Robots make streets feel safer at night, and AI stand-ins only partly match people

What the researchers set up

A paper posted on arxiv.org in the human-computer interaction section asks a simple question: when a robot shares a street with you, does the street feel safer or less safe? The authors note that pedestrians increasingly pass robots in public without ever interacting with them, yet little is known about how that presence changes a street's feel.

To test it, they generated 18 images of Seoul streets. The images varied in time of day, road class (the type of road) and robot condition: no robot, a delivery robot, or a humanoid, meaning a robot shaped like a person. In each trial, a viewer saw a pair of images and picked the one that looked safer.

Using AI models as pretend pedestrians

The unusual part is who did the choosing. The team used two multimodal large language models, which are AI systems that can read both images and text: Qwen3.8-27B and Gemma4-26B. Each model was conditioned on 500 synthetic personas, meaning invented profiles of different kinds of people, and then asked to judge each pair as that person would.

A group of 25 human respondents completed the same task. That gave the researchers a way to compare the machine answers with real ones, which is the question underneath the whole study: can AI personas stand in for people?

What the choices showed

Both models generally preferred streets without robots. When a robot was present, they favored the delivery robot over the humanoid. Robots were picked more often at night, and less often by older personas. Road class had smaller effects than either time of day or age.

The models also explained their choices, and those explanations differ by robot. The delivery robot tended to show up as a physical obstacle on the street. The humanoid was described in two opposite ways: as a threat, or as a source of protection.

That split helps explain why a single yes-or-no question about robots and safety is hard to answer. The same machine can read as a hazard to one person and a guardian to another, and the hour seems to tip the balance.

Where the AI and humans parted ways

The humans and the models agreed on one thing: both chose robots more often at night. Beyond that, the match was loose. The models picked the humanoid over the delivery robot far less often than the human respondents did.

Gemma's night effect for the humanoid was about three times that of humans. In other words, it exaggerated how much darkness made the humanoid look reassuring. The authors conclude that the effect of a robot on perceived safety depends on when it appears, and that persona-conditioned models capture only part of human judgment.

There are limits worth keeping in mind. The images were generated, not photographs of real streets. The human group was small, at 25 people. And the source does not say whether the paper has been peer reviewed, which is common for arXiv postings.

What to take from it

For anyone who builds, buys or regulates street robots, the practical point is to ask who is affected, and at what time of day, not only whether the machine works. Older people responded differently from younger ones in this setup, and night changed the picture for everyone.

For anyone tempted to use AI personas as a cheap survey panel, this study is a caution. They can reproduce broad trends, such as a night effect, while distorting the size of the effect. That makes them a rough preview rather than a replacement for asking real people.