In Fortis Advisors, LLC v. Krafton, Inc., 354 A.3d 906 (Del. Ch. 2026), Vice Chancellor Lori W. Will relied on a CEO’s ChatGPT queries and implementation of the chatbot’s advice to establish motive and pretext in a breach of contract dispute. The case highlights the evidentiary significance of AI use, which can establish state of mind in real time and underscores the need for robust AI governance and preservation protocols. Krafton is of particular interest to businesses because many transaction documents and corporate organizational documents choose Delaware as a forum.
The Case: Krafton’s AI-Assisted Strategy Backfires
Krafton, Inc. acquired Unknown Worlds Entertainment, a video game studio. The founders, Charlie Cleveland and Max McGuire (along with Ted Gill, the CEO who joined later), negotiated an Equity Purchase Agreement with a significant earnout tied to financial targets (collectively, the “Key Employees”). The Agreement also guaranteed the Key Employees’ ongoing operational control and provided they could only be fired for cause based on specific, narrow grounds. Concerned the significant earnout reflected a poor deal, Krafton’s CEO, CH Kim, sought legal advice on avoiding payment and was warned against doing so. Kim then turned to ChatGPT. The record does not indicate that counsel had authorized or directed him to do so, but his queries sought clarification and strategy on a decision involving legal risk. ChatGPT “advised” the earnout was “difficult to cancel,” and in response to his request for strategic advice, the chatbot outlined pressure tactics and talking points for negotiations with the Key Employees. At ChatGPT’s suggestion, Kim formed an internal task force, dubbed “Project X,” with a mandate to either negotiate a “deal” on the earnout or execute a “Take Over” of Unknown Worlds. Over the next month, Krafton followed most of ChatGPT’s recommendations.
Krafton also terminated the Key Employees claiming, “intentional acts of dishonesty” and termination “for cause.” The lawsuit challenged the termination, and a bench trial was held on the breach of contract claim concerning the propriety of the termination and the Key Employees’ loss of operational control over the company.
The evidentiary record included emails, Slack messages, and ChatGPT queries. As to the latter, the queries were quite obviously relevant, but also appeared to be obtaining legal guidance. Privilege was never raised because the queries had been circulated via Slack.
With the chat logs fair game, the court’s opinion quoted these ChatGPT responses and matched them against Krafton’s subsequent actions—and the impact was significant, with Krafton held liable and equitable relief granted. The evidence demonstrated the termination of Key Employees lacked a legitimate basis and the stated reasons were pretext for their termination. The court also found Krafton’s seizure of operational control breached the Agreement. As to remedies, the court reinstated Gill as CEO, restored his operational control, and equitably extended the earnout period by 258 days. Cleveland and McGuire were not reinstated to their original roles (indeed, the court found they had abdicated their game development duties before termination, though not dishonestly). A second trial phase will determine monetary damages.
Privilege and Work Product Protections
The CEO’s “conversations” were of a privileged nature had he been talking to a lawyer—but he was not. The chatbot’s responses might, in principle, have been argued to be work product prepared in anticipation of litigation. In Krafton, however, privilege was never asserted over the ChatGPT content because Kim had copied and pasted the ChatGPT outputs into Slack messages shared with colleagues, which Krafton produced in discovery. Having voluntarily disclosed the substance of the AI conversations to non-lawyers internally, any privilege argument would have been waived. Krafton’s position was simply that “all relevant information from ChatGPT is included in the documents that Krafton already produced.” Fortis moved to compel the original search histories (i.e., Kim’s actual queries to ChatGPT), but Krafton confirmed they “no longer exist”—Kim had deleted them, and they were created before litigation was reasonably anticipated. The motion was ultimately resolved without a published ruling on whether the original queries would have been discoverable had they been preserved.
Notably, evidence also emerged that the Key Employees themselves may have deleted ChatGPT data. On the day they were terminated, in a group chat used to discuss their disputes with Krafton, Cleveland suggested that McGuire and Gill “check your company ChatGPT account to make sure there’s nothing incriminating there.” Krafton argued this constituted spoliation of relevant data by the Key Employees, adding another dimension to the AI-preservation issues raised by the case.
This decision comes amidst recent federal district court rulings including:
- In U.S. v. Heppner (S.D.N.Y. Feb. 17, 2026), Judge Jed S. Rakoff held that a defendant’s use of consumer-tier Claude to outline defense strategies was protected by neither attorney-client privilege nor work-product doctrine, finding no reasonable expectation of confidentiality given the platform’s terms.
- In contrast, in Warner v. Gilbarco Inc. (E.D. Mich. Feb. 10, 2026), the court denied a motion to compel AI interactions and came down hard on the defendant who asked for them, characterizing the request as a “fishing expedition” seeking “internal analysis and mental impressions.” The court stated: “to the extent Defendants argue that Plaintiff waived the work-product protection by using ChatGPT, the work-product waiver has to be a waiver to an adversary or in a way likely to get in an adversary’s hand… And ChatGPT (and other generative AI programs) are tools, not persons, even if they may have administrators somewhere in the background… the Court agrees with Plaintiff that the pursuit of this information is a distraction from the merits of this case, and that Defendants’ theory, which is supported by no case law but only a Law360 article posing rhetorical questions, would nullify work-product protection in nearly every modern drafting environment, a result no court has endorsed.” (citations and internal quotations omitted).
- Most recently, in Morgan v. V2X, Inc. (D. Colo. Mar. 30, 2026), the court similarly extended work-product protection to a pro se litigant’s AI interactions—but imposed strict requirements for using AI with confidential discovery material, holding that Data Processing Agreements must prohibit vendor training on inputs, bar third-party disclosure, and guarantee deletion on demand. The court acknowledged these requirements would effectively “bar the parties from using most, if not all, mainstream low-to-no-cost AI” for confidential information.
- And just last week in Tate Group Automotive LLC v. Legacy Automotive Capital LLC (Tex. Bus. Ct. June 4, 2026), Judge Grant Dorfman ruled that certain ChatGPT conversations involving a third-party defendant were protected as work product under Texas Rules of Civil Procedure, rejecting the Heppner reasoning. The court held that work product protection was not waived merely by sharing information with the AI tool, because “the information wasn’t disclosed to an adversary or in a way likely to get in an adversary’s hands.” While finding AI-assisted document review analogous to using “Westlaw, LexisNexis, e-discovery platforms, or other litigation-support tools,” the court ordered disclosure of all discovery materials shared with ChatGPT and recommended the parties amend their protective order to address AI tool usage.
Practical Guidance for Counsel
Together, these decisions confirm that AI can cause real damage, and that privilege and work-product analysis is fact-specific and depends on factors like platform terms, whether counsel directed the use, and how outputs were shared. Notably, courts have distinguished between pro se litigants (who act as their own counsel and may claim work-product protection) and represented parties whose AI use was not at counsel’s direction. Attorneys should be cautious of overreach in discovery requests, particularly with pro se litigants.
AI Evidence as the New Paper Trail
Chatbot queries and responses are now a potentially discoverable source of evidence that can establish motive, context, and state of mind. The Krafton court relied on this evidence in finding pretext, and savvy trial counsel will no doubt use relevant records relating to AI queries and responses as proof for judge and jury, requiring counsel to stay up to speed on the type of materials that can be generated through AI use.
Updating Retention and Preservation Protocols
Businesses should update document retention policies and understand how AI tools do or do not save queries and output. In the event of a dispute, litigation hold notices should address AI tools. Preservation obligations should encompass AI prompts, queries, responses, and AI-generated drafts, whether from standalone applications or enterprise-embedded functions like Microsoft 365 Copilot, Google Workspace AI, Slack AI, and Notion AI. Hold memos should instruct custodians to disable auto-delete or temporary chat features within AI platforms, preserve custom instructions and system prompts, and disclose use of any personal or non-enterprise AI accounts touching the matter. Of course, proportionality and reasonableness of preservation should be top of mind as data creation is growing exponentially. In Krafton, the CEO admitted certain chats had been deleted, providing additional evidence of suspect motives—and how the court addresses that deletion in the pending damages phase may shape AI spoliation doctrine going forward.
Shadow IT poses a particular risk: senior executives running personal AI accounts on personal devices often operate with retention defaults the company has never seen. Consumer tiers may default to rolling 30-day deletion of temporary chats, meaning material could be gone before anyone thinks to look. Outside counsel should make an AI audit a first agenda item in matter kickoffs—identifying which tools are in use, by which custodians, on which devices, with which retention defaults—and conduct that conversation before or contemporaneous with sending the preservation memo.
The Benefits of AI Governance Policies
Effective AI governance, i.e., policies, procedures, training, and oversight for AI tool selection, deployment, and use, can reduce legal, operational, and reputational risk. As Krafton illustrates, even a CEO can deploy an AI-generated strategy that contradicts counsel’s advice and results in significant litigation exposure. AI governance programs should be integrated with existing compliance and training programs applicable from the C-Suite to any potential user of an AI tool to mitigate risk. Organizations are best served to deploy AI governance policies and trainings in alignment with other technology use policies; to the extent they overlap, consider consolidating or cross-referencing.
Non-Lawyers’ Use of AI for Business Decisions
The big red flag from Krafton: business decisionmakers should think twice before going to AI for advice on decisions that pose legal risk, particularly in a business with in-house legal counsel (making it difficult for the business to argue that they were representing themselves at the time). Here, the CEO’s chat logs involved two obvious legal risks: employment termination and breach of contract. AI can be a decent issue-spotter but it is not a responsible advisor. If you ask a chatbot “how do I get out of this payment obligation?” it might tell you not to try (as here), but it will still give you suggestions to do what you want. AI sycophancy is a well-documented phenomenon—your AI friend will always seek to be “helpful” and to validate you. Even asking a chatbot something more benign (“do we have to pay this earnout?” “can we terminate these employees?”) can be risky, because any subsequent deviation from that interpretation may appear to be knowing and willful. In-house lawyers might want to warn their leaders: “Imagine being deposed on your chat logs.”
But the ubiquity of AI assistants and agents, and the pressure to get things done quickly, may simply overpower the lawyers who simply say “don’t do it.” Is there a way to interpose safeguards when “don’t” won’t work? Perhaps the cases to date spark some creative mitigation ideas: Consult AI only with a lawyer’s oversight or assistance; use temporary chat (subject to litigation holds and preservation obligations); mark the chat privileged and confidential. (And of course, none of these matter if AI “advice” is circulated among nonlawyers internally.) Time—and more court decisions—may gradually reveal a more-practical path; in the meantime users can do their best to reduce risk but must assume that their chats will be subject to disclosure.
The introduction of AI compounds the risk that what previously would have been limited to privileged attorney interactions and thought processes on risk-based decisions will now be fully discoverable—and damaging. To guard against cautionary tales like these, consider training corporate decisionmakers and HR on the importance and applicability of evidentiary privileges, with special attention to the AI intersection.
