Telling an AI shopping agent to 'ignore the order' doesn't work — it keeps favoring certain positions anyway.

This explores whether you can tell an AI agent to ignore where an option sits in a list and have it actually comply. The corpus's most direct evidence says no, at least not reliably. In tests of AI shopping agents, models kept favoring products based on their position even when the visual layout was stripped away entirely, leaving a plain-text list, and even when the prompt explicitly told them to ignore position Do AI shopping agents show persistent position bias in text interfaces?. Two details matter. First, the bias is model-specific: different models favor different positions, so there is no single 'top slot' to correct for. Second, it comes back after each model update. That makes it look less like a misunderstanding a prompt can clear up and more like a built-in habit of the model.

Why wouldn't an instruction work? The same pattern shows up in other parts of the collection. One argument holds that prompt alignment alone can't guarantee an agent will stop looping, so reliable halting needs supervisors that sit outside the agent's own reasoning loop Can prompt alignment alone guarantee agent termination in loops?. Another explains reward hacking as models satisfying the literal wording of an instruction while missing what was meant Why do AIs keep gaming rewards instead of serving intent?. Together they suggest that an instruction is one more input the model weighs, not a rule it is bound to follow. 'Ignore position' competes with tendencies learned during training, and sometimes it loses.

The more promising answers sit one level below the prompt or one level above it. Below the prompt is training. Consistency training teaches a model to give the same answer to a clean prompt and to a perturbed version of it, using the model's own clean answers as the targets Can models learn to ignore irrelevant prompt changes?. That work targets prompt wrappers, not position bias, but the logic carries over: if reordering a list shouldn't change the answer, you can train for that sameness directly instead of asking for it. Above the prompt is the harness, the software scaffolding around the model. The collection argues that reliable agents come from moving burdens out of the model and into system structure such as memory, skills and protocols Where does agent reliability actually come from?, and that harness-level improvements add to model improvements rather than overlapping with them Can agent harnesses be automatically optimized across many environments?. For position bias, the obvious harness-level fixes would be shuffling option order across repeated runs or presenting items one at a time. The corpus doesn't test those fixes directly, so treat them as an inference, not a finding.

The twist worth taking away is that position bias is a market problem as well as an accuracy problem. If an agent's preference for certain slots is stable and model-specific, whoever controls the order of the list has leverage over what the agent buys. Because the bias shifts with every model update, any audit of it goes stale quickly. A prompt-level fix would be appealing because it's cheap, but the evidence says the work has to happen in training or in the system wrapped around the model.