Most AI tools today operate like answer vending machines. Insert a prompt and out comes a polished response. But legal training has long been built around almost the opposite interaction.

The Socratic method does not hand students the answer. It questions, challenges, changes a fact, and asks whether the conclusion still holds. The frustration is part of the design. Students learn the rule, its limits and, eventually, which questions matter in the first place.

What would it look like for AI to borrow from that tradition?

That question is becoming urgent. AI increasingly lets us get an answer without doing the thinking. What happens when AI-native professionals never learn to do the thinking in the first place?

The temptation is obvious. It is in our nature to make thinking easier. We write things down, use calculators, and Google facts the moment we cannot recall them.

This is not laziness. It is human nature.

A landmark study of the “Google effect” found that when people expected information to remain available on a computer, they remembered less of the information and more about where to find it.

That can be a good trade. Lawyers need not spend limited attention memorizing document locations or reconstructing project status from six email chains.

But generative AI goes further than Google. It does not just find information. It appears to do the reasoning for us. It gives us the conclusion before we have framed the question, developed an initial theory, or learned enough to recognize what is missing.

Microsoft researchers, including Advait Sarkar, surveyed 319 knowledge workers about their use of generative AI, and the results are striking. Use of AI resulted in lower levels of critical thinking, memory, creativity, and metacognition. The evidence is still emerging, but the warning signs are accumulating, and we should be sounding the alarm.

Yet there may be a different path: one where AI does not inevitably make us less capable.

In this moment of monumental change and building, design matters most.

Law schools are already wrestling with the tradeoff. The University of Chicago is creating AI-free spaces in required first-year classes and exams while introducing AI into legal writing and advanced courses. Stanford students analyze legal problems independently, then critique AI-generated answers. These schools are separating the work that establishes a cognitive foundation from the work where AI can extend it.

This issue often manifests as a conversation about training junior talent in the age of AI. That is at the heart of the problem, but not the entire problem. Judgment is not acquired once. Lawyers keep developing it through unfamiliar facts, changing law, new clients, and the consequences of prior decisions.

And AI can interrupt those feedback loops at every level of seniority.

Another policy reminding lawyers to “think critically” and “check the AI’s work” will not be enough. If an instant answer is on the screen, people will use it. A lawyer who clicks approve is technically in the loop. That does not mean meaningful judgment occurred.

As builders, we have to confront how people actually behave and design around that reality.

That does not mean we know exactly what the right systems look like. But one possible framework begins by distinguishing between work that merely consumes attention and work that develops judgment.

For routine work, the easiest path probably should be easy. AI can find documents, track status, organize information, route requests, and remember what happened. None becomes more valuable because a lawyer struggled through it manually.

Judgment-intensive work may benefit from something different.

Perhaps an AI tool could ask a lawyer to state an initial view before displaying its own. Maybe it could expose a hidden assumption, identify a missing fact, offer a counterargument, or ask what would change the analysis. It might help a lawyer engage more intentionally with the cases, documents and decisions that make up the actual craft.

Sarkar calls interventions like these “provocations.” Instead of completing our thoughts, AI can challenge them. His research at Microsoft Research in Cambridge suggests that tools can preserve direct engagement with the material, introduce productive resistance and prompt people to reflect on their own reasoning.

These may not be right for every lawyer, task, or moment. A system that turns every routine task into a law school seminar would not make anyone better.

But we should at least be asking different design questions.

Are we helping people understand the work, or only finish it? Are we strengthening their ability to evaluate an answer, or quietly weakening it? Are we preserving the mental repetitions through which expertise develops?

The risk is that we turn lawyers into professional validators of a machine’s opinions. The opportunity is to explore tools that flex the cognitive muscles lawyers need to challenge those opinions, improve the work, and sharpen their critical-thinking skills along the way.

Efficiency matters. Better thinking should too.

The most valuable AI may not always give us the fastest answer. Sometimes, it may help us ask a better question. Which leaves us with a choice.

Do we want a tool that thinks for us, or one that makes us think?