AI made the reasoning free. It didn’t make the choice any cheaper.
The conventional worry about AI is that it produces falsehoods. Hallucinations. Confident nonsense in fluent prose. That is a real problem, and there are real people working on it.
It is not the hard problem.
The hard problem is what happens when everything the machine produces holds up. Internally consistent, well-sourced, logically sound — and pointing toward a different action than an equally valid reading of the same facts. Ask an AI to analyze why a company failed and you will get a rigorous narrative about market timing. Ask again with a different emphasis and you will get an equally rigorous narrative that reaches a different conclusion. Same data. Different weights. Both check out.
The difference between them is not in the evidence. It is in what the evidence is asked to mean.
That has always been true. What is new is who can now do it, and how fast. Getting a second serious interpretation of anything used to take time, money, or access — a lawyer, a second consultant, a team willing to run the numbers a different way. Most people went through life encountering one or two framings of any given situation. The limited supply felt natural, as if the facts were speaking for themselves.
They never were. And the cost of producing them was doing something no one noticed until it disappeared: it forced commitment. When you could only afford one interpretation, you acted on it. Commitment, it turns out, was doing all the work.
To see what this costs when that constraint disappears, walk through a custody evaluation.
Two parents have separated. Both are competent. Both want primary custody of their eight-year-old daughter.
An AI-assisted evaluation produces three analyses from the same case file.
The first emphasizes continuity and stability. The daughter has attended the same school for three years, has established friendships in the mother’s neighborhood, and shows strong academic performance. Disrupting that environment, this analysis argues, risks real developmental harm. Recommendation: primary custody to the mother.
The second emphasizes the father’s increasing involvement over the past eighteen months — coaching her soccer team, attending every parent-teacher conference, rearranging his work schedule around her medical appointments — and cites research linking active paternal engagement to stronger emotional resilience in girls. Undermining that trajectory, this analysis argues, sends a message about which parent’s investment counts. Recommendation: expanded shared custody.
The third takes the child’s own statements as its starting point. She said she wants to spend more time with her father. She also said she doesn’t want to change schools. She is eight. She does not know that these two preferences may be incompatible given the distance between the households, but she expressed them clearly. Recommendation: a flexible arrangement built around what the child wants.
Same daughter. Same parents. Same file. Three frameworks, each grounded in real research, each internally coherent, each reaching a different conclusion about where this child wakes up on Monday morning.
The judge has to choose. Not which analysis is correct — they all are, within their own terms. The case file does not make that choice. The research does not make it. The AI, which produced all three with equal confidence and no preference, certainly does not.
The judge does. And the choice reflects not the data but something the data cannot contain: what “the best interest of the child” actually means when stability, involvement, and the child’s own voice are all real goods, all supported by evidence, and all pointing in different directions.
But notice what forces the judge to decide: the child needs somewhere to sleep tonight. She needs a school on Tuesday. The calendar moves whether or not the analysis is complete. Not every decision comes with that kind of deadline. Consider the manager who keeps scheduling another round of input before committing to a direction. The committee that orders another review. The institution that has been “revisiting the question” for eighteen months. They face no Tuesday morning. When there is no external pressure demanding a resolution — when the world will not obviously suffer if you keep looking at it from another angle — the abundance of coherent interpretations stops being clarifying. It becomes permission. Permission to keep deliberating indefinitely, while appearing to take the complexity seriously.
To understand why that permission is new, you have to understand what it replaced.
For most people, for most of history, this kind of abundance was not available.
A legal brief took weeks to produce. A competing narrative needed institutional backing and a platform willing to carry it. The work of generating a serious alternative interpretation was expensive enough that most people went through any given situation encountering only one framing, maybe two. The limited supply felt natural, as if the facts were speaking for themselves.
The facts were never speaking for themselves. Someone was always choosing which story to tell. The expense just hid the choosing.
But expense did something else, something nobody noticed until it was gone: it forced commitment. When you could only afford one interpretation, you acted on it. You built on it. You discovered what it got wrong, and you revised. Not because you chose wisely — often the only available framing was narrow, biased, or self-serving — but because you had no alternative to acting on something, and acting was the only way to learn where the framing failed.
That cycle — commit, act, encounter consequences, revise — is how understanding gets built over time. It is how courts learn from their own decisions, how a scientific field corrects itself, how a culture arrives at shared frameworks for what matters. Every step requires someone to hold a position long enough for reality to push back.
AI makes the first step optional.
When a machine can produce a dozen defensible framings in minutes, committing to one becomes a choice rather than a necessity. You can switch to whichever framing suits the moment. You can hold every interpretation at arm’s length, adopt whichever one supports your position, and discard it when the situation shifts.
This temptation is not evenly distributed. It is strongest where being visibly wrong is costly but being visibly uncommitted is invisible — the advisor whose value lies in analysis rather than calls, the consultant who can always produce one more perspective, the committee whose mandate is to study rather than decide. For these people, the abundance of coherent framings is not a burden. It is a kind of cover: you are never wrong because you are never committed. You are never committed because there is always another angle worth considering. And a world where commitment is permanently deferred is a world where interpretation never accumulates weight — because weight requires standing behind a reading long enough to find out what it costs.
We thought we were good at interpreting the world. We weren’t. We were good at operating under constraint. Scarcity did the choosing for us. We mistook its constraint for our competence.
AI didn’t create the interpretation problem. It revealed that we had never actually practiced choosing — only enjoyed the luxury of never being asked.
Building that practice now probably looks less like acquiring better frameworks and more like choosing before it is forced. Committing when alternatives are still available. Letting a reading accumulate weight before reality demands a reckoning. Not a technique. Closer to a discipline: the deliberate acceptance of the cost that cheap coherence would otherwise let you keep deferring.
The custody judge still has to decide. That is the irreducible fact underneath all of this. The framings are all defensible and the child still has to live somewhere. She needs a school on Tuesday. She needs a bedroom tonight. The abundance of coherent analyses does not change the fact that someone must choose — and that the choice will be lived with by the judge, by the parents, and most of all by a girl who will not understand for years, if ever, why Monday morning looks the way it does.
Reality is not self-interpreting. It never was, even when scarcity made it look that way. The facts are abundant. The framings are abundant. The coherence is free. And the thing that was never about coherence — the act of committing to one reading, accepting what it leaves out, and bearing what follows — remains exactly as expensive as it ever was.
Nobody figured out how to automate that.