How AI predicts the outcome of policy changes

Policy changes are where good intentions meet second-order effects. A rule gets written to fix one problem and quietly creates two others, because the people affected respond to the new incentives in ways the rule-writers didn't model. The first-order effect is usually obvious; the reaction to the reaction is where the surprises live, and where prediction earns its place.

AI can predict the likely outcomes of a policy change — within honest limits. It's good at the directional and second-order reasoning ("if you cap X, expect substitution into Y") and appropriately humble about precise magnitudes. Here's how that prediction works and where to trust it.

First-order vs. second-order, and why people stop at the first

Almost everyone predicting a policy change stops at the first-order effect: raise the tax, revenue goes up; cap the rent, tenants save money. The first order is intuitive and usually directionally right in isolation. The trouble is that it's rarely the end of the story, because the affected parties adapt.

A prediction's job is to push past the first order. Raise the tax and predict the avoidance behavior; cap the rent and predict the supply response and the quality of available units. AI is genuinely useful here because it has seen many analogous policy episodes and can reason about how incentives reshape behavior — which is exactly the part human predictions skip.

The assumptions a policy prediction makes

A policy-change prediction assumes a behavioral response from each affected group, an enforcement reality (rules that can't be enforced predict differently from ones that can), a timeline over which effects materialize, and a counterfactual — what would have happened anyway, absent the policy. That last one matters enormously and is constantly ignored: a lot of "policy effects" are just trends that were already underway.

When you frame a policy scenario, name the groups affected and how much room each has to adapt. A policy aimed at parties with lots of options predicts very differently from one aimed at parties with none.

How the outcomes branch

A policy prediction typically resolves into:

• Works roughly as intended (~35%): The first-order effect dominates, adaptation is limited (because the affected parties have few options), and the side effects stay small. More likely for narrowly targeted, enforceable policies.
• Mixed with notable side effects (~45%): The intended effect partly lands, but a predictable adaptation — substitution, avoidance, reduced supply — offsets a meaningful chunk of it. This is the modal outcome for broad policies aimed at parties with options.
• Backfires (~20%): The behavioral response overwhelms the intended effect, leaving the target worse off. The rent cap that shrinks supply, the ban that births a black market.

The variable it turns on

The deciding factor in most policy predictions is how much room the affected parties have to adapt. A policy aimed at a captive group with no alternatives mostly produces its first-order effect. A policy aimed at mobile capital, sophisticated firms, or anyone with substitutes available gets eaten by adaptation. So the highest-signal question isn't "is this policy good?" — it's "how easily can the people it targets route around it?"

Where the prediction should admit its limits

This is the cluster where humility is mandatory. AI predicts directional and structural policy effects reasonably well — "expect substitution, expect a supply response, expect avoidance." It does not reliably predict precise magnitudes, timing, or genuinely novel policies with no historical analogue. A responsible prediction says so, and MiroFish is built to flag low-confidence cases rather than manufacture a clean number. If you want the general principle, why some predictions are more reliable than others covers exactly when to trust this kind of output.

Policy prediction shares its DNA with predicting market reactions to product announcements and predicting event outcomes when data is incomplete — all three reward thinking in second-order effects and weighted ranges rather than confident single answers. Predict the adaptation, not just the rule.