12 Aug 2020
While stuck at home, I’ve done a lot of thinking on the kind of research I should do. It seems that if you want to do AI work, while preserving a sense of ethics, you can work on: AI “for social good”1, critiques of AI2, or other technical work that isn’t Evidently Evil3.
This list bothers me: is “not picking Evidently Evil projects” enough? Are we confined to only do things that are Not Bad? Wouldn’t you want to do something that’s inherently Good? I don’t know. And I’m struggling to figure that out.
I’ve come to question whether the only way to Do Good is to do AI for social good, or critical AI. Why don’t I just do that?
One answer could be the constant gaslighting of ML fairness experts who, when not outright dismissed, have to endure explanations of “AI bias” by people who decided it was an important topic mere months ago.
Maybe it’s because I’m afraid4. I want to be respected in the research field and have a stable career (key for imigration). Doing justice work means being unjustly branded as “less technical” and “controversial,” and by extension less employable5. This fucking sucks. Would I be able to get a job, or an investor buy-in if I have a reputation for “not watching my tone” because I’m speaking truth to power?
Yet, most importantly, it’s because there’s work I want to do just because I find it intellectually beautiful. I got into AI research because I genuinely want to understand how the learning process works, and whether we can artificially replicate it. Of course I’ll continue doing D&I work, but it’ll be pointless unless we, as marginalized people, are hired to work on the questions about cognition and learning that we find elegant6.
So why does it feel impossible to do anything other than critical AI without contributing to our own surveillance and oppression?7 I’d guess most of us started working in AI because of the elegance of the inquiry, not to improve classification accuracy by 1%. At some point along the way, we forgot that classification is merely a simplification created to simulate one biological ability the human brain has, and we started min-maxing classification benchmarks8.
It’s common to hear that “biological inspiration should not constrain our solutions for intelligence,” which I agree with! The hardware is different, so why shouldn’t the software? Yet, letting go of the biological inspiration gave machine learning permission to do whatever. Under capitalism, this means the original goal - understanding the learning process - was corrupted in favor or “usefulness.”
And this is the root of the problem: historically, the questions we work on have been picked by humans under capitalism9, who had an easier time not questioning them because they turned out to be useful. Today, knowing the ethical implications of these frameworks, many projects feel Evil because they feel handed down to us by this system10.
But there’s no need to work on heirloom problems. We do it because we underestimate the importance of our own ideas. In 1832, Évariste Galois, a 20-year-old mathematician widely rejected by academia at the time died in a duel over a girl. Before his fate with destiny, he had been writing about his frustrations with academia by fleshing out the ideas he found most interesting in mathematics. Once these manuscripts were discovered, his ideas turned out to yield a hugely influential new way to look at polynomial equations.
Questioning the questions of our time, and the assumptions built into them, is a key part of science. We should, as marginalized peoples, trust our guts more11. Our lived experience will lead us to questions that truly matter, as long as we seek them out. This, in itself, is a radical act.
So here’s my answer to this existential crisis: I will work on problems that are beautiful to me, and that genuinely tackle understanding how learning happens. I’m starting by writing down my own hamming problems, like Galois did. I’ll be reading more cognitive psychology literature and thinking about the similarities and discrepancies between the definitions of intelligence in humans, and those in computers. I’ll refine the questions I ask and continuously re-evaluate their alignment with my values. This way, they’ll never be merely Not Evil.
I have no plans to get into any duels, but I hope that creating a research agenda guided by my own values will be much more impactful and enriching than any SOTA-chasing ever could.
Thanks to Katherine Lee, Natalia Bilenko, Ria Kalluri, Keren Gu, Ben Poole, Joshua Morton, and William Agnew for thoughtful comments and discussion.
This is usually applied work. Often on things that are perceived by many as uncontroversial, to the point where they are often done as a PR strategy. This results in limited impact. “Good” isn’t good enough. ↩
Works on ML fairness, accountability, transparency and justice fall in this category. It can be theoretical (some great pieces out there read as pure philosophy), or applied (often demonstrations of how biased/abusable AI systems are, and how they perpetuate systemic injustices). Both are very technical, and heavily shape policy. ↩
There are a lot of good strategies you can adopt today to have your work be Not Evil:
Tons of justice work is extremely technical. But, much like HCI and design, it is perceived as inferior to those doing “true technical work,” such as fucking facial recognition… ↩
As you might have guessed by now, the feelings in this post started when research work became massively silly to work on, due to the current mismanaging of the pandemic, immigration restrictions, and the oppression of Black Lives Matter protests. ↩
“As I’ve dived into research wrt billions of affected people & massive systems, I’ve found it harder to review for/enjoy academic conferences. Many papers have the same issue: a sol’n that misses the bigger picture using datasets curated by colleagues also missing the bigger picture. Cynically, it seems like most papers have their main goal as simply putting something out towards a PhD, or aligning to grant funding to keep the people involved employed.” is another gem, which you can read here. ↩
“When people talk about ‘unintended consequences,’ says @png_marie, it sounds like they’re saying the consequences couldn’t have been predicted. But AI’s unintended consequences are in fact highly predictable if you just look back at history.” Read more here. ↩
“It seems that the influence of your teacher has been to give you a false idea of what are worthwhile problems. The worthwhile problems are the ones you can really solve or help solve, the ones you can really contribute something to. (…) I would advise you to take even simpler, or as you say, humbler, problems until you find some you can really solve easily, no matter how trivial. (…) You must not take away from yourself these pleasures because you have some erroneous idea of what is worthwhile. (…) No problem is too small or too trivial if we can really do something about it.” Read more here. ↩