The legal industry's relationship with artificial intelligence has settled into a comfortable groove. Law firms adopt AI for document review and legal research. Bar associations issue guidelines about "appropriate use." Vendors promise transparency and compliance. Everyone nods approvingly at the responsible integration of technology into practice.
This consensus is exactly the problem.
Recent surveys showing widespread AI adoption among law firms are framed as evidence of successful technology transition. The comfortable reading: lawyers are prudently modernizing, regulators are thoughtfully overseeing, vendors are behaving responsibly. But that framing mistakes acceptance for safety and adoption for understanding. The better question is not whether AI tools are popular in law firms. It is what liability structures break when everyone relies on tools nobody fully understands.
Consider the landscape that is quietly forming. An attorney uses an AI system to draft motions, review contracts, or predict case outcomes. That system operates through mechanisms even its creators struggle to fully explain. Something goes wrong. The work product fails. A client is harmed. Who is liable?
The comfortable answer is: the attorney. Professional responsibility already exists. Lawyers have duties of competence and care. Add AI to the equation, and those duties presumably expand to include understanding the tools you use. This feels logical and preserves traditional accountability structures.
But this answer assumes something that may not hold: that understanding is actually possible. If an AI system produces a result through statistical pattern matching across millions of parameters, can a lawyer meaningfully "understand" whether that system is reliable for a specific case? Can they validate its reasoning? Can they catch its failures in advance? Competence presupposes the possibility of competence. When tools operate as functional black boxes, that presupposition fractures.
The liability framework that works for traditional legal research breaks under this weight. A lawyer who relies on West or LexisNexis has access to the underlying data, the search logic, the limitations of what a keyword search will find. Those systems are comprehensible. A lawyer who relies on an AI system trained on terabytes of text and tuned through processes humans cannot easily trace is in a different epistemic situation entirely. The professional responsibility framework has not caught up to this difference.
What breaks next? Several things, probably in sequence.
First, the illusion that lawyers can discharge their competence duty simply by disclaiming AI use or documenting that they "verified" outputs. If verification is effectively impossible, this is theater. Regulators and courts will eventually see it as such.
Second, the assumption that liability flows only downward, from lawyers to clients. If AI vendors are training systems on attorney work product, charging law firms for access to those systems, and profiting from the patterns they extract, why should liability not flow upward as well? Vendors have information about their systems' limitations that lawyers cannot access. They have chosen the training data, the fine-tuning process, the deployment parameters.
Third, the vendor indemnity clauses that currently protect AI companies from the consequences of their products' failures. These will face pressure, probably through litigation. A widow suing over a cargo plane crash, a state suing over data collection from minors, a tribunal resisting regulatory capture: these cases hint at an emerging willingness to pierce the corporate veil protecting tech companies from consequences.
The comfortable consensus that AI is just another tool, governed by existing professional responsibility frameworks, assumes that tools are legible and that responsibility flows cleanly downward. Neither assumption survives scrutiny.
The question worth asking is not how to fit AI into existing legal structures. It is what new liability architecture becomes necessary when tools are neither fully transparent nor fully controllable.