When AI Meets Ethical Walls:
The Permission Problem Law Firms Can't Ignore
When a document leaves the system that governs who may read it and enters an AI tool, what happens to its permissions? In many setups, the uncomfortable answer is that they stay behind.
The Scenario
Nobody hacked anything. The tool simply did its job.
Picture an associate who has been formally screened from a sensitive acquisition. She opens the firm's new AI assistant and, in good faith, asks it to summarize the latest thinking on a related deal. In seconds it hands her a clean summary, drawn in part from the very matter she was walled off from.
Nobody hacked anything. No policy was defied. The tool simply did its job.
That is the question every firm should be able to answer before it lets artificial intelligence near a matter, and most cannot: when a document leaves the system that governs who may read it and enters an AI tool, what happens to its permissions? In many setups the uncomfortable answer is that they stay behind. The file moves. The access controls do not.
The Permission Problem
The walls are real. They are also invisible to most AI tools.
Firms have spent years building careful walls inside their document systems. This team can see this matter. That group is screened from it. A confidential deal is visible only to the few who need it. Those rules are the backbone of professional obligation, and they are invisible to most AI tools. When a protected document is pasted into a prompt or absorbed into a search index that feeds an AI, the wall that protected it does not travel with the words. What was carefully restricted becomes quietly available.
What happens to a walled document
Switch between the two approaches
Relevance Is Not Permission
A blind assistant is a dangerous one
A system that has ingested everything and forgotten who was allowed to see what will, when asked a helpful question, answer helpfully. It will surface a passage from a matter the person was screened from. It will summarize a document they were never entitled to open. The tool is not malfunctioning; it is retrieving relevant information, blind to the fact that relevance and permission are not the same thing. In a profession built on ethical walls and privilege, a blind assistant is a dangerous one.
Convenience Does Not Suspend Obligation
Regulators and courts have begun to notice
Regulators and courts have begun to notice, and the theme across recent guidance is consistent: convenience does not suspend obligation.
Not a Productivity Gain
If a tool cannot respect the access rules that governed a document in its original home, then feeding that document to the tool is not a productivity gain. It is an exposure, and one that is difficult to detect until it surfaces in exactly the wrong place.
Banning AI Is Not the Answer
Understandable, and over time untenable
The instinct of some firms is to ban the technology outright. That is understandable and, over time, untenable.
Reason 01
The value is too large to refuse
The value of AI that can reason over a firm's own knowledge is too large to refuse.
Reason 02
People will use it regardless
The people inside the firm will find ways to use it whether or not leadership approves.
Reason 03
The real task is inheritance
Not keeping AI away from the firm's knowledge, but making sure it inherits the rules that governed that knowledge all along.
The real task is not to keep AI away from the firm's knowledge. It is to make sure that when AI reaches that knowledge, it inherits the same rules that governed it all along.
The Fix
Make permission travel with the information
The fix is a change of principle. Permission should be a property of the information itself, enforced at the moment of retrieval, so a person only ever sees through the assistant what they could have seen without it. In practice that means AI that runs inside the firm's own environment rather than a public cloud, and that mirrors the firm's existing roles, security walls and ethical walls at the point of the query.
The One Test
The one test every firm should run
Reframe the AI conversation around a single test. Before any tool touches client matters, ask what a screened user would see if they asked the system a pointed question about a matter they are walled off from.
Read the Result Honestly
If the honest answer is that the tool would tell them, this is not a future risk to manage; it is a present exposure to close.
The firms that will use AI with confidence are not the ones that move fastest. They are the ones that made sure their most basic professional promise - that some information stays behind a wall - survives contact with the machine.
FAQ
