In a merger investigation, the next major witness might not be able to breathe. It may not even be aware that it is a witness. However, it will leave a trail, including drafts, summaries, suggested edits, and incomplete sentences that a chief strategy officer sent but never typed. Regulators are keeping a close eye on that area since no one is quite sure what to do with it yet.
You can see the same scene taking place in practically any large corporate legal department right now. Sorting through millions of documents pulled in response to a second request, paralegals hunched over screens, their coffee cold. Even before Copilot began writing half of the emails, the work was harsh. A more subdued query now looms over the space: who wrote this? Eighteen months ago, this question was nonexistent, but now regulators are beginning to ask it first.
| Subject | The role of generative AI artifacts in merger clearance reviews |
| Relevant year | 2026 |
| Primary regulators involved | U.S. DOJ Antitrust Division, FTC, European Commission, U.K. Competition and Markets Authority |
| Governing framework (U.S.) | Hart-Scott-Rodino Act, Second Request process |
| AI tools commonly in scope | Microsoft Copilot, Google Gemini, enterprise LLM assistants |
| Featured expert | Sean McDermott, digital forensics specialist and testifying expert |
| Publication referenced | Law360 |
| Acting AAG (Antitrust, DOJ) | Omeed Assefi |
| Core legal question | Distinguishing human-authored from AI-generated communications |
| Emerging discovery concept | AI engines treated as quasi-custodians |
Digital forensics expert Sean McDermott, who has testified in these cases, put it succinctly in a recent Law360 article. He contends that the question in merger clearance is no longer simply who stated it. Which model stated it, under what prompt, version, and audit trail are all important considerations. Experts in forensics believe that AI engines themselves are taking on an odd new role, akin to quasi-custodians with discoverable model states, version histories, and prompt logs. The majority of businesses haven’t kept up with this fundamental change.
Think about the actual factors that regulators consider when conducting a Hart-Scott-Rodino review. They are looking for standard business documents, such as unguarded information, casual Slack conversations, and open memos in which a competitor is referred to as “dominant.”
The language is important. A whole theory of harm can be shaped by it. However, what happens if the word “dominant” appears in a Copilot summary that was not thoroughly reviewed before being sent? An executive’s seeming admission might not have been an admission at all. Conversely, it’s also possible that AI covered up something that ought to have raised an alarm. The evidentiary record becomes unclear in either case.
The DOJ’s acting head of antitrust, Omeed Assefi, recently cautioned businesses not to use AI disruption as a handy justification for merger defenses. He said something that sounded almost tired while speaking at NYU: “We know when you’re trying to mislead us.” He pointed out that businesses are tempted to claim that since AI is revolutionizing their sectors, the deal should go through. “All right,” he said, “but bring proof.” The tone sounded like that of a regulator who had heard the AI narrative too many times.
It’s difficult to ignore how rapidly the ground has changed as you watch this happen. Metadata seemed to be at the forefront of e-discovery a year ago. Legal teams are now subtly incorporating provenance attestations and model-authentication language into collection protocols, much like they did with chain-of-custody clauses. It is necessary to rewrite the retention settings. Systems that delete prompt logs on a default schedule that no one bothered to check must be accounted for by legal holds.
The attorneys who deal with these cases believe that the upcoming years will be chaotic. There are no established standards. There isn’t much case law. Additionally, a language model has already written the email that will ultimately appear before a regulator, whether it is fair or not, somewhere in a server rack.