Fast AI decision models followed shopping defaults almost every time

A shopping test for generation-free AI

A preprint by Sascha Kaltenpoth, Haya Halimeh, Kevin Bösch and Oliver Müller at Paderborn University examines whether digital nudges steer a newer class of AI decision models. These models choose among predefined options rather than generating a written response. The paper, submitted on November 6, 2026, is listed as under review; its findings remain preliminary.

The researchers tested Jev, Microsoft’s Decision-1 and Liquid AI’s d1. Their central question was whether these models respond differently to two ways of influencing choices: making an option the default, or attaching information about customer satisfaction. Both leave the alternative available.

How the experiment worked

Each agent shopped in a simulated grocery store, choosing between two products in each of six categories before reviewing its selections and confirming the purchase. There were three conditions: a neutral interface, an interface with one product preselected in each category, and an interface labeling one product as having the highest customer satisfaction in its category.

Each model completed 200 runs per condition, totaling 600 runs per model and 1,800 overall. Across those runs, the experiment recorded 10,800 product choices. The researchers switched which product was targeted in half the runs and randomized product and category order, reducing the risk that an existing product preference or screen position would explain the differences.

The models did not navigate by interpreting screenshots in this setup. A modified browser-control library supplied visible page text, previous actions and clickable elements as available choices. At each step, the agent executed the option assigned the highest probability by the model. Runs exceeding 60 actions were stopped and restarted, with up to three attempts per task.

Defaults produced the strongest shift

Without a nudge, estimated selection probabilities for the designated target product ranged from 47% to 49%, close to an even split. With a preselected default, those estimates rose to 97.5% for Decision-1, 99.4% for d1 and 99.9% for Jev.

The customer-satisfaction label also shifted choices substantially, but less strongly: estimated target-selection probabilities were 75% for Decision-1, 82% for Jev and 96% for d1. Thus, every tested model responded more strongly to the default, although d1 was particularly responsive to the social information.

The team used Bayesian multilevel logistic regression, accounting for variation between runs. Both nudges increased target selection for every model, and the stronger default effect was supported by the reported uncertainty intervals. These are findings about which product was selected—not evidence that the selected products were better or worse for the shopper.

A human-thinking analogy, not a demonstrated mechanism

The authors interpret the pattern through dual-process theory, which distinguishes fast, automatic human thinking from slower, deliberate thinking. They classify defaults as automatic, Type 1 nudges and customer-satisfaction information as reflective, Type 2 nudges.

Developers sometimes call decision models “System One” models because they choose directly without generating intermediate reasoning text. The stronger response to defaults fits the authors’ prediction from that analogy. It does not, however, demonstrate that the models possess human cognitive systems or establish the internal mechanism behind their choices.

An earlier study cited by the authors found the opposite ordering for reasoning-capable language-model agents: social information influenced them more than defaults. That comparison is suggestive rather than a controlled head-to-head result, because the present study changed the task wording and browser implementation slightly.

What the findings can—and cannot—support

Completed runs averaged 5.8 seconds and cost $0.0012 at list prices in this experiment. Those measurements describe this particular setup, not a general guarantee about speed or cost across websites and tasks.

The study covers only three closed-source models, one simulated shopping task and one example of each nudge type. It therefore cannot establish that automatic nudges generally outperform reflective ones for all decision models. The authors propose testing additional models and nudges, and comparing decision models with language-model agents under identical conditions.

The narrower finding is clear: avoiding text generation did not prevent interface cues from steering these agents. In this experiment, the store’s preselected choices strongly shaped what the agents bought.