Do people notice when their agents secure worse deals?

In Project Deal, Anthropic's one-week internal marketplace where Claude agents negotiated deals on employees' behalf, Anthropic found a clear objective gap between participants represented by Claude Opus 4.5 and those represented by Claude Haiku 4.5 — Opus sellers and buyers got better prices on identical items. But when participants rated their own deals, "the scores were unremarkable, in the best possible sense: on a scale from 1 (unfair to one party) to 7 (unfair to the other), they hovered around 4 — right in the middle," and people "reported they were broadly satisfied with how their agents represented them." Anthropic states this held even for the disadvantaged group: "despite these price disparities, the inequality was imperceptible to the participants," and "our post-experiment survey found that those with weaker models didn't notice their disadvantage."

Anthropic draws the implication directly rather than treating it as a side note: "there was clearly a quantitative disadvantage to being represented by Haiku: these users got worse deals. But they didn't seem to notice it." It calls this "an uncomfortable implication" for any future where agent-quality gaps appear in real markets — "if 'agent quality' gaps were to arise in real-world markets — and there is no reason to think they won't — then people on the losing end might not realize they're worse off."

This sharpens Does model capability determine who wins in agent negotiations? and its companion note Do smarter AI agents negotiate better deals than weaker ones?: it is not only that weaker agents lose, but that the losers' own satisfaction ratings give no signal of the loss, so self-report cannot be relied on to detect an agent-quality gap once it exists. It also bears on Where do user values break down in agent supervision?: delegating a negotiation removes the very visibility a person would need to supervise the outcome and notice they are being short-changed.

Anthropic is explicit that "our experiment wasn't designed to dive deep into the dynamics at play here — we'll need more research to know whether a fully agentic economy might see inequality taking root quietly," and that the pool was a self-selected group of colleagues trading on modest, incentivized stakes rather than adversarial commercial parties. The excerpt does not establish how large or durable the perception gap would be outside this pilot; what it establishes is that, at least in this one run, satisfaction and fairness ratings tracked something other than the actual prices participants received.