review the original audio or video against the AI-generated transcript. Likewise, if AI states that a particular statute applies or that a case holds a particular proposition, the defender must confirm these assertions by reading the statute or the case. The hallucination problem documented in Part II means that even confident- sounding AI assertions may be entirely fabricated, propositions coming from actual cases may still be incorrect, and even legal-specific tools like CoCounsel and Lexis AI can hallucinate. Every citation in AI-assisted work product should be verified — and not by asking the AI whether its citations are real, which simply invites the AI to affirm its own errors. Legal analysis requires a different kind of verification. Defenders must evaluate whether AI-generated arguments are logically sound, whether they accurately characterize authorities, whether they address the strongest counterarguments, and whether they serve the client’s interests. This evaluation requires legal judgment that AI cannot provide about its own work. Of course, practical constraints impact verification expectations for all defenders. A solo practitioner with overwhelming caseloads cannot verify every sentence with the same rigor as an associate at a well- resourced firm. But this constraint argues for using AI judiciously rather than bypassing verification entirely. Defenders should focus AI assistance on tasks where verification is feasible and avoid using AI for tasks where verification would be impractical. When time permits only spot-checking rather than comprehensive verification, defenders should acknowledge this limitation to themselves and consider whether the risk is acceptable. The white paper uses “Verification Expectations” rather than “Verification Standards” intentionally. The defender’s verification obligation likely changes depending on the particular type of task or output, and reasonable defenders may disagree about what verification a particular task or output requires. What matters is that defenders approach verification thoughtfully rather than treating AI outputs as presumptively reliable.
38
Parity in Practice: The Defender’s Duty to Ethically Use AI
Made with FlippingBook Online document maker