When Each Handoff Makes the Next One Easier
OpenAI launched Astra for Law last week, and what struck me was not any particular feature but how much of the legal workflow tools like this are beginning to touch. They can help research an issue, develop an argument, draft the work, check it, and connect that work to many of the other systems lawyers already use. With each generation, a little more of the work that once belonged entirely to lawyers may be handed to the machine.
These are exciting times, and I think we should be open to using these tools to become more efficient and effective. I use GenAI myself, and I have spent much of the last few years encouraging judges and lawyers to learn how to use it responsibly. But the more capable these systems become, the more intentional we need to be about how we use them. We need to understand what we are handing over, what we are keeping for ourselves, and how the technology is changing the work around us.
As I read the Astra announcement, I kept coming back to a thought that has been bothering me for a while. People sometimes tell me the singularity is coming, or perhaps that it is already here. I always come back to the same question. Whose singularity?
At least in law, the answer is not as simple as saying the machine has surpassed the human. We talk about artificial intelligence surpassing human intelligence as though the machine arrived at that point on its own. But in law, what exactly is the machine standing on? Every statute it reads was written by a human being. Every brief reflects arguments developed by lawyers. Every opinion carries the judgment of judges applying those laws to real disputes. Layer by layer, generation after generation, human beings built the legal world the machine now appears so good at navigating.
When an AI system produces an extraordinary legal analysis, it is worth remembering where that ability came from. The machine did not create the law it is interpreting. It inherited a world built from human judgment.
So when someone tells me that AI may eventually become better at legal reasoning than a human judge, I find that troubling, not because I assume the machine will always be worse, but because of how easily human judgment could begin receding from the process one piece at a time. Perhaps we start with routine disputes or cases in which the law appears settled and the facts fit familiar patterns. The systems perform well, so we trust them with a little more, and then a little more after that. No single decision feels momentous enough to change the institution, but over time the human role becomes thinner.
And judges would not be changing in isolation. The rest of the legal system would be changing around them too. Lawyers would increasingly rely on agents to research issues, develop arguments, draft briefs, and respond to arguments generated by agents on the other side. Lawmakers would use AI to help draft legislation. Opinions that later agents read may themselves have been substantially shaped by earlier agents. What begins as assistance at each individual step can slowly become something much larger when all of those steps begin feeding one another.
Now move forward ten years. The statute an agent is interpreting may itself have been drafted with substantial AI assistance. Agents may have written the briefs arguing about what that statute means. Another agent may help analyze those arguments and draft the opinion interpreting it. That opinion becomes precedent, and the next agent uses it to interpret the next case.
There may still be a human at every step. A lawmaker votes for the bill, a lawyer signs the brief, and a judge signs the opinion. From the outside, the legal system may look much as it always has. But if machines increasingly perform the intellectual work between those human acts, the material feeding the next machine begins to change. A system that became extraordinarily good at law by drawing on generations of human judgment may eventually be interpreting laws drafted with AI, through arguments developed by AI, using precedent shaped by earlier AI analysis.
That is the feedback loop that concerns me because law is not a static answer key. The disputes coming before courts do not arrive in exactly the same form forever. Legislatures respond to new problems, lawyers frame new arguments, and judges apply existing law to facts that may not look quite like the ones that came before. Sometimes a precedent fits cleanly, and sometimes it does not. That friction matters because it forces people to decide whether a difference is legally meaningful or whether an argument exposes something the old rule did not anticipate.
The concern is not simply that an AI judge might make a mistake. Human judges make mistakes too, and these systems may eventually become extraordinarily good at getting legal questions right. I have no idea what they will be capable of ten or twenty years from now. What concerns me is that each reasonable handoff may make the next one easier while the human experience entering the loop recedes.
Earlier this month, I wrote about the engineer who knew where to put the chalk mark on a broken machine. My concern there was that if AI does more of the work through which lawyers traditionally developed expertise, eventually everyone may know how to operate the machine while fewer people understand where the chalk mark belongs. This is the same concern one generation later. The first machines learned where to put the chalk mark by studying generations of human beings who put it there before them. The next generation may increasingly study the chalk marks left by the machines that came before.
The danger is not that one day we wake up and discover that machines have replaced the legal system. It is that the replacement happens gradually, through thousands of reasonable efficiencies, while lawmakers still vote, lawyers still sign briefs, and judges still wear robes. By the time we notice what has changed, we may be left with a system increasingly refining its own prior interpretations while the human judgment and human experience that originally gave those interpretations meaning have receded from the loop.
None of this means we should resist these tools. But it does mean the work we keep for ourselves matters more, not less. The machine became good at law by learning from human judgment. If we want to keep that judgment in the loop, we have to keep exercising it.
