On Wednesday evening, I was the female participant on a panel about artificial intelligence. The event was hosted at the National Club on Bay Street in Toronto. At Friday’s lunch, a colleague who attended to support mentioned that the venue smelled like New York, carried the grime of time in its walls after so many rain storms. Indeed, upon entering, I rewalked into Manhattan’s Downtown Association, returned to April, 2017, before the move to Toronto, peered down from the attic of my consciousness to see myself gently placing a dripping umbrella in the back of the bulbous cloak room, where no one would find it, feeling the mahogany enclose me in peaty darkness, inhaling a mild must that could only tolerate a cabernet, waiting with my acrylic green silk scarf from Hong Kong draped nonchalant around my neck, hanging just above the bottom seam of my silk tunic, dangling more than just above the top seam of my black leather boots, when a man walked up, the manager, and beaming with welcome he said “you must be the salsa instructor! Come, the class is on the third floor!” I laughed out loud. Alfred arrived. Alfred who was made for another epoch, who is Smith in our Hume-Smith friendship, fit for the ages, Alfred who had become a member of the association and, a gentleman of yore, would take breakfast there before work, Acshenbach in Venice, tidily wiping a moist remnant of scrambled eggs from the right corner lip, a gesture chiseled by Joseon porcelain and Ithaca’s firefly summer, where took his time to ruminate about his future, having left, again, his past.
Upstairs we did the microphone dance, fumbling to hook the clip on my black jeans (one of the rare occasions where I was wearing pants). One of my father’s former colleagues gave the keynote. He walked through the long history of artificial intelligence, starting with efforts to encode formal logic and migrating through the sine curve undulations of research moving from top-down intelligent design (e.g., expert systems) to bottom-up algorithms (e.g., deep convolutional neural networks), abstraction moving ever closer to data until it fuses, meat on a bone, into inference. He proposed that intellectual property had shifted from owning the code to building the information asset. He hinted at a thesis I am working to articulate in my forthcoming book about how contemporary, machine learning-based AI refracts humanity through the convex (or concave or distorted or whatever shape it ends up being) mirror of the space of observation we create with our mechanisms for data capture (which are becoming increasingly capacious with video and Alexa in every home, as opposed to being truncated to bilps in clickstream behavior or point of sale transactions), our measurement protocol, and the arabesque inversions of our algorithms. They key thing is that we no longer start with an Aristotelian formal cause when we design computational systems, which means, we no longer imagine the abstract, Platonic scaffold of some act of intelligence as a pre-condition of modeling it. Instead, as Andrei Karpathy does a good job articulating, we stipulate the conditions for a system to learn bottom-up from the data (this does not mean we don’t design, it’s just that the questions we ask as we make the systems require a different kind of abstraction that is affiliated with induction (as Peter Sweeney eloquently illustrates in this post)). This has pretty massive consequences for how we think about the relationship between man and machine. We need to stop pitting machine against man. And we need to stop spouting obsequious platitudes that the “real power comes from the collaboration of man and machine.” There’s something of a sham humanism in those phrases that I want to get to the bottom of. The output of a machine learning algorithm always already is, and becomes even more, as the flesh of abstraction moves closer to the bone of data (or vice versa?), the digested and ruminated and stomach acid-soaked replication of human activity and behavior. It’s about how we regurgitate. That’s why it does indeed make sense to think about bias in machine learning as the laundering of human prejudice.
A woman in the audience posed the final question to the panelists: you’ve spoken about the narrow capabilities of machine learning systems, but will it be possible for artificial intelligence to learn empathy?
A fellow panelist took the Turing Test approach: why yes, he said, there has been remarkable progress in mimicking even this sacred hallmark of the limbic system. It doesn’t matter if the machine doesn’t actually feel anything. What matters it that the machine manifests the signals of having felt something, and that may well be all that matters to foster emotional intelligence. He didn’t mention Soul Machines, a New Zealand-based startup making “incredibly life-like, emotionally responsive artificial humans with personality and character,” but that’s who I’d cite as the most sophisticated example of what UX/UI design can look like when you fuse the skill set of cinematic avatars, machine learning scientists, and neuroscientists (and even the voice of Cate Blanchett).
I disagreed. I am no affect expert (just a curious generalist fumbling my way through life), but believe empathy is remarkably complex for many reasons.
I looked at her directly, deeply. At not just at her, I looked into her. And what I mean by looking into her is that I opened myself up a little, wasn’t just a person protected by the distance of the stage (or, more precisely, the 4 brown leather bar stools with backs so low they only came up to vertebra 4 or 5, and all of us leaned in and out trying to find and keep a dignified posture, hands crossed into serenity, sometimes leaning forward). Yes, when I opened myself to engage with her I leaned forward almost to the point of resting my elbows on my thighs, no longer leaning back and, every few moments, returning my attention to the outer crevices of my eyes to ensure they were soft as my fellow panelists spoke. And I said, think about this. I’m up here on stage perceiving what I’m perceiving and thinking what I’m thinking and feeling what I’m feeling, and somehow, miraculously, I can project what I think you’re perceiving, what I think you’re thinking, what I think you’re feeling, and then, on top of that, I can perhaps, maybe, possibly start to feel what you feel as a result of the act of thinking that I think what you perceive, think, and feel. But even this model is false. It’s too isolated. For we’ve connected a little, I’m really looking at you, watching your eyes gain light as I speak, watching your head nod and your hands flit a little with excitement, and as I do this we’re coming together a little, entangling ourselves to become, at least for this moment, a new conjoint person that has opened a space for us to jointly perceive, think, and feel. We’re communicating. And perhaps it’s there, in that entangled space, where the fusion of true empathy takes place, where it’s sound enough to impact each of us, real enough to enable us to notice a change in what we feel inside, a change occasioned by connection and shared experience.
A emotional Turing test would be a person’s projection that another being is feeling with her. It wouldn’t be entangled. It would be isolated. That can’t be empathy. It’s not worthy of the word.
But, how could we know that two people actually feel the same feeling? If we’re going to be serious, let’s be serious. Let’s impose a constraint and say that empathy isn’t just about feeling some feeling when you infer that another person is feeling something, most often feeling something that would cause pain. It’s literally feeling the same thing. Again, I’m just a curious generalist, but know that psychologists have tools to observe areas of the brain that light up when some emotional experience takes place; so we could see if, during an act of empathy, the same spot lights up. Phenomenologically, however, that is, as the perceived, subjective experience of the feeling, it has to be basically impossible for us to ever feel the exact same feeling. Go back to the beginning of this blog post. When I walked into the National Club, my internal experience was that of walking into the Downtown Association more than 1.5 years earlier. I would hazard that no one else felt that, no one else’s emotional landscape for the rest of the evening was then subtly impacted by the emotions that arose during this reliving. So, no matter how close we come to feeling with someone when our emotional world us usurped, suddenly, by the experience of another, it’s still grafted upon and filtered through the lens of time, of the various prior experiences we’ve had that trigger that response and come to shape it. As I write, I am transported back to two occasions in my early twenties when I held my lovers in my arms, comforting and soothing them after each had learned about a friend’s suicide. We shared emotion. Deeply. But it was not empathy. My experience of their friends’ suicide was far removed. It was compassion, sympathy, but close enough to the bone to provide them space to cry.
So then we ask, if it’s likely impossible to feel the exact same feeling, then we should relax the constraint and permit that empathy need not be deterministic and exact, but can be recognized within a broader range. We can make it a probabilistic shared experience, an overlap within a different bound. If we relax that constraint, then can we permit a Turing test?
I still don’t think so. Unless we’re ok with sociopaths.
But how about this one. Once I was running down Junipero Serra Boulevard in Palo Alto. It was a dewy morning, dewy as so many mornings are in Silicon Valley. The rhythms of the summer are so constant: one wakes up to fog, daily, fog coming thick over the mountains from the Pacific. Eventually the fog burns and if you go on a bike ride down Page Mill road past the Sand Hill exit to 280 you can watch how the world comes to life in the sun, reveals itself like Michelangelo reveals form from marble. There was a pocket of colder, denser, sweeter smelling air on the 6.5-mile run I’d take from the apartment in Menlo Park through campus and back up Junipero Serra. I would anticipate it as I ran and was always delighted by the smell that hit me; it was the smell of hose water when I was a child. And then I saw a deer lying on the side of the road. She was huge. Her left foot shook in pain. Her eyes were pleading in fear. She begged as she looked at me, begged for mercy, begged for feeling. I was overcome by empathy. I stopped and stood there, still, feeling with her for a moment before I slowly walked closer. Her foot twitched more rapidly with the wince of fear. But as I put my hand on her huge, hot, sweating belly, she settled. Her eyes relaxed. She was calmed and could allow her pain without the additional fear of further hurt. I believe we shared the same feeling at that moment. Perhaps I choose to believe that, if only because it is beautiful.
The moment of connection only lasted a few minutes, although it was so deep it felt like hours. It was ruptured by men in a truck. They honked and told me I was an idiot and would get hurt. The deer was startled enough to jump up and limp into the woods to protect herself. I told the men their assumptions were wrong and ran home.
You might say that this is textbook Turing test empathy. If I can project that I felt the exact same feeling as an animal, if I can be that deluded, then what’s stopping us from saying that that the projection and perception of shared feeling is precisely what this is all about, and therefore it’s fair game to experience the same with a machine?
The sensation of love I felt with that deer left a lasting impression on me. We were together. I helped her. And she helped me by allowing me to help her. Would we keep the same traces of connection from machines? Should empathy, then, be defined by its durability? By the fact that, if we truly do connect, it changes us enough to stay put and be relived?
There are, of course, moments when empathy breaks down.
Consider breakdowns in communication at work or in intimate relationships. Just as my memory of the Downtown Association shaped, however slightly, my experience at Wednesday’s conference, so too do the accumulated interactions we have with our colleagues and partners reinforce models of what we think others think about us (and vice versa). These mental models then supervene upon the act of imagination to perceive, think, and feel like someone else. It breaks. Or, at the very latest, distorts the hyperparameters of what we can perceive. Should anything be genuinely shared in such a tangled web, it would be the shared awareness of the impossibility of identification. I’ve seen this happen with teams and seen it happen with partners. Ruts and little walls that, once built, are very difficult to erode.
Another that comes to mind is the effort required to empathize deeply with people far away from where we live and what we experience. When I was in high school, Martha Nussbaum, a philosopher at the University of Chicago who has written extensively about affect, came and gave a talk about the moral failings of our imagination. This was in 2002. I call her mentioning that we obsess far more deeply, we feel far more acutely, about a paper cut on our index finger or a blister on our right heel, than we do when we try to experience, right here and now, the pain of Rwandans during the genocide, of Syrian refugees packed damp on boats, of the countless people in North America razed from fentanyl. On the talk circuit for his latest book, Yuval Harari comments that we’ve done the conceptual work required to construct and experience a common identity (and perhaps some sort of communal empathy) with people we’ll never meet, who are far outside the local perception of the tribe, in constructing the nation. And that this step from observable, local community to imagined, national community was a far steeper step function than the next rung in the ladder from national to global identity (8,000,000 and 7,000,000,000 are more or less the same for the measly human imagination, whereas 8,000,000 feels a lot different than 20). Getting precise on the limits of these abstractions feels like worthwhile work for a 21st-century ethicists. After all, in its original guise, the trolley problem was not a deontological tool for us to pre-ponder and encode utilitarian values into autonomous vehicles. It was a thinking tool to illustrate the moral inevitability of presence.
I received LinkedIn invites after the talk. One man commented that he found my thoughts about empathy particularly insightful. I accepted his invitation because he took the time to listen and let me know my commentary had at least a modicum of value. I’ll never know what he felt as he sat in the audience during the panel. I barely know what I felt, as two and a half days of experience have already intervened to reshape the experience. So we grow, beings in time.
 Loyal blog readers will have undoubtedly noticed how many posts open with a similar sentence. I speak at a ton of conferences. I enjoy it: it’s the teacher’s instinct. As I write today, however, I feel alienated from the posts’ algorithmic repetition, betokening the rhythm of my existence. Weeks punctuated by the sharp staccato of Monday’s 15-minute (fat fully cut) checkins, the apportioned two hours to rewrite the sales narrative, the public appearances that can be given the space to dilate, and the perturbations flitting from interaction to interaction, as I gradually cultivate the restraint to clip empathy and guard my inside from noxious inputs. Tuesday morning, a mentor sent me this:
 This is a loaded term. I’m using it here as a Bayesian would, but won’t take the time to unpack the nuances in this post. I interviewed Peter Wang for the In Context podcast yesterday (slated to go live next week) and we spoke about the deep transformation of the concept of “software” we’re experiencing as the abstraction layer that commands computers to perform operations moves ever closer to the data. Another In Context guest, David Duvenaud, is allergic to the irresponsible use of the word “inference” in the machine learning community (here’s his interview). Many people use inference to refer to a prediction made by a trained algorithm on new data it was not trained on: so, for example, if you make a machine learning system that classifies cats and dogs, the training stage is when you show the machine many examples of images with labels cat and dog and the “inference” stage is when you show the machine a new picture without a label and ask it, “is this a cat or a dog?” Bayesians like Duvenaud (I think it’s accurate to refer to him that way…) reserve the term inference for the act of updating the probability of a hypothesis in light of new observations and data. Both cases imply the delicate dance of generalization and induction, but imply it in different ways. Duvenaud’s concern is that by using the word imprecisely, we lose the nuance and therefore our ability to communicate meaningfully and therefore hamper research and beauty.
 Franco Moretti once told me that similar areas of the brain light up when people read Finnegans Wake (or was it Ulysses? or was it Portrait of the Artist? and the Bible (maybe Ecclesiastes?).
The featured image is Edouard Manet’s Olympia, unveiled in Paris in 1856. In the context of this post, it illustrates the utter impossibility of our empathizing with Olympia. The scorn and contempt in her eyes protects her and gives her power. She thwarts any attempt at possession through observation and desire, perhaps because she is so distanced from the maid offering her flowers, deflecting her gaze out towards the observer but looking askance, protecting within her the intimations of what she has just experienced, of the fact that there was a real lover but it was and will never be you. Manet cites Titian’s Venus of Urbino (1534), but blocks all avenues for empathy and connection, empowering Olympia through her distance.