If AI has touched any part of your casework, someone on the other side is preparing to ask you about it under oath. Get the 12 questions the way counsel will actually phrase them, what a strong answer sounds like next to a weak one, and a scoring rubric to find your gaps before someone else does.
By mid-2026, a public tracker had logged roughly 1,500 court decisions worldwide involving AI-fabricated citations. Courts have excluded expert declarations over AI use, and a federal court has ordered an expert to produce his AI prompts as discoverable methodology. Counsel does not need to understand AI to take a report apart. They need one question you cannot answer with a document.
This is the opening question because it sets the trap for everything after it. Many courts now have standing orders on AI use in filings, and ABA Formal Opinion 512 applies candor duties to generative AI use, so a false or hedged “no” is checkable. If use surfaces after you denied it, counsel will spend the rest of the cross on your credibility instead of the technology.
Most examiners prepare for the chatbot question and get blindsided by this one. The AI embedded in your forensic suite counts even if you never opened a chat window. And because SWGDE has published no best-practice standard on AI-assisted digital forensic examination, there is no standard to point at. There is only you, explaining what you relied on and what you confirmed yourself.
Enter your email to unlock the full test on this page and download the branded PDF.
Counsel asks this because the answer separates a managed tool from a personal habit. Account type determines who administered the settings, what the terms of service said, and whether anyone in your organization could even see the activity. “My personal account” opens an entire line of questioning about where case-adjacent material went and who agreed to what terms on the organization's behalf.
Typing case material into a chat window can amount to disclosing it to a third party, depending on the terms and configuration, and ABA Formal Opinion 512 puts confidentiality squarely on the professional using the tool. This question is where privilege problems start, with protective-order and victim-privacy problems right behind them. Counsel is hoping you never thought about the boundary at all.
In the 2025 Clio Legal Trends Report, 53% of legal professionals said their firm has no AI policy or they were unaware of one. Counsel knows those numbers and is betting you are in the majority. An approval chain shows an organization exercising judgment. Its absence lets counsel paint every AI touchpoint in your workflow as freelancing.
In Matter of Weber, a New York court rejected an expert's valuation after he could not recall the prompts he had given a chatbot. And a federal court has ordered an expert's prompts produced as discoverable methodology, though that order is stayed pending review. The safe assumption is that prompts are bench notes: part of the record of how you worked. If they are gone, counsel will argue your methodology is gone with them.
This question tests whether you understood the agreement you were working under. Data handling differs sharply by vendor and by account tier, and the terms change over time. If you cannot answer, counsel will invite the jury to imagine your case material sitting in someone's training data, and you will have no document to answer with. Guessing confidently is worse than not knowing, because the terms are checkable.
Fabricated output built the court record on AI: the roughly 1,500 tracked decisions worldwide all involve AI-fabricated citations, and Kohls v. Ellison shows how it ends: exclusion, with no second chance. The examiners who get through this question are the ones who treated every AI output as unverified until checked against its original source. The ones who do not are the reason the tracker keeps growing.
In the Watson Grinding matter, opposing counsel obtained the expert's ChatGPT conversations and showed that most of the filed report tracked the chatbot's output. Compare Ferlito v. Harbor Freight, where testimony was admitted: the expert wrote his report first and used AI only to check it afterward. The order of operations is the whole question. Courts have accepted the second pattern. The first one hands counsel your signature next to a machine's sentences.
Most examiners prepare for the chatbot question and get blindsided by this one. In State v. Puloka, AI-enhanced video was excluded because the algorithm could not be explained or validated. The AI embedded in your forensic suite counts even if you never opened a chat window. And because SWGDE has published no best-practice standard on AI-assisted digital forensic examination, there is no standard to point at. There is only you, explaining what you relied on and what you confirmed yourself.
In the 2025 Thomson Reuters survey, 52% of professionals said their organization has no generative AI policy at all. Counsel asks for the date because a policy written after the examination is nearly as damaging as no policy: it concedes the work was done in an unmanaged environment. The document itself matters less than what it proves about when your organization started paying attention.
The same Thomson Reuters survey found 64% of professionals received no AI training. Counsel uses this question to convert every earlier answer into a competence problem: you used a tool you were never trained on, under a policy you may not have read, and you want the court to rely on the result. Documented training closes that door.
Models change underneath their users, and vendors retire versions on their own schedules. Counsel asks this to suggest your work is unreproducible: whatever you validated no longer exists. You cannot freeze time, but you can keep a record that stands on its own, backed by a policy that treats every model change as a trigger to revalidate before the tool touches casework again.
Count how many of the 12 you could answer today, on the stand, with a document to back the answer. The document is the standard. An answer you cannot support in writing counts as a miss.
10 to 12: Ready. You have the makings of a disclosure-ready AI record. Ready means prepared to answer, not guaranteed to prevail; no honest document promises a court outcome. Your remaining work is keeping the record current as tools and policies change.
6 to 9: Exposed. You have discipline in places and gaps in others, and you do not get to pick which question counsel opens with. The gaps are usually the quiet ones: prompt retention, vendor-embedded AI, the policy date.
0 to 5: At risk. If AI is anywhere in your workflow right now, including inside your forensic tools, the record being built about your work is being built by accident. Fixing this before your next report costs far less than explaining it on cross.