Conflict Checks, AI Tool Memory, and the Cross-Matter Leak Risk

Conflict Checks, AI Tool Memory, and the Cross-Matter Leak Risk

Law firms have been running conflict checks for decades. The process is well understood, if sometimes tedious: before taking on a new client or matter, you search your records to make sure representing this party won't create a conflict with existing or former clients. Model Rules 1.7, 1.9, and 1.10 lay out the framework. Firms invest heavily in conflict databases, intake procedures, and lateral hire screening. It works reasonably well for the traditional problem.

But there is a new vector that most conflict systems were never designed to catch, and it is sitting inside the AI tools your attorneys are already using.

The Memory Problem

Many AI tools, particularly large language model platforms, maintain some form of persistent memory or context across sessions. OpenAI's ChatGPT introduced explicit memory features in early 2024. Microsoft Copilot retains context within organizational tenants. Even tools without formal "memory" features can retain information through conversation histories, fine-tuning on user data, or retrieval-augmented generation (RAG) systems that pull from stored documents.

Here is where it gets interesting for conflicts purposes. When an attorney uses an AI tool on Matter A for Client X, and then uses the same tool (or the same organizational instance) on Matter B for Client Y, information from Matter A can bleed into the work product for Matter B. This is not hypothetical. It is a natural consequence of how these systems are architected. The tool does not know it is supposed to maintain an ethical wall between matters. It is optimizing for helpfulness, and "helpfulness" means using everything it knows.

This creates two distinct problems that map directly onto existing conflict rules.

Problem One: Duty of Confidentiality Violations Under Rule 1.6

If an AI tool trained on or retaining data from Client X's matters surfaces that information while assisting with Client Y's work, you have a confidentiality breach. Full stop. Rule 1.6(a) requires informed consent for any disclosure of information relating to the representation of a client. The attorney may not even realize the AI tool is drawing on privileged or confidential material from another engagement. The leak is silent.

Consider a practical scenario. A litigation associate uses an AI drafting tool to prepare a motion for Client X, uploading strategy memos and deposition transcripts. Two weeks later, a colleague in the same firm uses the same tool to draft discovery requests against a party whose interests are adverse to Client X. If the tool's memory or document store retains Client X's materials, it could generate discovery requests that are suspiciously well-targeted. The opposing party might never know why, but the ethical violation is real.

Problem Two: Imputed Conflicts Under Rule 1.10

Rule 1.10 imputes one lawyer's conflicts to the entire firm. The rationale has always been that information flows within a firm, so if one attorney has a conflict, the whole firm is treated as conflicted. Traditionally, this was about human communication: hallway conversations, shared files, internal memos.

AI tools with shared memory or shared data stores function as a new channel for imputation. If Attorney A's client information is accessible to Attorney B through a shared AI system, the practical effect is identical to Attorney A having told Attorney B about the matter directly. The 2023 Florida Bar Ethics Opinion 24-1, while focused on confidentiality broadly, flagged that AI tools processing client data must be treated with the same care as any other repository of confidential information. The New York City Bar's Formal Opinion 2024-1 similarly warned that generative AI use requires attorneys to understand how client data is stored and whether it could be exposed to unauthorized parties, including other clients' legal teams within the same firm.

Why Traditional Conflict Systems Miss This

Your firm's conflict check system searches for party names, related entities, and known relationships. It was built for a world where information moves through people and documents with identifiable labels. AI tool memory does not work that way. The information is embedded in model weights, conversation logs, vector databases, or cached prompts. It is not tagged with a client name or matter number in any way your conflict database can query.

This means you can run a perfectly clean conflict check, get the green light, staff the matter, and still have a functional conflict because your AI infrastructure is leaking information between engagements. The conflict check passed because it was looking in the wrong place.

What Regulators and Courts Are Signaling

We are still in the early innings of regulatory response, but the direction is clear. The ABA's Formal Opinion 512 (July 2024) addressed AI use and reinforced that attorneys must ensure competence and confidentiality when using AI tools, specifically noting the risk of data being used to train models or being accessible in future sessions. Several state bars, including California, Texas, and New Jersey, issued guidance in 2023 and 2024 requiring attorneys to understand the technical architecture of AI tools they use in practice.

On the litigation side, judges are paying attention. The Mata v. Avianca debacle in the Southern District of New York in 2023 was about hallucinated citations, not conflicts. But it put a spotlight on attorney responsibility for AI-generated work product. The next wave of sanctions and malpractice claims will likely involve confidentiality breaches and undisclosed conflicts arising from careless AI tool usage. The professional liability insurers are already asking about it; Beazley and CNA both updated their risk advisories in 2024 to specifically mention AI-related confidentiality exposure.

Practical Steps to Contain the Risk

If you are a general counsel, managing partner, or compliance officer at a firm using AI tools (and at this point, most firms are, whether officially sanctioned or not), here is what actually reduces exposure:

• Matter-level isolation. AI tools must be configured so that data from one matter cannot be accessed, referenced, or influence outputs on another matter. This is an architectural requirement, not a policy one. Telling attorneys to "be careful" is not a control.
• No shared memory across clients. If your AI platform has memory or learning features, those features must be scoped to individual matters or disabled entirely. A firm-wide AI assistant that "learns" from all matters is a conflict machine.
• Audit trails. You need to be able to demonstrate, after the fact, what data an AI tool accessed when generating work product for a specific matter. If you cannot produce that record, you cannot defend against a conflict or confidentiality claim.
• Integration with conflict systems. AI tool access controls should be linked to your conflict database. When a new matter is opened and conflicts are cleared, the AI environment for that matter should be provisioned with appropriate walls already in place.
• Vendor diligence on training data. If your AI vendor uses client data to improve its models, you have a problem that extends beyond your firm. Contractual prohibitions on training are necessary, but so is technical verification.

How FirmAdapt Addresses This

FirmAdapt's architecture enforces matter-level data isolation by default. Every AI interaction is scoped to a specific matter with its own data boundary, so information from one engagement cannot surface in another. There is no shared memory, no cross-matter learning, and no ambient context that accumulates across clients. Audit logs capture exactly what data was accessed for each AI-assisted output, giving you the documentation you need if a conflict question arises after the fact.

The platform was designed for environments where regulatory obligations like Rules 1.6, 1.7, 1.9, and 1.10 are not optional considerations but hard constraints on system architecture. FirmAdapt treats ethical walls as infrastructure, not policy overlays. For firms that need to use AI without creating a new category of conflict exposure, that distinction matters in practice.