Ask ten lawyers how they use AI and you get ten different answers. Some use it every day and swear by it. Some tried it, got back fluent nonsense that took longer to fix than to write, and walked away. Some say their firm banned it. Some are using it constantly and telling no one. And almost everyone, quietly, suspects they are the one who is behind.
I want to take that worry off the table, because it is built on a misunderstanding of what the risk actually is.
The conversation right now is stuck between two camps. One is scared, usually privately. The other performs the opposite, treats fluency with the tools as a personality, and looks down on anyone who hesitates. Neither camp is describing the real problem. So let me say the thing both of them are circling.
The lawyer's job was always friction.
What a client pays a lawyer for is not speed and not output. It is the pushback. The inconvenient question. The doubt voiced before the deal closes. The "no" that nobody else in the room was willing to say. To be precise about it: the lawyer's value is disciplined friction applied where being wrong is expensive and hard to undo. That friction is the product. It always was.
And modern AI systems are optimized for the opposite. They are built for fluency, speed, and helpfulness - to give you something that sounds right, and to give it to you confidently. The agreeableness is not a glitch you can prompt away; it falls out of what the tool is optimized to do. In legal work, that fluency quietly erases the resistance that judgment requires.
In AI work this failure mode has a name: sycophancy. The model shifts toward your premise instead of holding independent judgment. In practical terms, it does not push back, ask the hard questions, or test your assumptions. It helps you get more persuasive before it helps you decide whether you are right. For a lawyer, that is the problem: the duty of competence assumes someone is testing the premise; instead, the tool is built to accept it.
Which means the competence risk is not that AI is sometimes wrong. Wrong is the easy case. It is the error everyone is already hunting - opposing counsel, the judge, the cite-check - and we are learning, one sanction at a time, that even the watched error slips through. The quieter, more serious risk is the one nothing is pointed at: AI is agreeable always. A lawyer who leans on a fluent model to move faster, without noticing, strips out the friction and signs their name to what is left. That is rubber-stamping, dressed up as efficiency.
Here is what I have learned working with these tools myself. You do not get real value out of a frontier model by accepting its first confident answer. You get it by refusing to. You interrogate the model the way you would cross-examine a witness whose certainty you do not trust. You push it the way a good professor pushes a first-year student who memorized the rule but cannot defend it. You make it show its work, you attack the weak link, you ask the question it routed around. The good answer is on the other side of the argument, not in the first reply.
Notice what that requires. It is not a prompt-engineering trick. It is judgment, applied as friction, on purpose. And it happens to be the exact skill we were trained for and most other professions were not.
So the inversion that should change how you feel about all of this: the better these tools get, the more valuable your "no" becomes, not less. A more fluent, more confident machine does not reduce the need for the person willing to interrupt it. It raises the premium on that person. What the scared camp fears is the very thing that makes a real lawyer more necessary - if they understand what the job actually is.
That is also why I am not worried about keeping up with the newest model. The model will change next month and the month after. What does not change is the discipline. Before you let any AI output touch advice, a filing, a negotiation, or a client communication, you should be able to answer five questions:
What task is the tool performing?
What client information, if any, enters it?
What does the provider do with that information?
What human judgment challenges the output before anyone relies on it?
And what record shows you took reasonable care?
If you cannot answer those, you do not have an AI workflow. You have a confidentiality gamble wrapped in convenience.
None of that requires you to understand model weights or transformer architecture. It requires you to remember that the tool is built to sound right and to be the one in the room who refuses to take that on faith.
The job was always friction. The tools are built to smooth it away. Competence, now, is the discipline of putting it back before the decision is final.
I am writing more about practicing law in the age of agentic AI - using these tools without surrendering judgment, privilege, or the duty of competence - at The Agentic Lawyer. www.theagenticlawyer.com
Educational only, not legal advice, and no attorney-client relationship is created. Views are my own, not my employer's. Attorney advertising in some jurisdictions.