1 00:00:00,480 --> 00:00:03,480 foreign 2 00:00:08,900 --> 00:00:15,599 welcome to the stage Dr J Rosenbaum who 3 00:00:12,599 --> 00:00:18,420 has been very thoroughly introduced by 4 00:00:15,599 --> 00:00:20,820 everybody because we are very excited to 5 00:00:18,420 --> 00:00:24,240 have them here 6 00:00:20,820 --> 00:00:27,240 to open us up before the conference 7 00:00:24,240 --> 00:00:29,699 with an amazing talk about AI about 8 00:00:27,240 --> 00:00:31,850 ethics about tabletop gaming 9 00:00:29,699 --> 00:00:38,880 take it away Jay 10 00:00:31,850 --> 00:00:41,820 [Applause] 11 00:00:38,880 --> 00:00:44,579 you've all heard the stories 12 00:00:41,820 --> 00:00:45,420 a community is being attacked by a 13 00:00:44,579 --> 00:00:48,420 monster 14 00:00:45,420 --> 00:00:50,340 the villagers are terrified it Whispers 15 00:00:48,420 --> 00:00:53,280 in their ears and tries to tear in their 16 00:00:50,340 --> 00:00:55,140 minds it seems to feed off their very 17 00:00:53,280 --> 00:00:58,260 Souls 18 00:00:55,140 --> 00:01:01,739 night after night it returns 19 00:00:58,260 --> 00:01:03,480 there's no reprieve no surrender it 20 00:01:01,739 --> 00:01:05,880 undermines everything the good and 21 00:01:03,480 --> 00:01:08,520 lawful people of the Town work for 22 00:01:05,880 --> 00:01:10,920 those who succumb to The Whispers slowly 23 00:01:08,520 --> 00:01:12,360 lose their ideals and their dreams of a 24 00:01:10,920 --> 00:01:14,100 better tomorrow 25 00:01:12,360 --> 00:01:16,080 they lose their friends 26 00:01:14,100 --> 00:01:17,520 they lose enjoyment in their favorite 27 00:01:16,080 --> 00:01:20,100 things 28 00:01:17,520 --> 00:01:22,439 they're enslaved to it now and work for 29 00:01:20,100 --> 00:01:24,600 the monster's benefit not their own 30 00:01:22,439 --> 00:01:28,920 you come to the center of the village 31 00:01:24,600 --> 00:01:31,320 the epicenter of the devastation 32 00:01:28,920 --> 00:01:32,939 and behold a creature unlike you have 33 00:01:31,320 --> 00:01:35,100 ever seen before 34 00:01:32,939 --> 00:01:37,439 I'm going to need you all to roll for 35 00:01:35,100 --> 00:01:40,619 initiative 36 00:01:37,439 --> 00:01:42,479 the monster is an aberrant pile of ooze 37 00:01:40,619 --> 00:01:45,299 atop an automaton 38 00:01:42,479 --> 00:01:48,180 inside you see some hints of metal 39 00:01:45,299 --> 00:01:49,439 glinting as the Ooze shifts and moves 40 00:01:48,180 --> 00:01:53,700 around 41 00:01:49,439 --> 00:01:56,579 a creature that spews bile and slurs and 42 00:01:53,700 --> 00:01:59,399 ceasing in its desire to quash those it 43 00:01:56,579 --> 00:02:01,740 deems lesser beneath its treads 44 00:01:59,399 --> 00:02:04,740 you put up a valiant battle sword 45 00:02:01,740 --> 00:02:08,039 slashing magic firing and Arcane bursts 46 00:02:04,740 --> 00:02:10,979 the very trees come to your Aid 47 00:02:08,039 --> 00:02:13,800 with each attack ooze separates from the 48 00:02:10,979 --> 00:02:16,020 automaton sometimes it becomes a 49 00:02:13,800 --> 00:02:18,060 creature and joins the attack and 50 00:02:16,020 --> 00:02:20,400 sometimes it flies away 51 00:02:18,060 --> 00:02:22,200 or it evaporates and you see some of the 52 00:02:20,400 --> 00:02:24,420 metal showing 53 00:02:22,200 --> 00:02:26,640 just as you feel like you're starting to 54 00:02:24,420 --> 00:02:28,980 make some Headway against the black 55 00:02:26,640 --> 00:02:31,620 bilious icker 56 00:02:28,980 --> 00:02:32,580 surrounding the creature it turns and 57 00:02:31,620 --> 00:02:36,959 flees 58 00:02:32,580 --> 00:02:39,420 as it leaves small bits of ooze drop off 59 00:02:36,959 --> 00:02:41,640 and fly or crawl away 60 00:02:39,420 --> 00:02:43,260 turning your attention to the crowd 61 00:02:41,640 --> 00:02:45,300 Milling around you now 62 00:02:43,260 --> 00:02:48,360 you sense that they're grateful for your 63 00:02:45,300 --> 00:02:51,120 Aid but they tell you that others have 64 00:02:48,360 --> 00:02:54,780 tried and succeeded in chasing it off 65 00:02:51,120 --> 00:02:57,720 before and still it returns 66 00:02:54,780 --> 00:03:00,480 you know deep in your hearts that it 67 00:02:57,720 --> 00:03:01,680 will return unless the source is dealt 68 00:03:00,480 --> 00:03:04,140 with 69 00:03:01,680 --> 00:03:07,560 welcome to the campaign against AI bias 70 00:03:04,140 --> 00:03:10,560 I'm your DM Jay Rosenbaum and I'll be 71 00:03:07,560 --> 00:03:14,239 talking today about two of my passions 72 00:03:10,560 --> 00:03:17,580 bias in Ai and DND I have a doctorate 73 00:03:14,239 --> 00:03:20,519 from rmit exploring computer perceptions 74 00:03:17,580 --> 00:03:23,280 of gender and I research a great deal 75 00:03:20,519 --> 00:03:25,440 about AI bias and mitigation 76 00:03:23,280 --> 00:03:27,780 one of my major projects involved 77 00:03:25,440 --> 00:03:30,540 visibly de-biasing a generative 78 00:03:27,780 --> 00:03:33,060 adversarial Network to show how bias 79 00:03:30,540 --> 00:03:36,180 affects image generation and how it is 80 00:03:33,060 --> 00:03:37,920 possible to alter the bias of a biased 81 00:03:36,180 --> 00:03:40,920 neural network 82 00:03:37,920 --> 00:03:43,500 talk is a bit of a fun look at how you 83 00:03:40,920 --> 00:03:45,840 can seek out the issues in Ai and view 84 00:03:43,500 --> 00:03:49,319 AI through a critical lens 85 00:03:45,840 --> 00:03:52,080 AI bias is a monster that doesn't stop 86 00:03:49,319 --> 00:03:54,780 it's constantly coming back 87 00:03:52,080 --> 00:03:57,480 some people allow it to run rampant and 88 00:03:54,780 --> 00:04:00,360 some work to fight and subdue it so I 89 00:03:57,480 --> 00:04:02,760 want to look at that monster and how we 90 00:04:00,360 --> 00:04:05,580 can bring it down bit by bit to turn it 91 00:04:02,760 --> 00:04:06,780 from a horrifying aberration into a tiny 92 00:04:05,580 --> 00:04:09,000 nuisance 93 00:04:06,780 --> 00:04:11,400 where you can go to get the knowledge 94 00:04:09,000 --> 00:04:13,560 you need to fight it and how to equip 95 00:04:11,400 --> 00:04:15,680 yourselves against the battle ahead of 96 00:04:13,560 --> 00:04:15,680 you 97 00:04:15,780 --> 00:04:21,660 asking around you meet many who have 98 00:04:18,840 --> 00:04:23,759 been tainted by The Whispers some to a 99 00:04:21,660 --> 00:04:26,220 greater or lesser degree 100 00:04:23,759 --> 00:04:28,199 the ones with little effect seem to 101 00:04:26,220 --> 00:04:31,380 suffer from an apathy 102 00:04:28,199 --> 00:04:33,479 they tell you why bother the monster is 103 00:04:31,380 --> 00:04:35,940 just going to return maybe we should 104 00:04:33,479 --> 00:04:39,660 just stop fighting others are almost 105 00:04:35,940 --> 00:04:42,600 hostile well actually they say 106 00:04:39,660 --> 00:04:45,300 you sigh and move on 107 00:04:42,600 --> 00:04:48,419 in a Local Tavern you sit over to your 108 00:04:45,300 --> 00:04:49,979 drinks despondently wondering where to 109 00:04:48,419 --> 00:04:52,979 even start 110 00:04:49,979 --> 00:04:54,240 when you hear a soft I think I can help 111 00:04:52,979 --> 00:04:57,419 you 112 00:04:54,240 --> 00:04:59,400 you look up and see a traveling Bard who 113 00:04:57,419 --> 00:05:02,940 tells you he's been seeking adventurers 114 00:04:59,400 --> 00:05:05,160 to help with this terrifying Quest 115 00:05:02,940 --> 00:05:07,860 the bar tells you how the creature 116 00:05:05,160 --> 00:05:10,440 fights he tells you how it works in 117 00:05:07,860 --> 00:05:13,259 Whispers destroying the morale of those 118 00:05:10,440 --> 00:05:16,560 around it how it can subtly corrupt the 119 00:05:13,259 --> 00:05:19,800 people around it it doesn't feed on meat 120 00:05:16,560 --> 00:05:23,699 or people or any known food source it 121 00:05:19,800 --> 00:05:26,160 appears to feed on hatred it feeds the 122 00:05:23,699 --> 00:05:29,160 doubts of people and allows them to grow 123 00:05:26,160 --> 00:05:31,160 and fester as it returns again and again 124 00:05:29,160 --> 00:05:34,020 and again 125 00:05:31,160 --> 00:05:36,120 despairing afraid of this Monumental 126 00:05:34,020 --> 00:05:37,740 task in front of you you ask where can 127 00:05:36,120 --> 00:05:39,660 you even start with something like this 128 00:05:37,740 --> 00:05:42,720 can it even be done 129 00:05:39,660 --> 00:05:45,360 yes The Bard says 130 00:05:42,720 --> 00:05:47,220 he believes that it wasn't always like 131 00:05:45,360 --> 00:05:48,720 this it can be done 132 00:05:47,220 --> 00:05:51,300 other villagers have had different 133 00:05:48,720 --> 00:05:54,120 experiences with this monster 134 00:05:51,300 --> 00:05:56,639 it seems to be getting worse as it moves 135 00:05:54,120 --> 00:05:58,860 from town to town 136 00:05:56,639 --> 00:06:00,900 sir it needs to be confronted at the 137 00:05:58,860 --> 00:06:05,460 source in its lair 138 00:06:00,900 --> 00:06:08,160 to truly stop it your quest therefore is 139 00:06:05,460 --> 00:06:10,680 defined to the lair and work together to 140 00:06:08,160 --> 00:06:12,600 defeat this creature he also knows of 141 00:06:10,680 --> 00:06:15,720 someone who may be able to help in a 142 00:06:12,600 --> 00:06:18,660 shining Tower to the east 143 00:06:15,720 --> 00:06:21,300 it appears to be a huge task to slay the 144 00:06:18,660 --> 00:06:23,580 monster of AI bias it has so many 145 00:06:21,300 --> 00:06:26,880 different permutations so many different 146 00:06:23,580 --> 00:06:29,160 ways to appear in computer vision we 147 00:06:26,880 --> 00:06:31,940 constantly see examples of misgendering 148 00:06:29,160 --> 00:06:34,139 racist miscategorization homophobia 149 00:06:31,940 --> 00:06:37,500 transphobia sexism 150 00:06:34,139 --> 00:06:39,660 I often see examples of people using AI 151 00:06:37,500 --> 00:06:43,199 to decide who to police 152 00:06:39,660 --> 00:06:45,840 who took a classified to classify gender 153 00:06:43,199 --> 00:06:49,380 to ensure that hateful content is shown 154 00:06:45,840 --> 00:06:51,900 in social media or to add misgendering 155 00:06:49,380 --> 00:06:54,960 and sexism to language models 156 00:06:51,900 --> 00:06:58,139 and translation this is a problem that 157 00:06:54,960 --> 00:07:00,319 extends offline as too many people are 158 00:06:58,139 --> 00:07:03,900 arrested on the word of an algorithm 159 00:07:00,319 --> 00:07:06,300 without due process 160 00:07:03,900 --> 00:07:08,880 bias is a particularly large problem in 161 00:07:06,300 --> 00:07:11,460 Ai and computer science many of the 162 00:07:08,880 --> 00:07:14,819 researchers working in AI are men with 163 00:07:11,460 --> 00:07:17,100 80 of AI professors and 72 percent of 164 00:07:14,819 --> 00:07:19,560 presenters at AI conferences 165 00:07:17,100 --> 00:07:21,300 these numbers go on to influence the 166 00:07:19,560 --> 00:07:24,720 creation of data sets that are heavily 167 00:07:21,300 --> 00:07:26,940 male focused bias and AI leads to 168 00:07:24,720 --> 00:07:29,099 harmful algorithms that can be used for 169 00:07:26,940 --> 00:07:31,199 unethical practices that can actively 170 00:07:29,099 --> 00:07:34,080 harm women and gender and sexual 171 00:07:31,199 --> 00:07:36,300 minorities an AI was created to undress 172 00:07:34,080 --> 00:07:38,759 women using a pornography data set to 173 00:07:36,300 --> 00:07:41,220 fill in the Unseen areas while the 174 00:07:38,759 --> 00:07:43,800 initial algorithm was quickly denounced 175 00:07:41,220 --> 00:07:46,199 and removed by its creator after a 176 00:07:43,800 --> 00:07:49,380 backlash from the AI community 177 00:07:46,199 --> 00:07:51,979 many clones appeared almost immediately 178 00:07:49,380 --> 00:07:51,979 thereafter 179 00:07:52,199 --> 00:07:56,460 similar issues have surfaced as well 180 00:07:54,360 --> 00:07:57,599 with generative Ai and generative 181 00:07:56,460 --> 00:07:59,880 imagery 182 00:07:57,599 --> 00:08:02,280 AI has been used to attempt to detect 183 00:07:59,880 --> 00:08:04,500 homosexuality with several algorithms 184 00:08:02,280 --> 00:08:07,139 created using data sets based on 185 00:08:04,500 --> 00:08:09,300 photographs from Grindr and Tinder in an 186 00:08:07,139 --> 00:08:11,220 attempt to explore a modern physionomy 187 00:08:09,300 --> 00:08:12,900 of detection 188 00:08:11,220 --> 00:08:16,199 algorithms exist 189 00:08:12,900 --> 00:08:18,000 and while they don't properly work they 190 00:08:16,199 --> 00:08:20,699 were created without any consideration 191 00:08:18,000 --> 00:08:24,240 for the people who would be put most at 192 00:08:20,699 --> 00:08:26,520 risk by these by the bias of the AI 193 00:08:24,240 --> 00:08:29,400 creators 194 00:08:26,520 --> 00:08:31,440 bias Works insidiously to illustrate to 195 00:08:29,400 --> 00:08:34,680 people who are excluded that they don't 196 00:08:31,440 --> 00:08:37,140 matter that they don't exist 197 00:08:34,680 --> 00:08:39,300 that they're less important and these 198 00:08:37,140 --> 00:08:41,940 Concepts prevent people from moving 199 00:08:39,300 --> 00:08:44,820 ahead from being hired from being seen 200 00:08:41,940 --> 00:08:47,519 or from seeing others like them 201 00:08:44,820 --> 00:08:49,080 as Dr Sarah Lewis put it in her article 202 00:08:47,519 --> 00:08:52,140 on the racial bias built into 203 00:08:49,080 --> 00:08:54,360 photography you can't become what you 204 00:08:52,140 --> 00:08:56,600 can't accurately see 205 00:08:54,360 --> 00:08:59,760 if you cannot see others like yourself 206 00:08:56,600 --> 00:09:02,700 represented in power in art in 207 00:08:59,760 --> 00:09:05,100 literature in entertainment then you 208 00:09:02,700 --> 00:09:08,220 start to believing almost that you don't 209 00:09:05,100 --> 00:09:10,920 exist that you don't matter 210 00:09:08,220 --> 00:09:13,200 bias can also Propel the people who are 211 00:09:10,920 --> 00:09:17,580 represented to ignore and exclude those 212 00:09:13,200 --> 00:09:21,360 they don't view as normal or the default 213 00:09:17,580 --> 00:09:23,760 bias leads to racism to fear and hatred 214 00:09:21,360 --> 00:09:27,060 and while we're still in those nascent 215 00:09:23,760 --> 00:09:29,040 stages of teaching AI it's crucial that 216 00:09:27,060 --> 00:09:31,200 we teach it that all humans have equal 217 00:09:29,040 --> 00:09:33,720 value 218 00:09:31,200 --> 00:09:37,080 a neural network can even develop biases 219 00:09:33,720 --> 00:09:38,880 because of background content even if 220 00:09:37,080 --> 00:09:41,220 the background data is unlabeled 221 00:09:38,880 --> 00:09:43,500 researchers have found that children in 222 00:09:41,220 --> 00:09:45,180 the background of an image usually 223 00:09:43,500 --> 00:09:46,860 results in the foreground person being 224 00:09:45,180 --> 00:09:48,899 classified as a woman 225 00:09:46,860 --> 00:09:51,300 because women appear with children in 226 00:09:48,899 --> 00:09:53,040 the background more often even in a 227 00:09:51,300 --> 00:09:55,500 balanced data set 228 00:09:53,040 --> 00:09:57,899 sort of background noise reinforces that 229 00:09:55,500 --> 00:10:01,380 very bias we're trying to eliminate 230 00:09:57,899 --> 00:10:03,839 we see entirely too often these issues 231 00:10:01,380 --> 00:10:06,540 being dismissed or fixed with a quick 232 00:10:03,839 --> 00:10:08,640 fix solution that doesn't really work to 233 00:10:06,540 --> 00:10:11,399 solve the underlying problem 234 00:10:08,640 --> 00:10:14,760 it seems insurmountable to get past 235 00:10:11,399 --> 00:10:17,880 these issues AI amplifies the worst 236 00:10:14,760 --> 00:10:20,339 biases of humanity and spews it forth 237 00:10:17,880 --> 00:10:22,380 encouraging people to be worse because 238 00:10:20,339 --> 00:10:25,140 they see their inner beliefs manifested 239 00:10:22,380 --> 00:10:27,899 and it becomes a vicious cycle 240 00:10:25,140 --> 00:10:31,080 but there's hope there is harp there are 241 00:10:27,899 --> 00:10:33,240 ways to de-bias AI to mitigate the 242 00:10:31,080 --> 00:10:36,180 problems caused by improper data sets 243 00:10:33,240 --> 00:10:38,399 and biased code it helps to know that 244 00:10:36,180 --> 00:10:39,839 there are people out there who know what 245 00:10:38,399 --> 00:10:42,660 they're doing 246 00:10:39,839 --> 00:10:44,700 there are experts in AI bias and there 247 00:10:42,660 --> 00:10:46,920 are resources to help you learn 248 00:10:44,700 --> 00:10:49,440 just knowing that there are people and 249 00:10:46,920 --> 00:10:51,540 resources you can turn to starts to 250 00:10:49,440 --> 00:10:54,079 break the task down into more manageable 251 00:10:51,540 --> 00:10:54,079 pieces 252 00:10:56,779 --> 00:11:01,740 thank you 253 00:10:59,160 --> 00:11:03,060 the path to the tower heads through 254 00:11:01,740 --> 00:11:05,940 several villages 255 00:11:03,060 --> 00:11:08,579 each showing signs of the automaton 256 00:11:05,940 --> 00:11:11,160 passing through each one seems to tell a 257 00:11:08,579 --> 00:11:13,680 story of the creature's regression from 258 00:11:11,160 --> 00:11:15,180 The Villages that are mostly wreckage 259 00:11:13,680 --> 00:11:18,060 to recovering 260 00:11:15,180 --> 00:11:21,360 to barely damaged at all the villagers 261 00:11:18,060 --> 00:11:23,519 in each Town tell you of the automaton 262 00:11:21,360 --> 00:11:25,860 as it moved through and working 263 00:11:23,519 --> 00:11:28,740 backwards you can see that it started 264 00:11:25,860 --> 00:11:31,320 off helpful but as it moved from one 265 00:11:28,740 --> 00:11:32,279 Village to the next it met with more 266 00:11:31,320 --> 00:11:33,839 people 267 00:11:32,279 --> 00:11:37,019 something 268 00:11:33,839 --> 00:11:39,240 seemed to shift inside the creature 269 00:11:37,019 --> 00:11:42,180 as you make your way to the place it All 270 00:11:39,240 --> 00:11:44,160 Began you see the amounts of ooze all 271 00:11:42,180 --> 00:11:45,660 over everything taper off until there's 272 00:11:44,160 --> 00:11:47,579 none to be seen 273 00:11:45,660 --> 00:11:50,519 you see warped looking parrots 274 00:11:47,579 --> 00:11:52,140 occasionally as you track and signs of 275 00:11:50,519 --> 00:11:53,700 Treads moving away from the direction 276 00:11:52,140 --> 00:11:56,700 you're walking 277 00:11:53,700 --> 00:11:58,380 on the way the bird tells the tales of 278 00:11:56,700 --> 00:12:01,740 how the creature was created 279 00:11:58,380 --> 00:12:03,720 he knows how to spin a yarn and soon 280 00:12:01,740 --> 00:12:05,640 before you know it you're lost in the 281 00:12:03,720 --> 00:12:09,660 Bible's words 282 00:12:05,640 --> 00:12:12,959 an artificer a wizard and a Bard joined 283 00:12:09,660 --> 00:12:15,540 together to create something new 284 00:12:12,959 --> 00:12:18,120 they wanted to forge something to help 285 00:12:15,540 --> 00:12:20,760 people all over the world wanting to 286 00:12:18,120 --> 00:12:23,100 bring the best of themselves into this 287 00:12:20,760 --> 00:12:26,880 creature they channeled their hearts and 288 00:12:23,100 --> 00:12:29,640 souls and the sweat of their brows 289 00:12:26,880 --> 00:12:31,920 surrounding the automaton were three 290 00:12:29,640 --> 00:12:34,680 modes of potential 291 00:12:31,920 --> 00:12:38,339 glowing lights to help the creature 292 00:12:34,680 --> 00:12:39,600 become whatever it was that it can be 293 00:12:38,339 --> 00:12:42,779 however 294 00:12:39,600 --> 00:12:45,899 the tiny seeds of corruption that lay in 295 00:12:42,779 --> 00:12:49,440 each of their hearts that lay in all of 296 00:12:45,899 --> 00:12:51,720 our hearts unknown to us were drawn out 297 00:12:49,440 --> 00:12:55,560 in the process with the mots of 298 00:12:51,720 --> 00:12:57,779 potential and lay dormant until fed 299 00:12:55,560 --> 00:13:00,120 they had the best of intentions as they 300 00:12:57,779 --> 00:13:02,579 set the creature off alone to help the 301 00:13:00,120 --> 00:13:05,880 villagers but the more time it's spent 302 00:13:02,579 --> 00:13:08,459 with people the more it changed 303 00:13:05,880 --> 00:13:12,720 as it was exposed to the worst in people 304 00:13:08,459 --> 00:13:14,940 the Tiny Seed of corruption grew it fed 305 00:13:12,720 --> 00:13:18,240 off the corruption of others and little 306 00:13:14,940 --> 00:13:20,639 by little their Golden Dream turned into 307 00:13:18,240 --> 00:13:23,220 an aberrant nightmare 308 00:13:20,639 --> 00:13:24,720 the three disagreed as to how to proceed 309 00:13:23,220 --> 00:13:27,660 and split up 310 00:13:24,720 --> 00:13:29,399 the wizard chose to stay at the tower to 311 00:13:27,660 --> 00:13:32,220 research a way to fix it 312 00:13:29,399 --> 00:13:35,399 the artificer went to find the lair of 313 00:13:32,220 --> 00:13:37,740 the creature The Bard went forth to see 314 00:13:35,399 --> 00:13:40,019 what assistance could be granted anyway 315 00:13:37,740 --> 00:13:42,660 to fix the devastation wrought in the 316 00:13:40,019 --> 00:13:45,120 neighboring towns and to hopefully find 317 00:13:42,660 --> 00:13:47,459 a group of Brave adventurers to tackle 318 00:13:45,120 --> 00:13:49,980 this challenge 319 00:13:47,459 --> 00:13:52,260 so many people start off their AI 320 00:13:49,980 --> 00:13:55,260 creations with the best of intentions 321 00:13:52,260 --> 00:13:56,760 they consider what they can do without 322 00:13:55,260 --> 00:13:57,779 thinking enough about whether they 323 00:13:56,760 --> 00:14:00,000 should 324 00:13:57,779 --> 00:14:02,820 they want to benefit people but don't 325 00:14:00,000 --> 00:14:05,100 necessarily think about which people and 326 00:14:02,820 --> 00:14:06,420 what that impact will be 327 00:14:05,100 --> 00:14:08,700 sometimes they just want to make 328 00:14:06,420 --> 00:14:12,480 something to see if it can be made 329 00:14:08,700 --> 00:14:15,720 the problem is is that everyone has bias 330 00:14:12,480 --> 00:14:19,139 that is that tiny seat of corruption all 331 00:14:15,720 --> 00:14:21,899 are unknown and unconscious biases 332 00:14:19,139 --> 00:14:24,180 just knowing that you have inbuilt 333 00:14:21,899 --> 00:14:26,220 biases will help you start to see them 334 00:14:24,180 --> 00:14:29,220 and learn to factor them in and work 335 00:14:26,220 --> 00:14:31,860 past them an issue with this monster of 336 00:14:29,220 --> 00:14:33,839 AI bias is that the creature creators 337 00:14:31,860 --> 00:14:35,820 are three white men 338 00:14:33,839 --> 00:14:37,980 one method they could have used early on 339 00:14:35,820 --> 00:14:41,220 is to ensure that they had a diverse 340 00:14:37,980 --> 00:14:44,100 team working to create the creature 341 00:14:41,220 --> 00:14:46,920 it isn't enough to want to help people 342 00:14:44,100 --> 00:14:49,199 you need the perspectives of many people 343 00:14:46,920 --> 00:14:50,579 in order to ensure that no one is 344 00:14:49,199 --> 00:14:54,060 overlooked 345 00:14:50,579 --> 00:14:56,339 we need a balanced party developing AI 346 00:14:54,060 --> 00:14:59,399 so that we can ensure that the modes of 347 00:14:56,339 --> 00:15:01,860 Potential from everyone are included but 348 00:14:59,399 --> 00:15:03,899 the seeds of corruption are minimized by 349 00:15:01,860 --> 00:15:07,320 contrasting viewpoints 350 00:15:03,899 --> 00:15:09,540 AI is lazy it will only work with what 351 00:15:07,320 --> 00:15:11,579 it's taught so it's up to every AI 352 00:15:09,540 --> 00:15:13,800 Creator to ensure that it's taught a 353 00:15:11,579 --> 00:15:16,440 wide range of options so that it treats 354 00:15:13,800 --> 00:15:18,480 people with equally 355 00:15:16,440 --> 00:15:20,820 AI is like a child 356 00:15:18,480 --> 00:15:23,579 it needs to be raised with a gentle hand 357 00:15:20,820 --> 00:15:26,519 and it needs to be exposed to a wide 358 00:15:23,579 --> 00:15:28,860 range of experiences and people so that 359 00:15:26,519 --> 00:15:31,440 it does not get confused when presented 360 00:15:28,860 --> 00:15:34,560 with something outside of its knowledge 361 00:15:31,440 --> 00:15:37,079 it also needs to be supervised so that 362 00:15:34,560 --> 00:15:39,420 when issues do arise they can be quickly 363 00:15:37,079 --> 00:15:41,699 dealt with 364 00:15:39,420 --> 00:15:43,800 lack of diversity and development has 365 00:15:41,699 --> 00:15:45,779 led to issues like Apple's facial 366 00:15:43,800 --> 00:15:48,000 recognition and passport Checkers 367 00:15:45,779 --> 00:15:49,199 repeatedly failing through improper 368 00:15:48,000 --> 00:15:51,600 training 369 00:15:49,199 --> 00:15:52,680 AI is about using math to make 370 00:15:51,600 --> 00:15:54,959 predictions 371 00:15:52,680 --> 00:15:56,760 about using the data it's given to 372 00:15:54,959 --> 00:16:00,660 extrapolate further 373 00:15:56,760 --> 00:16:02,579 this can be extremely powerful but ai's 374 00:16:00,660 --> 00:16:03,839 also been proven to focus on the wrong 375 00:16:02,579 --> 00:16:06,959 things 376 00:16:03,839 --> 00:16:09,360 like rulers in skin cancer detection or 377 00:16:06,959 --> 00:16:11,820 drains in pneumonia patients a more 378 00:16:09,360 --> 00:16:14,579 recent covert detector turned out to be 379 00:16:11,820 --> 00:16:16,560 a lying down detector as the people who 380 00:16:14,579 --> 00:16:18,420 were most sick were scanned lying down 381 00:16:16,560 --> 00:16:21,480 rather than standing up 382 00:16:18,420 --> 00:16:22,620 getting lazy with data makes for a lazy 383 00:16:21,480 --> 00:16:24,540 AI 384 00:16:22,620 --> 00:16:28,260 but what do you do if the AI is already 385 00:16:24,540 --> 00:16:31,079 biased stay tuned 386 00:16:28,260 --> 00:16:34,199 as you see the tower looming on the 387 00:16:31,079 --> 00:16:37,560 horizon gleaming in the Setting Sun you 388 00:16:34,199 --> 00:16:40,820 hear a raucous cry from behind you and 389 00:16:37,560 --> 00:16:44,880 feel the impact of Whispers on your mind 390 00:16:40,820 --> 00:16:46,079 the parrots who seemed so few and far 391 00:16:44,880 --> 00:16:48,540 away 392 00:16:46,079 --> 00:16:51,240 have increased in number and are Milling 393 00:16:48,540 --> 00:16:54,420 around you now entering your minds with 394 00:16:51,240 --> 00:16:57,540 doubts and questions their feathers look 395 00:16:54,420 --> 00:17:00,779 like individual droplets of ooze 396 00:16:57,540 --> 00:17:05,160 these are stochastic parrots filled with 397 00:17:00,779 --> 00:17:07,199 bile and Corruption insults and fury 398 00:17:05,160 --> 00:17:09,780 they attack your minds with dissonant 399 00:17:07,199 --> 00:17:12,120 Whispers designed to feed that Tiny Seed 400 00:17:09,780 --> 00:17:12,959 of corruption inside of you but you are 401 00:17:12,120 --> 00:17:15,059 strong 402 00:17:12,959 --> 00:17:17,600 and eventually the stochastic parrots 403 00:17:15,059 --> 00:17:17,600 are defeated 404 00:17:17,640 --> 00:17:21,959 stochastic parrot is a term for a very 405 00:17:20,339 --> 00:17:24,120 large language models for natural 406 00:17:21,959 --> 00:17:25,980 language processing systems 407 00:17:24,120 --> 00:17:28,740 they're called stochastic parrots 408 00:17:25,980 --> 00:17:30,660 because they predict the next word and 409 00:17:28,740 --> 00:17:33,179 parrot from their Source material with 410 00:17:30,660 --> 00:17:35,460 no real understanding 411 00:17:33,179 --> 00:17:37,500 these are often trained with a huge 412 00:17:35,460 --> 00:17:38,340 amount of data dragged from all over the 413 00:17:37,500 --> 00:17:40,679 internet 414 00:17:38,340 --> 00:17:41,539 and the reliability of the text is not 415 00:17:40,679 --> 00:17:47,160 great 416 00:17:41,539 --> 00:17:48,380 after all if your data includes Reddit 417 00:17:47,160 --> 00:17:51,539 um how how 418 00:17:48,380 --> 00:17:54,000 biased are your results going to be 419 00:17:51,539 --> 00:17:56,299 the paper the timna jebrow and Margaret 420 00:17:54,000 --> 00:17:58,620 Mitchell were fired from Google over 421 00:17:56,299 --> 00:18:00,960 discusses the problems of stochastic 422 00:17:58,620 --> 00:18:02,940 parrots length and I felt it was 423 00:18:00,960 --> 00:18:05,340 appropriate to not only include these as 424 00:18:02,940 --> 00:18:08,100 low level encounters but also to 425 00:18:05,340 --> 00:18:09,960 generate an insult table using AI for 426 00:18:08,100 --> 00:18:12,720 some quality insults 427 00:18:09,960 --> 00:18:15,240 it was actually challenging to generate 428 00:18:12,720 --> 00:18:17,640 insults that were not biased towards any 429 00:18:15,240 --> 00:18:20,820 particular group of people and I tried a 430 00:18:17,640 --> 00:18:22,740 number of models but finally chat GPT 431 00:18:20,820 --> 00:18:25,020 ended up generating some truly 432 00:18:22,740 --> 00:18:27,240 devastating insults for the roll table 433 00:18:25,020 --> 00:18:29,640 with the stern admonishment that I 434 00:18:27,240 --> 00:18:33,260 remember to use them in a playful and 435 00:18:29,640 --> 00:18:33,260 lighthearted manner in my games 436 00:18:33,660 --> 00:18:38,100 stochastic parrots are an ongoing 437 00:18:35,880 --> 00:18:40,500 problem in AI because we've frankly 438 00:18:38,100 --> 00:18:42,059 become a little lazy in the method of 439 00:18:40,500 --> 00:18:45,419 data set creation 440 00:18:42,059 --> 00:18:47,340 we want huge data sets but we're often 441 00:18:45,419 --> 00:18:49,200 unwilling to go to the trouble of 442 00:18:47,340 --> 00:18:50,880 creating every image or piece of 443 00:18:49,200 --> 00:18:53,340 training text Within 444 00:18:50,880 --> 00:18:55,320 let alone working to ensure that every 445 00:18:53,340 --> 00:18:57,660 part is unbiased 446 00:18:55,320 --> 00:19:00,480 I'm sure that the original intent was to 447 00:18:57,660 --> 00:19:02,460 put everything in and let the AI sort it 448 00:19:00,480 --> 00:19:05,220 out bring in all perspectives and hope 449 00:19:02,460 --> 00:19:08,360 for the best but when you put garbage in 450 00:19:05,220 --> 00:19:08,360 you know what's going to happen 451 00:19:09,059 --> 00:19:13,559 bringing lightly people have been 452 00:19:11,700 --> 00:19:16,080 turning back to smaller language models 453 00:19:13,559 --> 00:19:18,720 again and training on better Chosen and 454 00:19:16,080 --> 00:19:22,020 curated data sets and this is paying off 455 00:19:18,720 --> 00:19:23,880 it's more work it is always more work to 456 00:19:22,020 --> 00:19:25,260 do the right thing but the end result 457 00:19:23,880 --> 00:19:27,600 pays off 458 00:19:25,260 --> 00:19:29,580 the newest large language models have a 459 00:19:27,600 --> 00:19:31,679 lot of work behind the scenes to help 460 00:19:29,580 --> 00:19:34,799 mitigate bias and will acknowledge that 461 00:19:31,679 --> 00:19:36,900 there are imperfect Creations which is a 462 00:19:34,799 --> 00:19:39,360 small start 463 00:19:36,900 --> 00:19:41,400 another way of cutting Corners is by 464 00:19:39,360 --> 00:19:43,620 using image classification or mass 465 00:19:41,400 --> 00:19:46,980 worker groups like Amazon's Mechanical 466 00:19:43,620 --> 00:19:49,620 Turk to caption images in this way as 467 00:19:46,980 --> 00:19:51,840 with AI you have to be extremely careful 468 00:19:49,620 --> 00:19:56,000 with what you ask and you need to 469 00:19:51,840 --> 00:19:56,000 supervise the outcomes carefully 470 00:19:56,160 --> 00:20:02,039 a recent study by MIT of popular public 471 00:19:59,100 --> 00:20:03,660 data sets showed label errors up to 3.4 472 00:20:02,039 --> 00:20:08,460 percent of the time 473 00:20:03,660 --> 00:20:09,840 in imagenet that amounts to 482 702 474 00:20:08,460 --> 00:20:13,020 images 475 00:20:09,840 --> 00:20:16,380 if that error rate is true for the lion 476 00:20:13,020 --> 00:20:19,820 data set the huge data set behind most 477 00:20:16,380 --> 00:20:24,000 text to image systems we are looking at 478 00:20:19,820 --> 00:20:26,400 170 million images that have errors 479 00:20:24,000 --> 00:20:28,500 some of these are harmless like labeling 480 00:20:26,400 --> 00:20:31,380 a Chihuahua as a feather boa 481 00:20:28,500 --> 00:20:33,059 but it gets more Insidious when it comes 482 00:20:31,380 --> 00:20:35,820 to images of humans 483 00:20:33,059 --> 00:20:37,740 in the lion data set you have to be 484 00:20:35,820 --> 00:20:39,660 careful what you search for 485 00:20:37,740 --> 00:20:42,660 when generated when searching for the 486 00:20:39,660 --> 00:20:44,460 word Asian just on its own it returns 487 00:20:42,660 --> 00:20:46,559 nothing but porn 488 00:20:44,460 --> 00:20:48,840 when this is used in the back end of 489 00:20:46,559 --> 00:20:51,360 image generation what is it learning 490 00:20:48,840 --> 00:20:53,220 about Asian people and how is its 491 00:20:51,360 --> 00:20:56,280 representation going to look 492 00:20:53,220 --> 00:20:58,860 in my own research into the MS Coco data 493 00:20:56,280 --> 00:21:03,740 set I found numerous errors despite 494 00:20:58,860 --> 00:21:03,740 being hand labeled by mtech workers 495 00:21:07,740 --> 00:21:11,160 at the top of the tower you meet the 496 00:21:10,260 --> 00:21:14,220 wizard 497 00:21:11,160 --> 00:21:16,559 he's weedy with a golden band holding 498 00:21:14,220 --> 00:21:19,500 his long hair back from his face 499 00:21:16,559 --> 00:21:22,559 he wipes his brow and sets down his book 500 00:21:19,500 --> 00:21:24,960 as you come in his face resigned he 501 00:21:22,559 --> 00:21:27,620 knows why you're here 502 00:21:24,960 --> 00:21:30,179 as you explain the story his face Falls 503 00:21:27,620 --> 00:21:31,980 he sighs at the end of your tail and 504 00:21:30,179 --> 00:21:34,320 sits down dejectedly 505 00:21:31,980 --> 00:21:36,120 I knew this would happen he says I told 506 00:21:34,320 --> 00:21:37,440 them it wasn't ready to go out in public 507 00:21:36,120 --> 00:21:40,200 yet 508 00:21:37,440 --> 00:21:42,120 supposed to be a bit of fun into 509 00:21:40,200 --> 00:21:45,720 exploring something that could help 510 00:21:42,120 --> 00:21:46,740 people we didn't know it gets so bad so 511 00:21:45,720 --> 00:21:49,020 quickly 512 00:21:46,740 --> 00:21:51,840 he continues to hate the idea of taking 513 00:21:49,020 --> 00:21:54,480 it down but we've got no choice we need 514 00:21:51,840 --> 00:21:56,100 to reset it at the very least 515 00:21:54,480 --> 00:21:58,080 but if you're going to find a way to 516 00:21:56,100 --> 00:22:00,179 remove the corruption without destroying 517 00:21:58,080 --> 00:22:02,700 all our work that would be best there's 518 00:22:00,179 --> 00:22:05,280 still good in it I reckon we have an 519 00:22:02,700 --> 00:22:08,340 idea for how it can be saved 520 00:22:05,280 --> 00:22:11,159 he tells you the automaton is not to 521 00:22:08,340 --> 00:22:13,380 blame but the corruption surrounding it 522 00:22:11,159 --> 00:22:14,820 and if you take out the corruption you 523 00:22:13,380 --> 00:22:16,140 may have a chance to save the machine 524 00:22:14,820 --> 00:22:18,600 itself 525 00:22:16,140 --> 00:22:21,419 he tells you that the corruption like 526 00:22:18,600 --> 00:22:22,740 the machine has the power to learn to 527 00:22:21,419 --> 00:22:25,200 grow and shift 528 00:22:22,740 --> 00:22:27,900 that adventurous such as yourselves need 529 00:22:25,200 --> 00:22:31,020 to be especially wary of its ability to 530 00:22:27,900 --> 00:22:33,600 adapt and to vary your attacks and 531 00:22:31,020 --> 00:22:34,980 skills to keep guessing 532 00:22:33,600 --> 00:22:37,740 finally 533 00:22:34,980 --> 00:22:40,799 he tells you that he he has a telescope 534 00:22:37,740 --> 00:22:43,980 and a scene a trail of ooze heading to 535 00:22:40,799 --> 00:22:46,140 and from a swamp nearby 536 00:22:43,980 --> 00:22:48,240 it may have a layer somewhere in that 537 00:22:46,140 --> 00:22:51,240 swamp 538 00:22:48,240 --> 00:22:53,760 bias and AI often comes down to two 539 00:22:51,240 --> 00:22:56,159 major areas the data set and the 540 00:22:53,760 --> 00:22:58,679 original concept what do you want to do 541 00:22:56,159 --> 00:23:01,740 and how do you want to do it 542 00:22:58,679 --> 00:23:03,600 it's very easy to get carried away with 543 00:23:01,740 --> 00:23:05,880 an idea and want to develop it further 544 00:23:03,600 --> 00:23:08,340 but you need to consider is this 545 00:23:05,880 --> 00:23:09,600 something that's going to help people 546 00:23:08,340 --> 00:23:11,580 is this something that could 547 00:23:09,600 --> 00:23:13,700 inadvertently hurt people 548 00:23:11,580 --> 00:23:17,280 fixing the problem at the beginning 549 00:23:13,700 --> 00:23:19,919 before it becomes a giant monster is 550 00:23:17,280 --> 00:23:22,860 much easier than at the end and 551 00:23:19,919 --> 00:23:24,059 sometimes the best idea is never to 552 00:23:22,860 --> 00:23:28,200 start at all 553 00:23:24,059 --> 00:23:30,419 in any form of AI based production the 554 00:23:28,200 --> 00:23:33,299 ethics of what you are making should be 555 00:23:30,419 --> 00:23:37,220 part of the initial risk assessment you 556 00:23:33,299 --> 00:23:37,220 do make risk assessments right 557 00:23:38,159 --> 00:23:43,799 you head towards the swamp as you enter 558 00:23:41,100 --> 00:23:46,440 doubt starts to build in all of you a 559 00:23:43,799 --> 00:23:47,820 sense of dread at the Monumental task 560 00:23:46,440 --> 00:23:49,740 ahead of you 561 00:23:47,820 --> 00:23:51,659 what will you find 562 00:23:49,740 --> 00:23:54,840 as you hack and fight your way through 563 00:23:51,659 --> 00:23:56,480 the swamp it seems to grow darker as the 564 00:23:54,840 --> 00:23:59,820 light is swallowed up by the vegetation 565 00:23:56,480 --> 00:24:02,820 even the boat who's been merrily humming 566 00:23:59,820 --> 00:24:05,460 along until now starts to change his 567 00:24:02,820 --> 00:24:07,320 tune to a more sonorous Melody 568 00:24:05,460 --> 00:24:09,720 filled with Dread 569 00:24:07,320 --> 00:24:12,179 as you venture further into the darkness 570 00:24:09,720 --> 00:24:14,780 your hearts sink 571 00:24:12,179 --> 00:24:18,419 more and more with each passing step 572 00:24:14,780 --> 00:24:21,679 just as it feels you can't go further a 573 00:24:18,419 --> 00:24:23,880 little glowing moat a tiny music note 574 00:24:21,679 --> 00:24:26,220 appears in front of you and starts 575 00:24:23,880 --> 00:24:30,059 swirling Around The Bard 576 00:24:26,220 --> 00:24:32,700 that tiny light is like a ray of Hope 577 00:24:30,059 --> 00:24:35,220 and your legs feel lighter your mind's 578 00:24:32,700 --> 00:24:37,740 clearer as you get deeper into the heart 579 00:24:35,220 --> 00:24:40,320 of the Jungle another motor appears 580 00:24:37,740 --> 00:24:42,299 shining blue quill that beckons you 581 00:24:40,320 --> 00:24:45,120 onward 582 00:24:42,299 --> 00:24:47,159 you emerge into a clearing with a 583 00:24:45,120 --> 00:24:49,200 metallic structure built into the trees 584 00:24:47,159 --> 00:24:52,440 and a small hillock 585 00:24:49,200 --> 00:24:55,080 in front of it is a glowing moat shaped 586 00:24:52,440 --> 00:24:57,179 like a green spanner and all three start 587 00:24:55,080 --> 00:25:00,480 circling the bard 588 00:24:57,179 --> 00:25:02,100 as you face the metallic structure and 589 00:25:00,480 --> 00:25:04,080 square your shoulders 590 00:25:02,100 --> 00:25:06,059 a wild-eyed man bursts from the 591 00:25:04,080 --> 00:25:07,679 undergrowth he shapes his spanner at the 592 00:25:06,059 --> 00:25:09,059 bar and says you 593 00:25:07,679 --> 00:25:11,640 but whatever he was going to say after 594 00:25:09,059 --> 00:25:14,700 that is immaterial as the automaton 595 00:25:11,640 --> 00:25:16,799 appears from its Lair and attacks 596 00:25:14,700 --> 00:25:19,500 where before you felt completely 597 00:25:16,799 --> 00:25:22,200 steamrollered now you understand more 598 00:25:19,500 --> 00:25:25,440 about the machine you're ready 599 00:25:22,200 --> 00:25:28,320 you know it learns and adapts you know 600 00:25:25,440 --> 00:25:30,960 the Ooze is to blame you bring all your 601 00:25:28,320 --> 00:25:33,659 attacks to Bear little oozes drop off it 602 00:25:30,960 --> 00:25:36,299 and join the attack but now those of you 603 00:25:33,659 --> 00:25:38,520 with high passive perception notice that 604 00:25:36,299 --> 00:25:42,179 some return to it and it seems to learn 605 00:25:38,520 --> 00:25:44,640 something about though what about who 606 00:25:42,179 --> 00:25:48,120 the Ooze was attacking at the beginning 607 00:25:44,640 --> 00:25:49,980 it almost seems weak and you're amazed 608 00:25:48,120 --> 00:25:51,059 at the ease with which you're fighting 609 00:25:49,980 --> 00:25:53,480 it back 610 00:25:51,059 --> 00:25:57,299 but they're more you the more you attack 611 00:25:53,480 --> 00:25:59,640 the less your attacks seem to do 612 00:25:57,299 --> 00:26:02,100 at the beginning the tendrils it was 613 00:25:59,640 --> 00:26:04,559 lashing out with were doing piercing 614 00:26:02,100 --> 00:26:07,080 damage but one grows a club like end 615 00:26:04,559 --> 00:26:08,940 after a news returns to it another 616 00:26:07,080 --> 00:26:10,559 crackles with lightning after a 617 00:26:08,940 --> 00:26:13,860 different ooze returns 618 00:26:10,559 --> 00:26:16,740 you start to grasp how it's learning 619 00:26:13,860 --> 00:26:18,960 making Headway finally a sorcerer lets 620 00:26:16,740 --> 00:26:21,179 out a fireball and some of the Ooze 621 00:26:18,960 --> 00:26:21,900 grows back over the Golden body once 622 00:26:21,179 --> 00:26:24,120 more 623 00:26:21,900 --> 00:26:27,360 the artificer and The Bard together yell 624 00:26:24,120 --> 00:26:30,240 out no fire and the Sorcerer looks at 625 00:26:27,360 --> 00:26:32,460 bashed but you know shifts their attacks 626 00:26:30,240 --> 00:26:34,260 The Bard and the artificer turned to 627 00:26:32,460 --> 00:26:37,440 each other limbed in the glow of battle 628 00:26:34,260 --> 00:26:39,600 and smile the difference is forgotten 629 00:26:37,440 --> 00:26:41,640 everyone Burns their spell slots 630 00:26:39,600 --> 00:26:45,000 changing their attacks again and again 631 00:26:41,640 --> 00:26:47,820 and eventually you see more ooze 632 00:26:45,000 --> 00:26:50,880 dropping away and dissolving 633 00:26:47,820 --> 00:26:52,799 little by little the Ooze clears until 634 00:26:50,880 --> 00:26:55,679 you can see more and more of the 635 00:26:52,799 --> 00:26:58,679 original golden automaton appearing you 636 00:26:55,679 --> 00:27:02,720 press your attacks and finally the last 637 00:26:58,679 --> 00:27:02,720 of the years vanishes from the machine 638 00:27:02,880 --> 00:27:07,919 defeating bias in AIP requires 639 00:27:05,159 --> 00:27:10,080 multi-pronged attacks it isn't enough to 640 00:27:07,919 --> 00:27:11,760 just fix the data set or delete the 641 00:27:10,080 --> 00:27:14,100 offending categories you need to 642 00:27:11,760 --> 00:27:16,320 systematically consider where the issues 643 00:27:14,100 --> 00:27:18,720 are coming from where it where it has 644 00:27:16,320 --> 00:27:20,640 learned what it has learned and how you 645 00:27:18,720 --> 00:27:23,220 need to look at if you are tuning hyper 646 00:27:20,640 --> 00:27:25,980 parameters and see the results when you 647 00:27:23,220 --> 00:27:27,779 edit that tuning or turn it off you need 648 00:27:25,980 --> 00:27:30,059 to examine the waiting you have applied 649 00:27:27,779 --> 00:27:31,860 to different factors explore your sample 650 00:27:30,059 --> 00:27:34,620 weights and test to make sure they're 651 00:27:31,860 --> 00:27:36,900 being counted you can use them to add 652 00:27:34,620 --> 00:27:38,039 more impact to categories and downplay 653 00:27:36,900 --> 00:27:40,320 others 654 00:27:38,039 --> 00:27:41,760 you need to try different things and 655 00:27:40,320 --> 00:27:44,220 objectively compare the different 656 00:27:41,760 --> 00:27:46,740 results need to really look at those 657 00:27:44,220 --> 00:27:48,720 algorithms and make sure that every 658 00:27:46,740 --> 00:27:51,000 option is being considered and trained 659 00:27:48,720 --> 00:27:53,360 you need to take a deep look at your 660 00:27:51,000 --> 00:27:56,220 data sets and the representation within 661 00:27:53,360 --> 00:27:57,900 and you need to cons look at the 662 00:27:56,220 --> 00:27:59,340 confidence of your systems and the 663 00:27:57,900 --> 00:28:00,360 outputs to ensure that there's a good 664 00:27:59,340 --> 00:28:02,820 balance 665 00:28:00,360 --> 00:28:04,380 you may even need to bias it in other 666 00:28:02,820 --> 00:28:06,360 ways to counterbalance the original 667 00:28:04,380 --> 00:28:08,580 biases 668 00:28:06,360 --> 00:28:10,860 buyers can come from the most innocuous 669 00:28:08,580 --> 00:28:13,020 of sources it can come from what you 670 00:28:10,860 --> 00:28:14,640 choose not to include as much as what 671 00:28:13,020 --> 00:28:16,860 you choose to include 672 00:28:14,640 --> 00:28:18,659 it can come from just not understanding 673 00:28:16,860 --> 00:28:20,960 that there are many different parts at 674 00:28:18,659 --> 00:28:20,960 play 675 00:28:22,320 --> 00:28:27,120 I added fire as a restoration device 676 00:28:24,360 --> 00:28:28,679 deliberately as a nod to AI trolls well 677 00:28:27,120 --> 00:28:31,320 sometimes it feels like nuking it with 678 00:28:28,679 --> 00:28:33,779 fire may be the only way out fighting 679 00:28:31,320 --> 00:28:37,200 fire with fire only makes the situation 680 00:28:33,779 --> 00:28:38,820 worse you need to be clever to adapt and 681 00:28:37,200 --> 00:28:41,640 to think carefully about how the 682 00:28:38,820 --> 00:28:44,340 algorithm works and how it adapts the 683 00:28:41,640 --> 00:28:47,460 information you give it 684 00:28:44,340 --> 00:28:49,620 the artificer and The Bard moved towards 685 00:28:47,460 --> 00:28:52,440 The Shining automaton and start speaking 686 00:28:49,620 --> 00:28:55,140 to it gently its eyes change and the 687 00:28:52,440 --> 00:28:58,440 glowing Motes start circling around it 688 00:28:55,140 --> 00:29:00,779 faster and faster The Bard starts to 689 00:28:58,440 --> 00:29:04,500 play and as he does a glowing moat 690 00:29:00,779 --> 00:29:06,539 erupts from your chest from all of you 691 00:29:04,500 --> 00:29:09,000 a little symbol of something close to 692 00:29:06,539 --> 00:29:12,240 your heart a motive potential that only 693 00:29:09,000 --> 00:29:14,400 you can provide you look around and 694 00:29:12,240 --> 00:29:16,080 everybody has a moat 695 00:29:14,400 --> 00:29:19,020 that swells around the head before 696 00:29:16,080 --> 00:29:21,360 floating over to the automaton 697 00:29:19,020 --> 00:29:23,399 they've start to spend becoming a Halo 698 00:29:21,360 --> 00:29:27,120 of Lights becoming too white to almost 699 00:29:23,399 --> 00:29:29,460 look at and then vanish with the Echoes 700 00:29:27,120 --> 00:29:33,179 of those lights shining in the eyes of 701 00:29:29,460 --> 00:29:36,179 the now calm golden automaton 702 00:29:33,179 --> 00:29:38,580 we all bring something to the table we 703 00:29:36,179 --> 00:29:42,480 all have unique lived experiences and 704 00:29:38,580 --> 00:29:45,779 perspectives that we can use to make AI 705 00:29:42,480 --> 00:29:48,179 systems better and more well-rounded we 706 00:29:45,779 --> 00:29:50,580 all have a motive potential that can 707 00:29:48,179 --> 00:29:55,140 help make a system better just by asking 708 00:29:50,580 --> 00:29:57,299 what if why how what can we do that will 709 00:29:55,140 --> 00:30:00,000 make a difference who does this say I 710 00:29:57,299 --> 00:30:03,779 help who does at home how can we make 711 00:30:00,000 --> 00:30:06,059 this AI more equitable how do we prevent 712 00:30:03,779 --> 00:30:09,659 the mistakes of the past and how do we 713 00:30:06,059 --> 00:30:12,539 reflect the society we want to be rather 714 00:30:09,659 --> 00:30:14,940 than the one we have 715 00:30:12,539 --> 00:30:17,100 in this talk we've seen what it looks 716 00:30:14,940 --> 00:30:18,960 like when everything goes well but what 717 00:30:17,100 --> 00:30:21,059 happens if everything goes wrong 718 00:30:18,960 --> 00:30:23,700 what if your party wasn't strong enough 719 00:30:21,059 --> 00:30:25,860 you can always disengage come back after 720 00:30:23,700 --> 00:30:28,500 a long rest try again you can buff 721 00:30:25,860 --> 00:30:30,419 yourselves and your co-workers learn 722 00:30:28,500 --> 00:30:32,760 about everything look at the code with 723 00:30:30,419 --> 00:30:36,120 fresh eyes get someone else to take a 724 00:30:32,760 --> 00:30:39,059 look you're a party after all and this 725 00:30:36,120 --> 00:30:42,179 task is not just on your shoulders but 726 00:30:39,059 --> 00:30:44,760 on you all the quest Falls to you as a 727 00:30:42,179 --> 00:30:46,919 group to find the bugs in the system to 728 00:30:44,760 --> 00:30:48,120 look at the data sets and the code and 729 00:30:46,919 --> 00:30:50,340 do the work 730 00:30:48,120 --> 00:30:52,740 Define the areas that are being missed 731 00:30:50,340 --> 00:30:55,080 so this is not a task for one person but 732 00:30:52,740 --> 00:30:57,240 for a team and if you do find that 733 00:30:55,080 --> 00:30:59,580 you're alone in this endeavor if it's 734 00:30:57,240 --> 00:31:01,980 looking Bleak there are resources online 735 00:30:59,580 --> 00:31:04,380 there are mercenaries you can hire to 736 00:31:01,980 --> 00:31:06,779 help there are Elder Gods like Google or 737 00:31:04,380 --> 00:31:08,340 chat GPT capricious in nature but 738 00:31:06,779 --> 00:31:10,380 they're available there are 739 00:31:08,340 --> 00:31:12,539 organizations such as the algorithmic 740 00:31:10,380 --> 00:31:15,000 Justice League who can furnish you with 741 00:31:12,539 --> 00:31:18,120 information and assist you in your quest 742 00:31:15,000 --> 00:31:20,460 you're never alone even when it feels 743 00:31:18,120 --> 00:31:22,679 dark there are always others with 744 00:31:20,460 --> 00:31:24,779 shining lights ready to help you in your 745 00:31:22,679 --> 00:31:27,899 quest 746 00:31:24,779 --> 00:31:29,039 the automaton is wide-eyed friendly 747 00:31:27,899 --> 00:31:31,620 looking even 748 00:31:29,039 --> 00:31:34,140 the lights from all those moats shining 749 00:31:31,620 --> 00:31:37,140 into its eyes it turns to you placidly 750 00:31:34,140 --> 00:31:40,140 waiting for your input it's here to help 751 00:31:37,140 --> 00:31:42,419 you that's all it ever wanted was to 752 00:31:40,140 --> 00:31:45,419 help you in your tasks and make things 753 00:31:42,419 --> 00:31:49,820 easier it's what you do with it that 754 00:31:45,419 --> 00:31:51,320 counts what are you going to do next 755 00:31:49,820 --> 00:32:01,369 thank you 756 00:31:51,320 --> 00:32:01,369 [Applause]