Tools, Not Beings

How the study worked

Seventeen robot images went on a board: industrial arms, quadrupeds, humanoids at trade shows and humanoids at home. Twenty-one people in the United States each got green dots and red dots and one instruction. Drag green dots on the robots that look helpful. Drag red dots on the robots that look creepy.

The platform sorts the result at a 60% consensus threshold. An image is helpful when at least 60% of the dots it drew were green, creepy when at least 60% were red, and contested in between. Of the 17 robots the report pictures, seven cleared the bar as helpful, nine crossed it as creepy, and one split the room. See the full report.

Every number on this page is a share of that image’s own dots. Green and red are the study’s own encoding, so the verdict is written next to each score.

The one variable: tool or being

Line the 17 robots up by the share of dots they took green and they do not spread out along a friendliness scale. They sort into two piles with almost nothing in between. One pile is equipment. The other pile is company.

Figure 1 · Every robot, ranked

The winning images are obviously machines. Exposed servos, cable runs, a brushed finish, a job being done on an object. The losing images are ambiguously people. Smooth white shells, a face that is almost a face, a person-shaped thing standing in a person-shaped space. The more a design blurred the line between those two categories, the more red it took.

What read as helpful: the arms

Five arm-based robots scored between 90.91% and 100% helpful, and four of them were unanimous. None has a face. None has a body. Every one of them is shown doing something to an object: welding, folding, picking, writing. Where a human appears, the human is supervising, not being served.

The palette is white, grey, black and metallic silver. Orange appears only as a safety accent. Nothing is glossy, nothing imitates skin, and the Dyna machines make a point of showing their motors. Visible mechanism is not a flaw to be hidden here. It is the reason the room trusted them.

The exception that proves the rule

One humanoid cleared the helpful bar, and it did so in a kitchen, which is exactly where every other humanoid on the board died. Sunday Robotics’ Mamo scored 92.31% helpful serving food at a home counter, and its head alone scored 85.71%.

Look at the head. A rounded dome, an orange cap, two camera slots where eyes would go. It is a 1960s idea of a robot, and that is the whole trick. The face reads as an appliance, not an entity. Mamo never asks the viewer to decide whether it is a person, so the viewer never has to.

That is the only humanoid strategy the data supports. If the form has to be human-shaped, make the head unmistakably a machine. Vintage futurism reads as equipment. A soft, minimal, almost-face reads as something else, and the next two sections show what happens to it.

What read as creepy: the humanoid problem

Every robot in the creepy array is humanoid, pseudo-humanoid or animal-form. Boston Dynamics’ Atlas went 100% red. Figure’s studio unit with a black visor for a face went 83.33% red. The iCub, a child’s face on a metal body, went 84.62% red. Even the Toyota partner robot playing a violin, the most obviously harmless thing on the board, took two thirds of its dots red.

The quadruped extends the pattern. A grey Unitree dog with no face at all scored 76.92% creepy. Animal form is as much of a problem as human form. The issue is not the face. It is the suggestion of a creature.

The gentle-helper trap

Here is the finding that should worry a consumer robotics marketing team. Domestic settings and caregiving activities make humanoid robots more creepy, not less. A Figure unit watering a plant at a home sink scored 81.82% creepy. A 1X robot being admired by an elderly man scored 87.50% creepy. Placing a human-shaped machine in a human-inhabited room sharpens the question of what it is, and the room answered.

The fabric face is the failure mode. The 1X units with white fabric heads and two dot eyes score between 87.50% and 100% creepy. The intent is obvious: soft materials, rounded forms, minimal features, nothing threatening. The result is the strongest rejection in the dataset. Making a face softer is precisely an attempt to make it more like a being, and being-like is what people reject.

Compare the plant waterer with the whiteboard arm. Both are doing a small, useful chore. One is a machine at work. The other is a stranger in your bathroom. Same task, opposite verdict, and the only difference is the body.

The whole study in two images

The cleanest pair on the board is two robots built for the home. Both are shown in soft, domestic terms. Both were unanimous.

Same domestic ambition. One says equipment. The other says someone. The room needed no instruction to tell them apart.

One image split the room, and it is the one that treats humanoids as merchandise. The X1 product lineup, three fabric-faced units in grey, white and black, went 58.33% helpful and 41.67% creepy. Presented as objects for sale rather than actors in a scene, the same design reads as slightly less like a being. Slightly. It still could not clear the bar.

What to do with this

The decision logic translates directly into product design and into how a robot is photographed. Show robots doing things, not being present. Task completion beats social context, and a technical-documentation aesthetic beats lifestyle photography every time.

One warning before you run with this. The sample is 21 people, which the report rates as directional with moderate confidence. The patterns are consistent and internally coherent, but treat them as hypotheses to validate with a larger sample and demographic cuts before a design programme is built on them. The same mechanism, category clarity beating decoration, shows up in the orange juice study, where every layer of styling between the fruit and the picture read as distance.

Frequently asked questions