Making AI fair | Osonde Osoba | TEDxManhattanBeach — Transcript

Full transcript

1. 0:14So, when I was a teenager growing up in
2. 0:17Nigeria,
3. 0:19I was not directing music videos.
4. 0:32I did decide however to build an
5. 0:34artificial intelligence.
6. 0:38I know. I know.
7. 0:41I was taking computer science and logic
8. 0:43courses at the local university, and I
9. 0:45figured it would be easy being a
10. 0:4715-year-old to just create an AI that
11. 0:50could reason and argue logically.
12. 0:56I was inspired by the same dreams the
13. 0:58early pioneers of artificial
14. 0:59intelligence had. That we could we could
15. 1:02create this supermind, an intelligence
16. 1:05of pure logic and objectivity. And that
17. 1:07this supermind would be free of human
18. 1:10traits, traits like subjectivity and
19. 1:12bias.
20. 1:16I was working without guidance with
21. 1:19really, really old computers
22. 1:21and with irregular electricity. If
23. 1:24you've been to Nigeria, you might know
24. 1:25what that's about.
25. 1:27So, let's just say I was less than
26. 1:30successful.
27. 1:32I did however go on to study in the
28. 1:34United States
29. 1:35and do a lot more research as an
30. 1:37engineer on artificial intelligence.
31. 1:43Around 2013, I was reminded of my dreams
32. 1:46of objective AI when I heard of a
33. 1:49Harvard professor's experiments.
34. 1:52Latanya Sweeney at the university at
35. 1:54Harvard University was exploring how
36. 1:57search engine ads change when you look
37. 1:59up different types of names.
38. 2:03When she looked up names more predictive
39. 2:06of being black,
40. 2:08she got ads for criminal justice
41. 2:11services.
42. 2:14So, she'd look up names like Deshawn or
43. 2:17even Latanya, and she'd get ads like,
44. 2:20"Has Deshawn been arrested?"
45. 2:24It's not funny, but it's kind of funny.
46. 2:26"Do you need a bail bondsman?"
47. 2:30When she looked up names less predictive
48. 2:32of being black, so names like Jeffrey or
49. 2:36Emma,
50. 2:37not so much.
51. 2:41Eventually, she calculated that
52. 2:42black-identifying names were 25%
53. 2:45more likely to generate ads suggesting
54. 2:48that the person had an arrest record,
55. 2:51even when they did not.
56. 2:54But, these are algorithms serving up
57. 2:56ads.
58. 2:58Can algorithms be prejudiced?
59. 3:01Can a piece of software be racist? Can
60. 3:04it be sexist?
61. 3:09Learning algorithms are recipes for
62. 3:12teaching AI how to learn.
63. 3:14Modern AI learns by consuming vast
64. 3:16amounts of data. They learn by spotting
65. 3:20hidden correlations and patterns that
66. 3:21humans cannot see. They learn by
67. 3:24applying neutral math equations.
68. 3:27But, the more researchers have looked,
69. 3:30the more they've found examples of
70. 3:32artificial intelligence systems
71. 3:34producing significantly biased outcomes.
72. 3:38And by bias, we mean generally
73. 3:41violations of ethical or social norms.
74. 3:44For example,
75. 3:46a recent study found that search engines
76. 3:49show more ads for higher-paying jobs to
77. 3:52men than to women.
78. 3:55They've also been There was a recent
79. 3:57study last year showing extensive bias
80. 4:00in algorithms used in the criminal
81. 4:02justice system.
82. 4:06So, just using AI does not eliminate
83. 4:09bias in in decision-making.
84. 4:12But, how exactly do these biases get
85. 4:14into our algorithms in the first place?
86. 4:19Modern AI
87. 4:20is only as good as the data on which it
88. 4:23is trained.
89. 4:24AI eats
90. 4:26and processes data, and uses the
91. 4:28patterns it finds in the data to make
92. 4:30future decisions. Now, these decisions
93. 4:32could be as trivial as distinguishing
94. 4:34between cats versus dogs.
95. 4:36Or, they could be extremely important.
96. 4:39An insurance company might use
97. 4:40artificial intelligence to help it
98. 4:42decide who is insurable versus who is
99. 4:45not insurable.
100. 4:48Consider the case of an artificial
101. 4:49intelligence trying to find the best
102. 4:52nurses.
103. 4:54If we decide to feed this system
104. 4:56stereotypical training data, so data
105. 4:58consisting mostly of female nursing
106. 5:01profiles,
107. 5:02then chances are the system would more
108. 5:05confidently judge future female
109. 5:06candidates as better fits,
110. 5:09unless it's carefully designed not to do
111. 5:11so.
112. 5:13The system isn't being inherently
113. 5:15sexist. It's simply learning the biases
114. 5:18present in our data, and applying them
115. 5:20more consistently into the future.
116. 5:24Now, occasionally,
117. 5:27AI develops a case of what we might call
118. 5:29uh food poisoning.
119. 5:32Back when IBM was training up the future
120. 5:35AI Jeopardy champion, that's Watson,
121. 5:38they had to feed Watson vast amounts of
122. 5:40data.
123. 5:42And then somebody decided it might be a
124. 5:44good idea to
125. 5:46feed Watson the Urban Dictionary.
126. 5:52So, after ingesting all that
127. 5:56let's just say colorful data,
128. 5:58Watson developed a little bit of a
129. 6:00swearing habit.
130. 6:04He began in inserting R-rated
131. 6:07four-letter words in his responses.
132. 6:10Now, personally, I would have loved to
133. 6:11watch that version of Watson
134. 6:13on Jeopardy.
135. 6:17But, the moral here is
136. 6:19we need to be careful what we feed our
137. 6:22algorithms and our AI.
138. 6:24And the thing is
139. 6:26AI is not going away.
140. 6:28AI already determines what you see on
141. 6:30your Twitter feed, what you might want
142. 6:32to watch on your on Netflix, who you
143. 6:34might want to date.
144. 6:37And it's going to keep making more and
145. 6:38more of these decisions in the future
146. 6:40because the amount of data we have and
147. 6:43we create is vast.
148. 6:45And it just keeps growing.
149. 6:48Artificial intelligence and algorithms
150. 6:50present the only viable way of making
151. 6:52sense of this much data.
152. 6:56So, if you want better medical
153. 6:58diagnosis, if you want fewer car
154. 7:00crashes, or if you want to improve the
155. 7:01quality of human lives and prevent
156. 7:03needless suffering more generally,
157. 7:05you're going to need AI systems that can
158. 7:07learn quickly, tackle complicated
159. 7:09problems, and make or inform hard
160. 7:11decisions.
161. 7:13But, we also want these systems to play
162. 7:15fair.
163. 7:18So, it's great that we're beginning to
164. 7:19get behind or at least discard this
165. 7:21illusion
166. 7:22that AI, modern AI, is objective or
167. 7:25infallible, especially now as AI is
168. 7:28ascendant.
169. 7:31And we can all get behind this idea that
170. 7:33we do not want AI to make decisions
171. 7:35that run contrary to our values.
172. 7:39But, how exactly do we teach artificial
173. 7:42intelligence to abide by social or
174. 7:45ethical norms?
175. 7:47And who gets to decide what norms? Who
176. 7:49gets to decide what is fair?
177. 7:54To me, these are fascinating questions.
178. 7:57And researchers have started looking at
179. 7:58a few approaches to tackling these
180. 8:00questions.
181. 8:02So, first,
182. 8:04we could disclose to consumers when AI
183. 8:07is making decisions that affect them.
184. 8:11Top-down government regulation may not
185. 8:13be desirable or even feasible or
186. 8:15effective as a way to tackling bias in
187. 8:17algorithms.
188. 8:19Making consumers aware
189. 8:21is an easy first step.
190. 8:25This also makes us more aware of the
191. 8:27thousand little ways in which artificial
192. 8:29intelligence affects our lives and more
193. 8:31aware of our choices as consumers of
194. 8:34services that rely on automated
195. 8:37decisions.
196. 8:39Second,
197. 8:40we could create processes for consumers
198. 8:42to appeal automated decisions.
199. 8:46They'd be appealing to humans, of
200. 8:47course.
201. 8:50Third,
202. 8:51we could have more human oversight when
203. 8:54artificial intelligence is making
204. 8:55decisions in high-risk domains. This
205. 8:58includes domains like uh
206. 9:00defense or the criminal justice system.
207. 9:04These have been policy solutions so far.
208. 9:07There are a couple of technical fixes in
209. 9:09the works.
210. 9:10For example, we've been exploring ways
211. 9:12of making AI models more transparent,
212. 9:15more observable.
213. 9:17Other researchers have been exploring
214. 9:19ways of making AI models explain their
215. 9:23decisions.
216. 9:25The idea behind these approaches, these
217. 9:28technical approaches, is to create a
218. 9:30trail of breadcrumbs
219. 9:32so that we can follow the trail of what
220. 9:34we might call intermediate inferences
221. 9:36forward from the inputs all the way
222. 9:38through to the final decisions.
223. 9:40Basically, show your work. Or we can
224. 9:42follow the decisions the trail backward
225. 9:44from the decisions all the way down to
226. 9:46underlying causes or reasons.
227. 9:51These solutions
228. 9:53they help us see where AI bias exists,
229. 9:57but they don't always fix them.
230. 10:00And more fundamentally, don't get to the
231. 10:01more fundamental question of what
232. 10:04qualifies as bias in specific domains.
233. 10:11As an engineer,
234. 10:13I fully own my tendency to think of or
235. 10:17focus on technical virtuosity as the way
236. 10:21to tackle and solve problems,
237. 10:23just as I tried to do as a teenager all
238. 10:25the way back then.
239. 10:29But
240. 10:30I'm not
241. 10:32convinced that we can just engineer our
242. 10:35way out of this particular problem.
243. 10:37Because the problem is less about the
244. 10:39technology AI, it's more about all the
245. 10:43social and cultural context in which we
246. 10:45apply the technology.
247. 10:47At the moment, we have AI being applied
248. 10:49in fields as diverse as medicine, law,
249. 10:51criminology, and even education. Each of
250. 10:54these fields have different norms,
251. 10:56different ethics.
252. 10:57And AI is like a very smart child. It's
253. 10:59not always clear on the context of its
254. 11:01decisions.
255. 11:03For example,
256. 11:05in many states, it is perfectly legal
257. 11:08to discriminate on the basis of
258. 11:10demographic characteristics
259. 11:13when you're setting auto insurance
260. 11:15rates.
261. 11:17Younger male drivers
262. 11:18tend to be riskier to insure, at least
263. 11:21so I'm told.
264. 11:24But the ability to calibrate this kind
265. 11:26of risk is crucial for insurance.
266. 11:30On the other hand, it's illegal to
267. 11:32discriminate on the basis of demographic
268. 11:34characteristics when you're issuing
269. 11:36mortgages.
270. 11:38Context determines the relevant norms.
271. 11:41And engineers cannot always teach AI how
272. 11:44it ought to behave in every single
273. 11:48application.
274. 11:49We need to work closer with social
275. 11:51scientists, so lawyers, economists,
276. 11:54anthropologists, linguists. And I can
277. 11:56tell you from personal experience,
278. 11:58working across these different fields is
279. 12:00difficult.
280. 12:02We are trained differently,
281. 12:03and this is not a joke. We really do
282. 12:05speak different languages.
283. 12:09At my day job at the RAND Corporation,
284. 12:12we try to use
285. 12:14as an example, we try to use artificial
286. 12:15intelligence models to study behaviors
287. 12:19on social networks.
288. 12:20This includes behaviors like
289. 12:22radicalization, political polarization,
290. 12:25and even cyberbullying.
291. 12:28What you might consider fair or
292. 12:31normative in any of those applications
293. 12:33will depend on what you're focused on.
294. 12:35If you're focused on radicalization,
295. 12:38then you might not think it's fair that
296. 12:40we treat every opinion expressed on
297. 12:42social media exactly alike.
298. 12:48Here's the crux of the problem.
299. 12:51We have created artificial intelligence
300. 12:53in our own image,
301. 12:55and this is not a compliment.
302. 12:59AI too easily reflects human
303. 13:01shortcomings it finds in our data
304. 13:03streams.
305. 13:04It amplifies our flaws.
306. 13:08It's going to take the full range of
307. 13:10human intelligence to correct for these
308. 13:12for these shortcomings.
309. 13:15But if you want to build better, fairer
310. 13:17societies,
311. 13:19we need AI systems that reflect and
312. 13:22amplify the better parts of our nature.
313. 13:25Thank you.