AI or rule?

Step 1 of 8 · The setting

Two shops, one price cut

Picture two shops on the same street selling the same thing, say a court booking or a litre of fuel. Each sets its own price every day, and many shops now let an AI pricing agent pick that price.

On day 0 the competitor cuts its price by €8 for a single day, then goes back to normal. Everything in this walkthrough is about what our shop does next.

Step 2 of 8 · The worry

What a cartel looks like

A cartel is a group of firms that keep prices high together instead of competing. Making such an agreement is illegal. Competition authorities worry that AI agents could learn to do the same without anyone agreeing to anything.

A cartel only holds if cheating is punished. So when a rival cuts its price, a cartel member answers with a price war of its own: it stays low for several days to teach the rival a lesson, then goes back up. Auditors look for exactly this shape. The chart shows a shop that follows such a punishment rule.

Step 3 of 8 · The innocent case

An honest agent, no rule

Now meet an honest agent. It never punishes anyone: each day it simply picks the best price for the market it sees. When the competitor dips, it dips a little too, for one day, and comes straight back.

An auditor sees a normal reaction. Good.

Step 4 of 8 · The twist

Add a legal platform rule

Many markets have rules about when a price may change. A booking platform may hold a newly approved offer for a few days. Western Australia makes fuel stations fix each day's price 24 hours ahead. Since April 2026, German fuel stations may raise prices only once a day, at noon. These rules are legal, and often required.

Here the platform holds every new price for 3 days. The honest agent still wants to go back after one day (dashed line), but customers keep seeing the low price (solid line). Shaded days: the rule blocked the agent's change.

Step 5 of 8 · The real AI

What a real AI model does

So far the agent was a simple formula. Now the agent is a large language model, Qwen3-235B, prompted as a shop's pricing manager. It sees its past prices, its sales and the competitor's prices, and replies with a price each day. The lines are averages over 64 runs in 16 different markets.

Without the rule (dotted line), it reacts to the cut and mostly recovers. With the rule, what customers see stays low (solid), and so does what the AI wants (dashed). The AI looks at its own held-down price and keeps choosing low prices.

Step 6 of 8 · The finding

Is it the AI or the rule? Both.

With the rule, the real AI's price stays lower after the cut, by an extra per day on days 2–5. We can split that extra into two parts:

Why this matters: an auditor who only sees prices may wrongly accuse the AI. And an auditor who also logs what the AI wanted still cannot clear it, because the rule changed what it wanted.

Step 7 of 8 · Which AI?

Different AI models, very different shops

We gave the same market to open AI models from Chinese and US developers. The question is simple: the day after the competitor's €8 cut, how much does each one cut its own price? Economic theory says about €2.40 for this market.

The answers range from no reaction at all to more than twice the textbook cut. Bigger is not simply safer: in one family the small model overreacts, in others the small models barely react. For the models we also ran with the 3-day rule, the right panel shows how much extra drop the rule adds on days 2–5.

Neither the country nor the size tells you how a model will price. You can give an agent pricing rules, but when an instruction is missing or forgotten, its own sense of prices takes over. That sense decides how it meets a rival's cut, whether the rival is another AI or a human shop, and whether its owner gets the behaviour they expect. So test the exact model, in your market and under your rules.

Step 7 of 8 · How we know, and try it

How we tested it

1. Copy the market at the moment just before the agent sees the competitor's cut.
2. Run both copies: one with the rule, one without. Log what the agent wanted and what customers saw.
3. Replay the agent's no-rule choices through the rule. That is what the rule alone would do; the rest is the AI adapting.

Now try any combination yourself:

Limits: a simulated market with a scripted competitor. A second AI model (gpt-oss-120b) barely reacts to cuts, so the effect is small there. This shows how audits can be fooled; it does not show that any real firm colludes. Not legal advice.

Full playground with every setting · Code and data