Read on Twitter
1\\ A Case for Ethical AI Optimism:
A lot of people are hurting right now. Everyone needs to hear something different. If this thread is going to feel like toxic positivity, I do not want to shove that on you. Please ignore.
A LONG thread…
A lot of people are hurting right now. Everyone needs to hear something different. If this thread is going to feel like toxic positivity, I do not want to shove that on you. Please ignore.
A LONG thread…
2\\ You all know of @jovialjoy and @timnitGebru's landmark Gender Shades paper which pretty much cracked open the field of AI Ethics to the broader public. I "contributed" to that paper, and not in a good way. I was a contributor to the underlying problem.
3\\ IBM Watson’s Visual Recognition was one of the services tried and found severely wanting in its ability to detect diverse faces, especially Black women.
4\\ The genesis of IBM’s face detection service was its acquisition of AlchemyAPI, a small Denver company picked up in 2015 for NLP, vision, and other AI tech.
5\\ I, along with @phdharp and a couple folks not on Twitter, was part of a small team that built the original technology in 2014, much of which was still in use in early 2018.
6\\ And it was bad at detecting diverse faces, especially Black women! And of course part of that was the training data. But ultimately, it’s because we were careless. We kicked the tires a bit on diverse faces with nothing remotely approaching rigor, and it seemed ok.
7\\ It’s not like ethics weren't front of mind. @eturner303, our CEO, is an incredibly ethical person, maybe the most ethical business leader I know. We focused on use case ethics. We knew we'd never allow ourselves to be involved in surveillance, for example (and never were).
8\\ So we put it out. We didn’t have tons of traction; use cases for facial detection without recognition of an individual weren’t easy, so the problem stayed quiet.
9\\ And then we got acquired, and for those of you in the know, that is the frantic lens through which everything is viewed for the next couple of years. You get pulled in a million directions every day.
10\\ By the time Gender Shades came out, my team had long moved on to other projects. So even though @IBMWatson responded in *days* after that paper with a system that reduced the bias ten-fold, I can claim precisely zero credit for any of it. https://www.ibm.com/blogs/research/2018/02/mitigating-bias-ai-models/
11\\ The incredible speed was possible due not only to reprioritization but also to the fact an @IBMResearch team had been diligently working on a new model to solve that precise problem for nine months. I wasn’t even aware, I just heard it after-the-fact from @ruchir_puri.
12\\ Two years later, IBM was also the first major company to entirely pull out of facial recognition (detecting individuals). https://fortune.com/2020/06/09/george-floyd-ibm-exits-facial-recognition-bias-human-rights/
13\\ Even though I’ve moved on from IBM Watson, I’m proud of their decisions here. But I’m not naive. What drives these responsible decisions? It’s knowledge and it’s social incentives.
14\\ I can tell you that *down to the last man and woman*, every person in AI that I know personally cares a ton about ethics. But people will stumble if they haven’t considered what’s right.
15\\ And huge companies will only walk away from money on the table if there is at least some sort of pressure, be it PR or regulatory.
16\\ We all know the path forward is more complicated than “don’t build things”. We need education so that people start off on the right foot and avoid mistakes.
17\\ Perhaps even more, we need incentives for those who admit their mistakes *and fix them*, because it’s hard to choke down pride (humans) or give up money (businesses).
18\\ But if we do it right…you know human bias injected into AI is not a foregone conclusion, right? That means it is technically possible (while technically and socially difficult) to build AI that is MORE ETHICAL THAN HUMANS.
19\\ That’s the endgame. Not “not worse”. Better. Much better.
20\\ And getting there will take a million imperfect steps. AI replaces tasks. Tasks done today by humans, programs, or other AI, all of which are almost certainly already very biased. So I myself am strongly biased in favor of putting things out…IF it’s a step forward, not back.
21\\ @timnitGebru, you don’t need me to tell you, but what you do matters. I’ve watched it matter. Having lived the pervasive fire drill of an IBM-sized company (much love, IBM gang :P), it's a powerful force indeed that cuts through long enough to help us remember what’s right.
22\\ @JeffDean, it's not big coming from me, but your choices and the choices of @GoogleAI leaders matter. I believe the world is far better because of Google. And since there's no natural law that says this must be so, it will only continue based on the choices of its leadership.
23\\ I’m not here to judge anybody. I believe only God can judge what is in a person’s heart. And I’m an entrenched atheist, but the math still checks out!
24\\ Since I can only speak for myself, then, I’ll say that I am proud of how the work I am doing is continually becoming a better reflection of my values.
25\\ So, to anyone who made it to the end: if hope is what you need right now, I hope you find it. And if it’s not, I hope you find what you do need. Because we need all of you in order to make AI a force for global good.
Yes, I’m talking to you.
Yes, I’m talking to you.
