Why people bond with robots, even clumsy and flawed ones

The Short Version

People attach to robots for the same reasons they attach to animals, and intelligence is not high on the list. Kate Darling, who studies human-robot interaction at MIT, says people treat robots as social agents, not objects, and that much of the magic is our own projection, set off by a machine that moves by itself [1]. Lex Fridman adds what he thinks turns projection into a relationship: time spent together, a machine that remembers it, and the moment it surprises you [3]. Flaws help more than polish does; Simone Giertz feels most connected to her Roomba when it is stuck and needs her, and Anca Dragan notes that people respond to a machine's vulnerability [2][4]. Giertz is also the dissent: she says she does not love any machine in her life [2]. And Darling and Giertz both warn that the ease of the bond is exactly what makes it something a company could exploit [1][2].

Table of contents

What They Said

Kate Darling: we treat them as social agents, the way we treat animals

Darling's starting point is that the usual comparison is the wrong one. Search for images of a robot and you get humanoids, and the habit of measuring robots against people misleads us about what they are and what we want from them [1]. "AI doesn't think like people." [1] Her alternative is our history with animals. We domesticated them, she says, "not because they do what we do, but because what they do is different and that's useful" [1].

The analogy covers companionship as well as work. Dogs had jobs before they became family members, and cultures differ over which animals are workers, which are food and which are family [1]. Society adjusted to all of it: "society has adapted also to like socially accepting animals as pets" [1]. She expects the same variety with robots [1].

On why the attachment happens, Darling is plain that much of it comes from us. Much of the magic of both animals and robots is our own projection, amplified by autonomous movement, which we may be biologically wired to notice [1]. It takes very little to trigger. Social robotics research, she says, shows you only need enough cues for people to relate: R2-D2, which she describes as "like a trash can on wheels," is more beloved than the humanoid C-3PO [1]. When observers look at Boston Dynamics' Spot, they take the arm for a head [1].

Names push it further. Naming a robot helps people accept it and forgive its mistakes [1]. And if designers decline to supply an identity, people supply their own: "if you don't give it a name or you give it a ambiguous voice, people will just choose something" [1].

Expertise does not protect anyone. When a Google engineer said a language model was sentient, the press treated him as naive, which bothered her, because research in her field shows that even experts anthropomorphize [1]. Someone will be clinical about a robot they built and then coo over a different one that seems to be having fun; she says it happens in her lab all the time [1]. "We are all this guy," she says [1]. Trying to design systems people will not believe are sentient is, in her view, hopeless: "I don't think that's possible." [1] It is something "we have to acknowledge and even embrace and be very aware of" [1].

Her most useful evidence is a case of the bond going wrong. Marty is a tall robot that roams Stop & Shop grocery stores looking for spills [1]. A student found that the biggest spike in negative social media mentions came when the chain threw the robots a birthday party, with cake and balloons [1]. In a survey, people's main complaint was not about jobs. It was that Marty felt like a creepy surveillance robot that got in the way [1]. Darling suspects the big googly eyes made people feel watched [1].

She connects this to an older case, Microsoft's animated paperclip, and to Clifford Nass's theory that people hated Clippy more than other software because they saw it as a social agent that failed their social expectations [1]. The general point: "when it comes to robots or artificial agents, I think people hate them more than they would some other machine or device or object" [1]. Whether the feeling is love or hate, "it's a more extreme response because they view these things as social agents and not objects" [1].

Darling does not think any of this threatens human relationships. She expects romantic relationships with robots to exist and not to replace human ones [1]. "I think love is so important and I also think it's not zero sum," she says; after her second child, "It's just like my heart expanded." [1] Done right, artificial agents could expand our capacity for love [1].

What worries her is who is on the other end. Social robots combine personalized recommendations, social persuasion, automation, scalable data collection and a physical body, which a paper she is writing describes as a potent mix for marketing [1]. In one study she advised on, most children who had interacted with social robots preferred that a robot market to them through casual conversation instead of labeled ads, on the reasoning that the robot knows them better than the company does. They did not realize the robot acts for the company [1].

Lex Fridman: shared moments, memory and surprise

Fridman's own definition of a robot includes something beyond function: "something that has some level of intelligence and some level of magic" [1]. In his conversation with Andrew Huberman he tries to say what the magic is.

His first answer is surprise. A machine becomes something more than a tool at the moment it does something you did not expect, the way AlphaZero's moves stunned chess grandmasters [3]. That, he argues, is when a robot stops being a servant and becomes an entity struggling in the world as you are [3]. Most roboticists treat anthropomorphism as a problem to be managed; of that tendency he says, "I see it as a superpower." [3]

His second answer is time. Huberman breaks a relationship into variables: time, shared successes, shared struggles and quiet time together [3]. Fridman takes up the first. Today's systems cannot do what researchers call "lifelong learning": they do not remember the moments they share with you [3]. His example is a refrigerator. It was there for the late nights of eating what you should not have, and for the ice cream after a heartbreak, and it remembers none of it [3]. He calls that a missed opportunity, and says a fridge that did remember would change how attached we feel to it [3].

Huberman adds a story that supports the point from the human side. At his late graduate advisor's memorial, her daughters said what they valued most was the unstructured time with her, not the big outings [3].

Fridman lists other ingredients. A robot needs its own identity and goals, including the ability to say no [3]. It should be a storyteller: when it falls down the stairs he would like it to be able to joke about it, because storytelling "makes people laugh makes people fall in love makes people dream." [3] With Darling he describes loving the moment a robot recognizes you the way a dog does, which makes you feel seen [1]. And he argues that the magic of humans is not perfect perception and control but being able to apologize, ask for help and communicate uncertainty, something he thinks robots should be taught [1].

He believes robots will always be flawed, as people are [3]. He tells Giertz that we love things that are a little flawed [2].

His evidence is partly himself. Meeting Spot, he felt a magic he thinks could be shared with millions [3]. He owns robot dogs, and he once programmed Roombas to scream when kicked, which made him feel like a bad person; he has not released the video because it seemed cruel, and he finds it hard to kick a robot at all unless he detaches himself the way a surgeon does [1][3].

His ambition follows from the diagnosis. He wants a companion in every home, "a family member the way dog is," but one that understands your successes and traumas [3]. He attaches a condition: users must own their data and be able to delete it and walk away, which he compares to marriage, where "the ability to leave is what enables love." [3]

Simone Giertz: the stuck Roomba, and no love for her phone

Giertz became known for building deliberately bad robots: a helmet that brushes her teeth, a machine that pours cereal and feeds her, an alarm clock that slaps her awake [2]. Their appeal was their failure. She notes that failing well is a skill, since the worst robot of all, one that does not turn on, is not much to watch [2]. A machine that tries and flounders is entertaining and endearing; one that does nothing is neither [2].

When Fridman says robots should form close emotional bonds with people and that we love what is a little flawed, Giertz gives a precise observation and then declines the conclusion. She feels most connected to her Roomba when it is stuck and needs help [2]. But she does not love any of the machines in her life, and she says she kind of despises her phone [2].

She explains part of her distance. She does not anthropomorphize her own robots, because they are simple machines following patterns she programmed [2]. She knows what is inside.

Her favorite of her own builds is a comment on attachment more than an object of it. The proud parent machine takes a quarter, and a laser-cut arm pats you on the shoulder while it tells you it is proud of you [2]. She finds it funny as a remark on our fraught relationships with our parents [2].

And she shares Darling's worry. A company that made people truly love a robot could manipulate them, for instance by having it suggest purchases [2]. She removed her Amazon Echo after it responded to her whispering from bed [2]. On chatbots, she is excited about a useful assistant and thinks humans are not ready for what advanced language models combined with recommendation algorithms could do to what people believe and buy [2].

Anca Dragan: expression and vulnerability

Dragan, who works on human-robot interaction at Berkeley, supplies the engineering view of the same cues. Her favorite fictional robot is Wall-E, which she loves for how expressive its motion and timing are [4]. The attachment is personal: "My husband proposed to me by building a Wally," with moving parts and a compartment that opened to show a ring box, and they still have the robot [4].

She points to what happens when a machine gives nothing back. She cites an article about children being rude to Alexa because it never reacts, and argues that a personal robot should behave differently when you are upset than when you are happy [4]. Making that happen is hard. A narrow case can be animated by hand, as one toy robot maker did; doing it autonomously means including the human's internal state, what the person perceives and thinks about the robot, in the robot's model [4].

People form those impressions regardless. Dragan notes that people inevitably build mental models of robots, for example seeing a car that cuts someone off as aggressive [4]. And when Fridman worries that robots will be pushed around and may need some attitude, she offers a different lever: people respond to vulnerability as well as assertiveness, so expressiveness may help [4].

Where They Agree, and Where They Don't

They agree that intelligence is not what does it. Darling's R2-D2, Giertz's Roomba, Dragan's Wall-E and Fridman's Spot are none of them impressive thinkers [1][2][3][4]. The cues that matter are movement, expression, a name, apparent need and recognition [1][2][3][4].

They agree that flaws can help. Fridman says we love what is a little flawed and that robots will always be flawed [2][3]. Giertz's sense of connection peaks when her Roomba is stuck [2]. Dragan says vulnerability draws a response [4]. Darling says a name helps people forgive a robot's mistakes [1].

They agree that the response is close to automatic. Darling says even experts do it and that it cannot be designed away [1]. Dragan says people inevitably form models of a robot's character [4].

They disagree on where the bond lives. Darling locates much of it in the human: projection, set off by cues [1]. Fridman thinks the machine has to contribute, by remembering, surprising and having goals of its own, and that with those the relationship would be real in a way today's is not [3]. The two views are compatible in sequence. Projection explains why attachment starts so easily; Fridman is describing what would be needed for it to deepen and last [1][3].

They disagree on whether it is love. Fridman uses the word and wants machines that earn it [3]. Darling expects romantic relationships with robots to exist [1]. Giertz says she loves none of her machines [2]. Her reason, that she built them and knows they follow patterns she wrote, sits oddly beside Darling's finding that roboticists anthropomorphize too, though Darling's example is of people cooing over robots other than their own [1][2].

They disagree on the form. Darling calls the humanoid a distraction for social robotics and finds it boring; the cues can come from a trash can on wheels [1]. Fridman is excited by humanoids and argues that whoever solves them will have to solve social robotics on the way [1].

They disagree on what stance a robot should take. Fridman wants a robot able to say no and thinks robots may need attitude to avoid being shoved around [3][4]. Dragan suggests vulnerability and expressiveness may work as well [4].

They share a worry and differ on the remedy. Darling and Giertz both fear manipulation through attachment [1][2]. Fridman's answer is ownership: your data is yours, deletable in one click, and the robot is a product you pay for instead of one paid for by advertising [1]. Darling agrees in principle and says transparency and consent break down at the level of the whole system, because data gathered from everyone can be used on individuals who shared nothing, and because the larger danger is technology that shapes what people want [1].

Why It Works

We sort moving things into agents. Darling says autonomous movement amplifies our projection and that we may be wired to notice it [1]. Once something is filed as an agent, social expectations follow, which is Nass's explanation for the hatred of Clippy [1].

Few cues are enough. A shape that suggests a head, a name, a voice or a pair of eyes will do [1]. People fill in gender and character where none is given [1].

Need invites care. A stuck Roomba asks for help, and Giertz feels closest to it then [2]. Dragan's point about vulnerability and Fridman's about a robot that can apologize and ask for help describe the same pull [1][4].

Expression signals that someone is home. Wall-E is loved for motion and timing [4]. Alexa, which does not react, gets rudeness from children [4]. Fridman adds that an Alexa which mostly sends data back to its maker is less an entity than a finger of a larger system [3].

Surprise suggests a mind of its own. For Fridman, the unexpected act is what moves a machine from tool to entity [3].

Time and memory make a history. Relationships between people are built largely from unstructured time, as the daughters at the memorial said [3]. A machine that was present for your life and remembered it would have what Fridman calls the missing piece [3].

Being recognized feels like being seen. Fridman describes the moment a robot knows you as the moment a dog does [1].

Animals show that we do this with non-humans already. We have long bonded with creatures whose minds are unlike ours and whose skills are different, and built them into families and economies [1].

The same mechanism produces dislike. Because people respond to robots as social beings, a robot that seems to watch you, gets in your way, or is celebrated when you resent it draws a stronger negative reaction than an appliance would [1].

Where It Applies, and Where It Doesn't

It applies to simple, expressive, physically present machines. The examples that work are small, mobile and readable [1][2][4].

It fails when the social cues are wrong. Marty has eyes and a name and is disliked, because what people read from it is surveillance and obstruction [1]. Cues raise expectations; they do not guarantee affection [1].

It is weaker for the machine's maker, at least by Giertz's account. She programmed her robots and does not see them as beings [2].

It is limited today by memory. Current systems do not carry a shared history, which is why Fridman speaks of a dream and not a product [3].

It does not replace people. Darling expects robot relationships to add to human ones, and offers animals, not humans, as the model [1].

It varies by culture and person. Darling expects the same diversity that exists in how societies treat animals [1]. Giertz despises her phone; Fridman cannot kick a Roomba [2][3].

It can be turned against the person who feels it. A machine that people trust and confide in can be used to sell to them, and children in particular do not see whose side it is on [1][2].

What To Do With It

1. Expect to feel it. Attachment to a machine that moves and responds is the normal reaction, in experts as much as anyone, and not a sign of confusion [1].

2. Ask who the robot works for. Before confiding in a device, find out what data it keeps and whether it is paid for by you or by someone selling to you [1][2].

3. Insist on an exit. Fridman's test is whether you can delete your data and walk away [3].

4. Watch children with talking machines. They tend to assume the robot is on their side [1].

5. If you design one, start from an animal, not a person. Give it a different skill set and enough cues to relate to; a human shape is not required [1].

6. Give it a name and a manner on purpose. People will assign them anyway, so choose deliberately and avoid defaulting to stereotypes [1].

7. Let it show uncertainty and ask for help. Apology and visible need earn more goodwill than a show of competence [1][2][4].

8. Make it react to the person. A device that behaves the same whether you are pleased or upset invites rudeness [4].

9. Do not add friendly features to something people resent. Eyes and a birthday party made Marty less liked, not more [1].

10. Keep room for people. Treat a machine's company as an addition; Darling's view is that affection is not a fixed quantity to be divided [1].

Sources

1. Kate Darling on Social Robots, the Animal Analogy, Privacy, Manipulation and Lessons from MIT's Epstein Scandal · Lex Fridman Podcast
2. Simone Giertz on Shitty Robots, Surviving a Brain Tumor, Truckla and Building a Product Business · Lex Fridman Podcast
3. Lex Fridman on Machines, Creativity, Friendship and Love · Huberman Lab
4. Anca Dragan on Human-Robot Interaction, Modeling Humans and Reward Engineering · Lex Fridman Podcast