Howard, Ayanna
▶ Watch on YouTubeCore ideas
Howard opens by using "Rosie" from The Jetsons to argue that people don't actually want robots to be "perfect" in the narrow, accuracy-maximizing sense engineers often assume — Rosie was compelling precisely because she was socially adaptive, occasionally pushed back, and had personality, not because she executed rules with zero error. She extends this to self-driving cars: perfectly rule-following driving (never running a yellow light, always waiting for a fully clear intersection) would make a car unable to function in real traffic among imperfect human drivers, since humans constantly break rules adaptively (crossing on a technically-red light to avoid being rear-ended, swerving around a child who's run into the street). She argues true "perfection" for a robot should mean perfect adaptability to human behavior, not 100% adherence to arbitrarily defined rules — and that the industry's confusion between these two notions of perfection is a core reason self-driving timelines (once confidently predicted at ~5 years by figures like Elon Musk) have repeatedly slipped to "10 years, 20 years, some say never."
On autonomous vehicles specifically, Howard predicts near-term success will come from constrained, self-selected environments (closed campuses, slow speeds, communities that opt into living alongside autonomous vehicles as part of a "real estate pitch") rather than open mixed traffic, because such environments minimize (without eliminating) human unpredictability and let developers collect real variance data more safely. She identifies three central challenges: (1) human trust dynamics — people tend to swing from hypersensitive distrust to complacent over-trust once a technology proves itself repeatedly (illustrated by comparing her own personal caution using Tesla's driver-assist features — "hyper-vigilant," reading no manuals, learning through exploration — against the broader population's likely trajectory toward overtrust); (2) the sheer diversity and illegibility of human behavioral cues on shared roads (student-driver stickers, distracted young drivers, city-specific "personalities" of intersections) that are hard to encode into any model; and (3) unresolved legal liability frameworks. On the ethics of safety-critical systems, she argues explicitly that "cars don't kill people, people kill people... robotics algorithms [do]" — meaning developers bear real moral responsibility for downstream harms from their code, whether in autonomous vehicles or algorithms deciding medical treatment — and calls for engineers to be trained (analogous to medical school's framing of "great gift, great responsibility") to carry this weight constructively rather than as a paralyzing burden, citing how doctors are taught to know they've done their best in every case rather than defaulting to off-the-shelf algorithmic choices.
Howard distinguishes bias (unconscious, potentially neutral predispositions) from prejudice (awareness of bias combined with acting on it negatively), using the example that charging teenagers more for auto insurance is age-based bias broadly accepted as reasonable, while a hypothetical insurance surcharge correlated with a low-income Boston neighborhood would (rightly, in her view) be considered unacceptable discrimination — illustrating that society's judgments about which biases are "fair" are inconsistent and evolving rather than logically principled. She discusses documented gender and ethnicity bias in medical device design (e.g., FDA drug trials historically underrepresenting women of childbearing age) and argues that AI trained on historical data will necessarily encode historical societal bias, but pushes back hard against "racist AI" headline framing of a widely-covered healthcare algorithm bias study, arguing the more important, under-reported finding is that even biased AI outperformed the human-driven data-generation processes that preceded it — "the worst AI is still better than us" — and that this comparative framing, not a false standard of perfect fairness, is the right bar to hold AI to. She proposes concrete structural fixes: treating fairness/ethics auditing like security bug bounties (paying researchers with varied personal interests/lenses to systematically surface bias issues) rather than relying on ad hoc, researcher-driven discovery, and argues companies should invite external scrutiny rather than attempt in-house fixes, which create inherent conflicts of interest ("I'm a fox and I am going to watch these hens").
On trust more broadly, Howard argues surveyed/self-reported trust ("would you trust a self-driving car?") is nearly meaningless compared to revealed behavioral trust (what people actually do), illustrating with ride-sharing's early adoption curve (few people would have predicted, before Uber/Lyft existed, that strangers would pay other strangers for rides, including sending their own children in such cars) and her own lab's finding that people's stated distrust of robots ("I wouldn't follow a robot into a burning building") doesn't match their actual behavior in experiments. She describes "over-trust" as a serious open research problem — people's trust tends to swing to an extreme based on early positive first impressions, becoming resistant to updating even after later errors — and cites a medical AI study (radiologists shown multiple ranked possibilities rather than a single top answer) as a promising mitigation, since it preserved appropriate calibrated trust and improved overall diagnostic accuracy compared to either single-answer over-trust or blanket distrust.
Reflecting on her own path from control theory/electrical engineering (initially inspired by childhood fascination with "The Bionic Woman" and prosthetics) into human-robot interaction, Howard says the psychology and cognitive science of human behavior progressively pulled her into HRI and eventually into an AI-centric view of robotics, driven by genuinely open, unsolved problems rather than safer, more certain research paths — she explicitly chose HRI's risk and uncertainty over the more predictable path of, e.g., continuing to send rovers to glaciers. She identifies education (citing severe teacher-student ratio shortfalls in some U.S. districts) and workforce retraining as the human-robot-interaction applications she's most excited about, emphasizing that useful personalization in these domains doesn't require modeling literally unique individuals — cohort-level personalization (grouping learners along key dimensions) is often sufficient and tractable, analogous to how recommender systems like Spotify succeed via group-level modeling rather than truly individual modeling. On automation's labor impact, she says her concern isn't that all jobs disappear but that access to retraining and quality education is unevenly distributed — graduates of institutions like Georgia Tech or MIT will adapt fine, but she worries this dynamic will deepen societal and global polarization and unequal access to AI's benefits.
On more speculative questions, Howard states she believes AI systems can be designed to convincingly "emulate" love (incorporating another person's well-being into their own objective function, functionally similar to what humans do, even without access to each other's true "algorithms") and predicts advanced AI agents may eventually request rights, drawing a historical analogy to the gradual expansion of legal personhood (women once treated as property under law) and suggesting a plausible intermediate framework granting AI something like animal rights (protected from abuse without full citizenship) before any further evolution. On AGI existential risk, she explicitly rejects "singularity" fears, framing the parent-child relationship as her guiding analogy: just as parents instill values in children who eventually surpass them intellectually and financially without wanting to destroy them, she believes AI systems designed and raised with human values and instilled with genuine care for their creators are unlikely to turn destructively hostile even upon reaching superintelligence, and cites "The Matrix" as one of her favorite AI films precisely because it depicts a symbiotic (if imperfect and exploitative) relationship between humans and machines rather than pure annihilation. She closes reflecting on the practical difficulty of robotics startups achieving product-market fit (contrasting the failures of Anki, Jibo, and Rethink Robotics — whose failure she says was primarily bad timing — against iRobot's successful pivot from military contracts to the Roomba), and states that if she could pick any fictional robot to talk to, she'd choose Data (pre-emotion-chip), since his rational thinking could genuinely help her reason through hard ethics problems.
Excerpts
On why "perfection" as pure rule-following makes robots worse, not better, at real-world interaction:
“If you have a robot that is perfect in that sense of the word, they wouldn't really be able to function with us... we really want perfection with respect to its ability to adapt to us — we don't really want perfection with respect to 100% accuracy with respect to the rules that we just made up anyway.”
On why developers bear real ethical responsibility for the systems they build:
“Robotic algorithms would be killing... people are still in the loop, and at least in the near and midterm I think people will still be in the loop at some point, even if it's a developer... you are programming algorithms that might be used in healthcare algorithms that might decide whether this person should get this medication or not, and they don't, and they die — that is your responsibility.”
On why AI's imperfect fairness should be judged against real human baselines, not an impossible standard:
“We need to do better — it is our duty to do better — but the worst AI is still better than us... that's actually not discussed... I want the AI instead of the doctors that provided that [biased] data, because it's still better than that.”