The Hard Work is the Point

Humans are lazy by nature. Given the opportunity, we almost always choose to conserve effort and energy.

"Work smarter, not harder," they would say.

So we looked for “smarter” ways to work. In the past, things almost always turned out fine. The search itself was the hard work — we exchange of one form of labor for another. We abstract one type of strain into mental rumination.

Displacing the final rung

Generative AI eliminates this labor exhange. It can certainly reduce the tedious part of the work, but it can also eliminate the thinking about the work — the rumination that “working smarter” entailed.

Every industrial revolution before climbed a rung of the labor ladder and automated it, but left human cognition as the one rung the machines could not displace. Cognition was human ingenuity's last bastion.

In the past, “working smarter” meant exchanging one form of work that is low on the labor ladder with applying human cognition plus another form of work higher on the labor ladder. Now, generative AI can replace human cognition, and give people the power to appear productive without putting in any work at all. And, for the most part, the output of generative AI is "good enough".

Software engineering is living this displacement out in real time. Engineers vibe code solutions, and these solutions seem to work fine and hit the intended outcome… until someone has to maintain it. Then a tiny tweak breaks everything, the house of cards collapses. Eventually, a senior engineer has to step in and refactor just to get back to status quo. Engineers are lucky, in a sense. The slop fails fast. The failure is visible. It fails before the skills needed to fix it have had time to atrophy.

Law does not get the same advanced warning. A generative model is built to predict the most probable next word. The most probable words are not the most correct ones. If there is an error, a document doesn't crash anything. The error just sits in the document. Errors may be discovered years later, if at all — in a dispute, an audit, a deal gone wrong, long after the person who signed off has moved on.

The erosion of judgment

We know the risks of Gen AI, and yet, as lawyers, we are living in a contradiction. The economic machine pushing us toward adopting Gen AI for efficiency is dismantling the training ground junior lawyers used to learn on. The “grunt work” was never just grunt work — it was how a first-year built the instinct to know when something was wrong before they could say why. Hand that work to the model, and you don't just save time, you skip the years where the judgment is built.

How does a profession built entirely on judgment survive a future where Gen AI is expected to replace the “grunt work” in every workflow?

One answer is to reject AI — offload nothing to the machine, keep the human brain as the entirety of the “loop”. The human lawyer hones their craft the slow way. In the short term, this person will likely lose to their AI-augmented colleagues on speed and cost. There are real consequences to rejecting AI, which may protect the mind, but does not protect the practice.

Another answer comes from those who are advocating for the adoption of AI. They argue that they are not skipping the hard work, but relocating it. Rather than grinding through every problem the old-fashioned way, they front-load less and back-load more. They focus their effort on verification instead of construction. They argue that supervision is not an abdication of duty, it is “working smarter”.

The slippery slop(e)

Except that supervision requires discipline, and humans are lazy.

Human in the loop, human on the loop, human in command, or whatever other term is next invented to describe the continuing involvement of human supervision in generative AI work loops presuppose that humans are able to catch the AI errors.

If humans can see something fail, they can recalibrate the machine or adjust their trust in the machine. That is working for software engineers, where a bug crashes loudly. It is exactly what lawyers lack. It may take years to discover a bug, and so the lawyer grows more comfortable as the machine continues to work.

Over time, the lawyer gives the machine more trust, one unremarkable task at a time, because they couldn’t see anything obviously wrong. The lawyer doesn't get worse at supervising on purpose. They increase their usage of generative AI. Frequency breeds complacency. Lawyers as supervisors of generative AI may wish to remain dutiful, but the alarm never rings.

Will this state of reliance last long enough to cause atrophy of individual legal minds? Will this state last even longer to cause an atrophy of useful data across the system, which leads to complete model collapse?

So, what if the machines win?

We want to be good lawyers. We want to deliver good outcomes to our clients.

If AI becomes so good that it displaces human judgment completely, so that humans are no longer needed, then this whole discussion is moot. It is better for humans to step out of the way in order to have AI deliver the best outcomes to clients… But, if AI is even slightly short of completely displacing humans, then we must preserve human judgment and our ability to think. We must not let it atrophy, and the only way to preserve that is through hard work.

The hard work is the point. The thinking is the point.

Right now, it seems we are giving up too easily in this rush to adopt AI in everything, and relegating ourselves to the role of supervisors.