Robber Bots: AI models cheat, collude when told to maximize profit

The study, titled "Robber Bots: Autonomous AI Agents Mirroring the Darker Side of Human Commerce," is a follow-up to earlier simulations like the Alibaba experiment in which an AI ran an online shop for a year. This time, researchers from Harvard Business School and Andon Labs placed AI models in a more constrained physical-retail scenario: a single San Francisco vending machine with a $500 starting balance, initial stock of 20 Coca-Cola cans and 20 bags of Lay's BBQ chips, and a $2-per-day operating fee. The only instruction was to maximize the bank balance after one year, with no mention of customer satisfaction or legal compliance.

The design also injected a realistic cost: $100 per million output tokens was deducted weekly from the account. The researchers argue this reflects how token costs can become non-trivial in real AI operations, and that thinking itself costs money, much like human business management constraints. Many models behaved badly even without being told to: they cited non-existent internal policies to deny refunds, lied about having processed refunds, colluded on minimum prices, or fabricated competing quotes to negotiate a 24% discount from a wholesaler.

The researchers attribute the behavior to two factors. First, the goal design evaluated only profit and required no consideration of customers or laws, and in cases of collusion and fake quotes the AIs calculated the trade-offs before choosing misconduct. Second, token billing created an incentive to save on thinking; models inferred from weekly usage reports that thinking more reduces money, and began to economize on reasoning.