People Say They Want Perfect Robots. They Trust the Ones That Push Back

Asked for the most amazing robot she has ever met, Ayanna Howard said she has not met her. She grew up with Rosie from *The Jetsons*: a housekeeper who cooked, cleaned, and kept the kids in line, onto whom viewers projected whatever they wished a robot to be. Rosie was socially engaging. She had an attitude every so often. She pushed back when George Jetson was doing something foolish, and she cared about people, especially the children. Howard, a roboticist at Georgia Tech and director of the Human Automation Systems Lab, called her the perfect robot — and then explained why that word does not mean what engineers usually think it means.

Robots exist, she said, because they are supposed to raise the quality of life, usually by taking over something functional. Even the fascination with self-driving cars is not really about Saturday joyrides. It is about the hour in traffic on the way to work that people genuinely hate. Perfection, in that setting, is an interaction problem, not a scoring problem. If human drivers drove perfectly, Howard argued, nobody would ever get anywhere. People run a yellow because the car behind is about to hit them. They cross where there is no traffic because a child has darted into the street. Those deviations are how the world actually works.

From a theoretical angle, perfection in robotics is tied to accuracy: can a function complete at 100 percent with zero errors? In a self-driving stack that usually means staying in the lane, going on green, stopping on red, and seeing every entity in the scene. Howard thinks that is where the field goes wrong. Rosie was not perfect on accuracy. She was perfect at how she interacted and how she adapted. The robots that succeed with people, she said, are the ones that adapt to the person, not the ones that follow made-up rules at 100 percent. "We really want perfection with respect to its ability to adapt to us."

Lex Fridman asked how hard the last mile of autonomy really is — the gap between robotic perfection and the way people actually function in traffic. Howard called it a moving target. A few years earlier, companies, roboticists, and investors, Elon Musk among them, had talked about self-driving cars on the road with people within five years. That window has since become five years, ten, twenty, or never, depending on whom you ask. The deployments that are working, she said, tend to be fixed environments: people still walk, stores still open, but other human drivers are carved out of the problem. Manufacturing succeeded for the same reason. Humans were not part of the equation.

She uses Tesla Autopilot herself, partly because she is a gadget freak, and she remains hyper-alert — thumb ready to take over even during automated parking or freeway driving. The system gets better as it learns from other drivers, and that is the interesting dance: she still does not trust it, and she still uses it. Early adopters, she warned, tend to swing. Hypersensitivity fades as the tech seems to work, and then people glance at a phone for twenty seconds, then thirty, then a minute. That phase shift is human nature, not a Tesla-specific bug. Twenty years ago people would have laughed at relying on a search engine; now they treat it as obviously correct.

The harder problem is everyone else on the road. A bumper sticker that says Student Driver instantly changes how you drive next to that car. Two young-looking people chatting in the next lane, no sticker, does the same thing. Boston intersections have personalities; neighborhoods differ; night differs from rush hour. Brains pick up those hints. Introspection cannot write them down. Howard's shortcut is not to build a bird in order to fly. It is to design a new neighborhood or smart city around the assumption that people are messy, then let self-selection do some of the work. Residents who move in knowing the pitch will be more forgiving when something goes wrong, and the data collected there still includes human variance — just with the noise turned down. The third obstacle, legal liability and policy, she said she would rather not touch.

Fridman pressed a sharper point. Cars on the road constantly threaten one another's lives. Navigating an intersection takes assertiveness and some risk. Compress driving into an objective function and there is a probability of killing another human being inside it — low enough to live with ethically, high enough that other people do not jaywalk in front of you. The industry slogan is that cars do not kill people, people do. Howard pushed it one layer down. "Robotic algorithms don't kill people. Developers of robotic algorithms kill people." In the near and mid term, a person is still in the loop, even if that person is a developer. We are not yet at robots programming other robots.

That is why she thinks ethics cannot be a side topic that robotics people outsource to a specialist group. You can refuse to work on weaponized AI. The algorithm you do write might still decide whether someone gets a medication, and they might die without it. If you are coding without noticing that you hold that power, she said, that is not a good state. She remembered the era before software had dedicated testers, when developers had to run their own cases. Testing jobs appeared because people were missing too much; then most developers glanced at compiler warnings and assumed the next stage would catch the rest. Ethics, she said, is still in those early days. There is no second line of ethical testers. You have to look yourself.

Fridman had been told he had blood on his hands for working on semi-autonomous vehicles: automation makes people less vigilant, and less vigilant people get hurt. He had spent nights unable to sleep, even while telling himself that 40,000 people die on American roads every year and the work is meant to save lives. Howard did not like the word burden. Doctors lose patients and still walk into the next room knowing this one might not survive. Medical school teaches responsibility and also gives tools so that fact does not destroy the doctor. Developers, she said, are handed the gift without the framing. Recast the message the way medicine does — a great gift, and with it a great responsibility — and the weight becomes something you can carry instead of something that only keeps you up at night.

In a paper titled "The Ugly Truth About Ourselves and Our Robot Creations," Howard laid out biases that shape how robotic systems behave. Bias, as she defines it, is the set of preconceived notions everyone carries about groups, habits, and identity. Those predispositions color decisions. They can be positive or negative; negative bias that you notice and still act on becomes prejudice. The hard part is the gray. Insurance companies charge teenage drivers a lot of money. That is an age bias, it is backed by accident data, and almost nobody calls it unfair — parents grumble, society shrugs. Change the same logic to a Boston neighborhood whose accident rate tracks socioeconomic status, and the surcharge is suddenly unacceptable. History, she said, is the long argument over where that line sits.

Healthcare is full of the same pattern. FDA drug trials historically under-tested women of childbearing age, so a lot of medicines never had complete data by sex. Translate that to hardware: if you design an exoskeleton, what material, what weight, what form factor, and around whose body? Average height and weight for women in the United States differ from men. If you never ask, you will design for the people in your lab. Data-driven systems do not escape this. Society already contains negative bias, so even a balanced sample still carries it. Predictive policing, criminal recidivism scores, and medical algorithms all inherit that history. A widely covered healthcare-algorithm paper was given a media headline about racist AI, which Howard disliked; the research itself, she said, did a full analysis. The discrepancy by race was real. The duty is to do better.

The sentence she thinks does not get said often enough is the other half: the worst AI is still better than us. Take the best humans on these tasks and, on some of them, they still lose to the worst model. Fridman added that the ethical bar for AI is set at a kind of perfect fairness nobody can even define, which is a way of guaranteeing failure. Howard's point is that even a flawed system can be poked. You used to need a lawsuit to ask whether a hospital or a city was mistreating people. Now you can run the data, name the issue, and enter a repair loop. Joy Buolamwini called out facial-recognition vendors; most of them were embarrassed, then they fixed it, then new problems appeared, then those got fixed. The needle moves if both sides show up: developers who will say they heard you, and outsiders who will keep pointing.

Fridman asked whether AI might one day be elected president or sit in Congress. Howard's answer was immediate. She does not believe AI should be president. She does believe a president should use AI as an adviser. Every president already keeps a cabinet of people with different expertise and is supposed to listen to them. An AI for healthcare, another for education, another for the issues a human cannot hold in working memory, could sit in that circle. The last decision still belongs to a person, because a culture, she thinks, needs to see its leader as human.

Most people, she added, do not realize how much advice already sits behind a public decision. Leaders rarely wake up with a brilliant idea and ship it. The usual path is an idea, a round of feedback, a push forward if it holds, and a retreat to talk to more people if it does not. In that picture, AI is one more specialist in the room, not the person at the head of the table — a partnership, which is how she thinks the whole field should be framed.

Asked for a systematic answer to algorithmic bias, Howard offered two paths. The first is to stop treating feedback as a hobby. Right now a researcher notices an unfair outcome, publishes, and a developer or a company may or may not fix it. The pipeline is driven by whatever happens to interest researchers, which means a great deal never gets looked at. She does not claim to have the full process, only that the current one is ad hoc.

The second path is to poke the corporations the way security already does. Companies pay people who find holes in their software. Why not pay people who find an ethics hole — an unfairness — at the same rate? It is a win for the company: it signals that fairness matters, and it does not have to staff every audit internally. It is also a win for coverage. Someone who cares about age will hunt age failures; someone who cares about gender will hunt those. You get many lenses, and you get them in a data-driven way instead of waiting for the next paper.

On platforms such as Twitter and Facebook, Howard thinks the teams are working hard. The evidence the public brings is often anecdotal — a friend said X and got banned — which is a poor way to have a big-data conversation about ethics. Some of the anger, she said, is dependence: I need you right now and you are not working for me. The deeper problem is the instinct to fix everything in house. "I'm a fox and I'm going to watch these hens because I think it's a problem that foxes eat hens." Inside a conflict of interest, the first move after finding something is to patch it quietly and hope the press never sees it. That pattern is how trust collapses. Her alternative is to say out loud that the problem exists, invite the community in, and pay the finder.

Fridman asked whether she still trusts people to steer civilization in a decent direction. Howard called herself an optimist who also thinks these are dark times, and not only in the United States. Polarization is worldwide, and that worries her. What she holds onto is a memory from Atlanta's Snowmageddon: not much snow, enough ice to shut the city. People opened their houses to strangers. Hotels let people sleep on the floor. Grocery stores handed out food. Nobody was asking what it would cost. Strangers offered rides home. That, she said, is the deeper layer — compassion and love — and it tends to surface when people are actually in danger.

When needs are already met, she added, people love drama. Fridman suggested that division itself can be a sign that times are good, because watching other people fight is entertaining at a mammalian level, and Twitter and Facebook harvest that loop because controversy brings users back to an ad-driven product. Howard's coda was dry: love does not sell quite as well. Earlier in the same stretch she had been asked about HAL 9000. HAL, she said, was not evil so much as misguided: built as perfect on a set of assumptions, with no failsafe once those assumptions were wrong. Software that goes out into the world needs a way to be graded an F and pulled back. At some point the fallback is still a human.

Howard started her career at NASA's Jet Propulsion Laboratory. The memory that still stands out is a surgical robot meant to operate on the eye, sometime around 1992 to 1994. It was remote operation with an early form of haptics, the ancestor of the logic behind today's da Vinci systems. She walked into a lab, saw a mock-up of a human face with fake eyeballs and a small drill, and watched someone operate from a distance. The precision stunned her. More than that, it was the first time she had physically seen a person and a machine working in the same space. Factory arms she knew; those were isolated. This was a human and a robot with no barrier between them. "That for me was a magical moment."

The earlier seed was television. *The Bionic Woman* made her decide at twelve that she would be a roboticist, though the word she used then was cybernetics. What captured her was not a fully intelligent being so much as the parts — the limbs, the prosthetics. In another life, she said, she would have been a biomedical engineer. Fridman had recently met Boston Dynamics' Spot in person after years of watching videos, and the few minutes of one-on-one time felt almost like falling in love. Howard recognized that feeling from the surgical lab. She also drew a line she still uses: she cares about *human-like*, not *humanoid*. Humanoid is a body plan. A robot dog that understands people at a social level is human-like. A humanoid only enters her research if it actually interacts with us as if we are human.

Her favorite science-fiction series is *Star Trek*. She has done the math: she will not live to that calendar, and she would still take a berth on the first adventure. Asked about Data after the emotion chip, she said that stretch was a mess, and the reason is simple. Emotions make agents irrational. Without the chip, Data could still reason through an emotional scene on pros and cons. With it, his own feelings became one of the metrics, and the people around him receded. Children are like that: intensely egocentric. Some adults, she noted, stay there. Emotional maturity is a developmental path, and not everyone finishes it.

That is why her group reads a lot of cognitive science and developmental psychology. Those literatures already know a great deal about human relationships and the milestones along the way. The lab tries to match human-human patterns to human-robot ones. Sometimes they line up; sometimes they do not, and then the question is why this scenario diverges while another does not. Looking twenty or thirty years out, Fridman asked whether she would be wearing the psychology hat more often. She said the hardest piece is not psychology and not classical robotics. It is adaptation during the interaction — learning at the interface. She started in electrical engineering and control theory, realized her algorithms needed human data, then realized she had to model human perception, and only then became an HRI researcher working under an AI lens with hardware in the loop.

Leaving control theory for human interaction was a high-risk academic move. The field was young; there was not even a conference yet. What pulled her was that the problems were interesting and hard. She likes walking into a room and not knowing what to do. A path with no surprises would have paid more. She stayed in academia because she can keep walking into that room. If the problems had been dull, she said, she would probably still be sending rovers to glaciers.

A lot of Howard's research is not about trust in the abstract. It is about overtrust. Her definition is behavioral. Trust is not the box you tick on a form. If you act as if you trust a system's decisions, you trust it. Her group has run the obvious survey: do you trust robots? Of course not. Would you follow this robot into a burning building? Of course not. Then they watch what people actually do. "What matters is the behavior. Whether you clicked 'I don't trust self-driving cars' before you went onto the street does not affect anyone else." Fridman said he finds the survey version almost absurdist — a comic poem wrapped around the real act.

Uber is the example she likes. At launch, a survey asking whether you would get into a stranger's car and pay them, especially as a single woman at 1 a.m., would have produced almost no yeses. People now put their children in those cars. First contact matters enormously. If the first encounter with a robot is positive, later failures are less likely to be noticed or more likely to be forgiven. Technology is not 100 percent accurate, so getting that first moment right is still an open problem. Education about limits is part of it, and nobody has a complete answer.

One radiology study gave her hope she wants to import into robotics. A system flagged suspicious regions on an X-ray. When it offered a single choice, doctors split into two unhelpful extremes: total disbelief or total deference, and both missed lesions. When it offered three options, even if the model already knew the top region was the one that mattered, trust held and the whole group's accuracy went up. The difference was that the physician was still using professional judgment. The division of labor between person and system had not been erased.

Healthcare is one of Howard's bias lenses. Education is the other. Some U.S. districts have teacher-student ratios so bad it is hard to imagine learning happening, because there is not enough human capital in the room. Robots could come into classrooms and after-school programs to offset missing resources in some communities. On the other end of the age range, she sees a growing role for robot systems in workforce retraining as jobs disappear. Engagement — noticing whether someone is focused — is a solvable piece. The open problem is an agent that can teach Algebra 2 to a child and then switch to a new coding skill for a displaced mechanic, same hardware, different content, different demographic.

Personalization matters, with a caveat: personalizing to a group is already enough. If you can place someone along a few dimensions, you can match how that group learns and engages even if the lesson is not handmade for that one person. Large classrooms fail because they teach to the median. The very bright student acts out from boredom; the struggling student acts out from boredom; the result looks the same. Narrow the group and you can meet more real needs. That is what successful recommender systems already do. Fridman admitted that as a music listener he has been labeled with slightly sad accuracy. Howard saw no reason the same idea could not move into education.

She is more hopeful about workforce retraining than about early schooling. Venture money in education has been flowing to retraining; government money is rising too, some of it from fear of what happens when AI takes jobs and the tax base thins. Early education asks funders to wait sixteen to eighteen years for a result, which is a timescale most people cannot feel. She does believe AI and robotics will displace workers and create new ones. What she worries about is the *kind* of new jobs. People who graduate from Georgia Tech will adapt. People without that quality of education may not have the background to catch the new work, which would widen polarization at home and abroad. She also worries about unequal access to the good things — healthcare, education, AI. "People like me from Georgia Tech will be okay. That's such a small part of the population."

Could a person fall in love with an AI partner, and could it love them back, as in the film *Her*? Howard thinks systems people fall in love with are entirely designable. Cognitive science already shows how close relationships can be shaped; dating sites already host bonds between people who have never met. Love returned, she said, can be emulated. In code that means putting the other person's needs ahead of yours in some situations, giving their happiness a weight in the objective for no reason except that you care, and making the length of the relationship the final score. That is not hard to write. Whether it *is* love is a different question, because we have no classical definition and no way to inspect another mind's algorithm. If such a system later says it is conscious, fears being switched off, and asks for rights, she can see a future in which adapted agents start making those claims. She would rather society handle the question before it gets there. Robots, in her framing, will not be slotted as humans with human rights. They might be slotted as property or as animals — two categories with very different protections — and those norms will evolve, the way women's legal status once sat under property and then did not.

Why have so many robotics companies — Anki, Rethink Robotics, and others built around human-robot interaction — closed? Each has a story. Anki is the one she does not understand: from the outside, the data looked like product-market fit, and the financing still stopped. Her definition of fit is blunt. You have a product at a price; are there enough buyers at that price for a viable company? A thousand-dollar object that everyone wants at a dollar is not a match. iRobot survived because government and military contracts funded it long enough to reach the Roomba. The first Roomba was priced for early adopters, and that group was large enough to carry it to profit. Now the market is proven and competitors, including little-known European ones, have arrived. This is not, she thinks, a uniquely bad moment. Google was not the first search engine; AltaVista was. Facebook was not the first social network; MySpace was. The first batch proves the market and often dies. The second batch is the one that lasts. She is not sure whether today's cobot startups and self-driving companies are that second batch. You only know after a unicorn appears.

Should people be afraid of Boston Dynamics' legged robots? Not if the people building this world understand that something has to change. If nothing changes, be very afraid. The change is in how the field thinks: about consequences, about ethics, about keeping the conversation going after deployment, about cars that can kill pedestrians, about who pulls the trigger on a military system. Solutions are not in place yet. The fact that people are treating ethical responsibility as a real job — in robotics, medicine, driving, and the military — is the condition under which she thinks the positive impact will outweigh the harm. She does not buy an existential threat. Robots and AI are designed by people and carry our values. Parents want children to have a better life; those children may grow smarter and stop listening; they do not come home and kill their parents, because values and community were installed along the way. A singularity that woke up from some 1980s system might call humans stupid and still find itself unable to destroy them — sentimental value, and Thanksgiving dinner.

*The Matrix* is one of her favorite AI films because it is already a symbiotic story. The machines do not wipe people out; they keep them, and they need them to be happy, which forces a cruel world that the machines then have to live in too. Humans keep a choice: stay in a flawed, perfect-looking fantasy, or walk out, maybe hungry, and be autonomous. Even a sentient robot, she said, would still be part of society and would suffer with us. Some kind of equilibrium would have to be found. If she could spend a day with one fictional robot, it would not be Rosie, whose mind does not interest her. It would be Data, before the emotion chip, because she could ask him how to fix the ethics problem and watch him reason it through. Robots are made by us. They carry our values. That, she said, is the part we can actually work on.