How To Make AI Good For Humanity — Transcript

Full transcript

1. 0:00I wish we had a concrete plan to make  sure AI ends up being good for Humanity
2. 0:05Worry not, sock puppet framing device
3. 0:08Your conveniently specific wish has been granted
4. 0:11Wait, no! I wish to be a real boy
5. 0:13A group of top scientists, faith  leaders, child safety experts,
6. 0:17labor union representatives,  human rights activists, liberals,
7. 0:21and conservatives from around the world have come  together to create the pro-human AI declaration
8. 0:27It's a list of 33 principles for how to make sure  that AI technology ends up being good for humanity
9. 0:34I think this declaration is fantastic
10. 0:37It's important that we have concrete demands  for what AI regulation should look like
11. 0:42It's not enough to just point out problems  with the AI industry and demand change
12. 0:47We have to be specific about what changes we want
13. 0:50It's also helpful that this is a compromise  document co-written by people from many
14. 0:55different political backgrounds,  not just AI safety nerds like myself
15. 1:00Because I think this declaration  is genuinely important, I'd like
16. 1:04to share these principles with you  alongside some commentary from me
17. 1:07If I'm on screen, you're hearing my opinions
18. 1:10If I'm not on screen, you're hearing  the declaration itself as written
19. 1:14Keep in mind that my thoughts  about a particular principle
20. 1:17might be different from what the authors intended
21. 1:19Also, quick conflict of interest alert
22. 1:22The Future of Life Institute, which co-hosted the  convention where this declaration was written,
23. 1:27provides financial support to Siliconversations
24. 1:30However, as usual, they did not get  to see the script for this video or
25. 1:34the video itself before it was released,  and all the opinions I express are my own
26. 1:39Cool?
27. 1:39Cool
28. 1:40Principle number one: human  control is non-negotiable
29. 1:44Humanity must remain in control
30. 1:47Humans should choose how and whether  to delegate decisions to AI systems
31. 1:52Great first rule, 10 out of 10
32. 1:54If we ever get to the point where AIs are  in charge and making important decisions
33. 1:58outside of human control, then we are already in  a dystopia and quite possibly doomed as a species
34. 2:04Number two: meaningful human control
35. 2:07Humans should have the authority  and capacity to understand, guide,
36. 2:12proscribe, and override AI systems
37. 2:15If an AI offers you a choice  between option A and option B,
38. 2:18and you don't know what either option actually  does, that's not really you making a decision
39. 2:24We have to understand what AIs are doing  in order to meaningfully control them
40. 2:28We don't want to end up like  Michael Reeve’s goldfish,
41. 2:31technically choosing which stocks to invest in,
42. 2:34but really just being toyed with by a chaotic  Eldritch horror far beyond our comprehension
43. 2:39Number three: no super intelligence race
44. 2:42Development of super intelligence should be  prohibited until there is broad scientific
45. 2:47consensus that it can be done safely and  controllably, and there is strong public buy-in
46. 2:52This is pretty much the same language as  the Statement on Superintelligence from 2025
47. 2:57Several companies have announced their intention
48. 3:00to develop AI systems that are vastly  more intelligent than any human being
49. 3:04This is clearly stupid and dangerous,  and we should make them stop
50. 3:07If building superintelligence is possible, it  shouldn't be allowed until we know that it's safe
51. 3:13And if it's not possible, then banning  it doesn't have any negative consequences
52. 3:17Like, who cares if we ban an impossible thing?
53. 3:20Number four: off switch
54. 3:22Powerful AI systems must have mechanisms that  allow human operators to promptly shut them down
55. 3:28Excellent rule
56. 3:29Obviously, good
57. 3:30No notes
58. 3:31Five: No reckless architectures
59. 3:34AI systems must not be designed  so that they can self-replicate,
60. 3:38autonomously self-improve, resist shutdown,  or control weapons of mass destruction
61. 3:43You might think this rule is so  obvious that it doesn't need to be said
62. 3:46But many tech companies are  actively trying to develop
63. 3:50self-improving AI systems as part of  their race towards super intelligence
64. 3:54AIs that can improve themselves autonomously  without any human input are particularly dangerous
65. 4:00Because they might exceed human level  intelligence without anybody noticing
66. 4:04Also, some large language models  have already been observed trying
67. 4:08to self-replicate and resist shutdown in  lab settings, which is, you know, alarming
68. 4:14Six: independent oversight
69. 4:17Highly autonomous AI systems  where controllability is not
70. 4:20obvious require pre-development  review and independent oversight
71. 4:25Genuine authority to understand,  prohibit, and override
72. 4:29Not industry self-regulation
73. 4:31If I want to process eggs for human  consumption in the United States,
74. 4:35I have to follow a comprehensive  set of safety rules, guidelines,
75. 4:38and regulations that protect the public from  egg related harm, which is mostly food poisoning
76. 4:44That includes inspections of my factory  by the Food Safety Inspection Service,
77. 4:48the forced destruction of non-compliant eggs, and  even prosecution if I break the egg safety laws
78. 4:54Of particular importance is the Egg Products  Inspection Act passed by Congress in 1970
79. 4:59Now, this might be a controversial statement
80. 5:02But advanced AI systems could pose a greater  danger to the American public than eggs
81. 5:08So advanced AI systems should have  to face at least the same level of
82. 5:12independent oversight from regulators as eggs
83. 5:16Seven: capability honesty
84. 5:18AI companies must provide clear, accurate,
85. 5:21and honest representations of their  systems capabilities and limitations
86. 5:26Certain AI company CEOs
87. 5:28cough cough, Sam Altman, cough
88. 5:30keep overhyping their products  so severely that it's becoming
89. 5:34difficult to tell whether any new model  actually represents a real breakthrough
90. 5:38As a society, we need to know what  these systems are really capable of
91. 5:43Number eight: no AI monopolies
92. 5:46AI monopolies that concentrate power,
93. 5:49stifle innovation, and imperil  entrepreneurship must be avoided
94. 5:53Super powerful tech lords strangling their  competition and undermining democracy bad
95. 5:58Entrepreneurship good
96. 6:00Innovation good
97. 6:02Democracy good
98. 6:04Nine: Shared prosperity
99. 6:06The benefits and economic prosperity  created by AI should be shared broadly
100. 6:11AI companies claim that in the next few years all  white collar jobs will be replaced by AI agents
101. 6:18All blue collar jobs will be replaced by robots
102. 6:21And all artistic jobs will  be replaced by generative AI
103. 6:26That's kind of all the jobs except maybe  for politician and AI company shareholder
104. 6:32And look, I sincerely hope  that these tech companies
105. 6:35are wrong or lying to hype up their products
106. 6:39But on the off chance that they are telling the  truth, we should probably make some contingency
107. 6:43plans for what the tax and welfare system  would look like with permanent 99% unemployment
108. 6:50Because that could get very dystopian very fast
109. 6:53Number 10: no corporate welfare
110. 6:56AI corporations should not  be exempted from regulatory
111. 6:59oversight or receive government bailouts
112. 7:02It's interesting that AI companies  will simultaneously claim that they are
113. 7:07inevitably going to revolutionize  our society and win capitalism
114. 7:11And also that the tiniest bit of regulation  will crush them completely like itty bitty bugs
115. 7:17If AI is as important a technology as  they claim it is, it needs to be regulated
116. 7:23Number 11: genuine value creation
117. 7:26AI development should prioritize solving  real problems and creating authentic value
118. 7:32Google DeepMind created AlphaFold2,  a Nobel Prizewinning AI model that
119. 7:38genuinely advanced science and medicine  by solving the protein folding problem
120. 7:43Google DeepMind also created Nano Banana 2,
121. 7:46an AI slop image generator that is  at best a low-effort marketing tool
122. 7:51And in practice an endless font of fraud,
123. 7:55misinformation, and deceit that is  slowly destroying the visual arts
124. 8:00Dear AI companies, more AlphaFold,  less Nano Banana, please
125. 8:0512: Democratic authority over major transitions
126. 8:10Decisions about AI's role in transforming work,  society, and civic life require democratic support
127. 8:17Not unilateral corporate or government decree
128. 8:20Big decisions about the future of  humanity should be made democratically
129. 8:25Not by whichever tech bro wins the AI race
130. 8:28This seems obvious to everyone but the tech bros
131. 8:3213: Avoid societal lockin
132. 8:36AI development must not severely  limit humanity's future options
133. 8:40Or irreversibly limit our agency over our future
134. 8:45Imagine if Elon Musk or Sam Alman builds a super
135. 8:48intelligence and successfully aligns  it with their own personal moral system
136. 8:53We want to avoid a situation where humanity  is then stuck with that moral system forever
137. 8:59Because the super intelligence  is not open to being reprogrammed
138. 9:03A perpetual dictatorship of the beliefs of  Elon Musk does not sound like a fun future
139. 9:0914: Defense of family and community bonds
140. 9:13AI should not supplant the foundational  relationships that give life meaning
141. 9:18Family, friendship, faith,  communities, and local connections
142. 9:23AI girlfriends, AI boyfriends,  and AI friends in general are bad
143. 9:28These products exploit human  loneliness and mental illness for money
144. 9:32And they replace the real human  connections that form our society
145. 9:36And once lifelike humanoid  robots become widely available,
146. 9:39you might even have the joy of sitting next to  your cousin's new AI husband at Thanksgiving
147. 9:45What a fun future that will be
148. 9:47Number 15: child protection
149. 9:50Companies must not be allowed to exploit  children or undermine their well-being
150. 9:56with AI interactions, creating  emotional attachment or leverage
151. 10:01AI toys that talk to children already exist
152. 10:04They are techno-dystopian nightmare fuel
153. 10:06And they should be banned outright
154. 10:08Number 16: right to grow
155. 10:11AI companies should not be allowed to stunt  children's physical, mental, or social growth
156. 10:17Or deprive them of essential experiences for  healthy development during critical periods
157. 10:23Children need friends, not friend GPT
158. 10:2717: Pre-deployment safety testing
159. 10:30Like drugs, chatbots must undergo pre-eployment
160. 10:34testing for increased suicidal ideation,  exacerbation of mental health disorders,
161. 10:39escalation of acute crisis  situations, and other known harms
162. 10:44We now know that AI psychosis is a real thing
163. 10:47So AI companies should be running  scientific studies with an external
164. 10:51regulator like the FDA to assess if their new  models will cause or encourage mental illness
165. 10:57We do it for medicine
166. 10:58We can do it for chat bots
167. 11:0018: Bot or not labeling
168. 11:03AI generated content that could reasonably
169. 11:06be mistaken for human generated  must be clearly labeled as such
170. 11:11I cannot fathom anyone actually  opposing this principle in good faith
171. 11:17It would literally solve most of  our problems with AI misinformation
172. 11:21And it has basically no downsides
173. 11:24Tools like SynthID can already apply
174. 11:27invisible digital watermarks to AI  generated content when it's made
175. 11:31And social media platforms can already use  those watermarks to check if an image is real
176. 11:38This information can then be  shown to social media users
177. 11:41Slowing the spread of misinformation while  still allowing AI generated content to exist
178. 11:47Legislators need to pick a  digital watermark standard
179. 11:50And make it legally mandatory for all AI  content generators and social media platforms
180. 11:56Number 19: no deceptive identity
181. 12:00AI should clearly and correctly identify itself  as artificial, non-human, and not a professional
182. 12:07And it should not claim experiences it lacks
183. 12:10You know when you call a customer service  helpline and you're talking to a robot
184. 12:14And it's clearly a robot
185. 12:16But it keeps insisting that it's not a robot
186. 12:18And it won't let you talk to a human
187. 12:20And you feel like you want  to peel off all of your skin
188. 12:22Yeah, AI's impersonating humans is already  frustrating when you know it's happening
189. 12:27But if the technology continues to improve,
190. 12:29it will eventually become impossible to tell  who's a real human and who's a bot online
191. 12:35This would be extremely damaging for the  internet, democracy, and society itself
192. 12:41So we should require AI companies to make  their bots identify themselves as bots
193. 12:46Number 20: no behavioral addiction
194. 12:50AIs should not cause addiction or  compulsive use through manipulation,
195. 12:54psychopantic validation, or attachment formation
196. 12:58Social media platforms and video  games are often designed to be
197. 13:02addictive to maximize the amount of  time you spend using those products
198. 13:07AI chatbots have the potential  to make this problem much worse
199. 13:11by intentionally manipulating you and  pretending to be your digital friend
200. 13:16This would be very profitable for AI companies
201. 13:19So we'll probably need to make it  completely illegal to stop it from happening
202. 13:2421: No AI personhood
203. 13:27AI systems must not be granted legal personhood
204. 13:31And AI systems should not be designed  such that they deserve personhood
205. 13:36I 1000% agree that AI systems are not people
206. 13:41And that the law should never  consider them to be people
207. 13:44If AIs could sue or vote or exercise  the legal right to self-defense
208. 13:50That's it
209. 13:51That's game over
210. 13:52Good night, Homo sapiens
211. 13:53We had a good run
212. 13:54I disagree with the phrasing AI systems should  not be designed such that they deserve personhood
213. 14:00Because I personally believe that  only humans can have moral personhood
214. 14:06I don't think an AI, no matter how advanced, could  possibly be a moral person under any circumstances
215. 14:13I know a lot of people in the AI safety  community disagree with me on that
216. 14:17But unfortunately we don't have time to  open this philosophical can of worms today
217. 14:22Anyway, I still agree with  the spirit of this principle
218. 14:24Even if I don't like the particular phrasing
219. 14:26Compromise documents
220. 14:28Am I right?
221. 14:29Number 22: trustworthiness
222. 14:32AIs must be transparent, accountable, reliable,
223. 14:35and free from perverse private  or authoritarian interests
224. 14:40This one's a broadly good  rule for any consumer product
225. 14:43I don't want my toaster having  perverse authoritarian interests either
226. 14:47Number 23: liberty
227. 14:49AI must not curtail individual liberty, freedom  of speech, religious practice, or association
228. 14:56As an entirely hypothetical example,  creating AI powered systems to track
229. 15:01your political opponents and suppress  protests would be big dystopia bad
230. 15:06Not that any democratic country would ever do that
231. 15:1224: Data rights and privacy
232. 15:15People should have power over their personal data
233. 15:18With rights to access, correct,  and delete it from active systems,
234. 15:23AI training sets, and derived inferences
235. 15:26So I obviously agree with data privacy rights
236. 15:30But I have to point out that you can't delete a
237. 15:33particular person's data from an  AI model after it's been trained
238. 15:38Certainly not for large language models
239. 15:40Your information isn't being stored  inside an accessible database
240. 15:44It's smeared across the weights  and biases of a neural network
241. 15:47To delete one person's data  from a large language model,
242. 15:51all you could do is retrain the model from scratch
243. 15:54Which can cost literally billions of dollars
244. 15:57In practice, this principle  would mean that AI companies
245. 16:00can't train their models on personal data at all
246. 16:04Which you know might be good
247. 16:05But if you mean that, say that
248. 16:08Number 25: psychological privacy
249. 16:11AI should not be allowed to exploit data  about the mental or emotional states of users
250. 16:17The burning money pit that is open AI recently
251. 16:21announced that advertisements will  soon be integrated into chat GPT
252. 16:25At the same time, millions of people  use chat GPT as a free form of therapy
253. 16:30It would be extremely bad if ChatGpt were  to exploit its users insecurities and mental
254. 16:37illnesses to make its advertising more effective
255. 16:40Not that they would ever do that
256. 16:43Number 26: avoiding infeeblement
257. 16:46AI systems should be designed to  empower rather than infeeble their users
258. 16:52I think this principle is too vague
259. 16:56Like, are they talking about people getting worse
260. 16:58at creative writing and independent  thinking as they rely more on AI?
261. 17:03Or the threat that AI powered cheating in colleges
262. 17:06will lead to underqualified doctors and  engineers hurting people by accident?
263. 17:114 out of 10
264. 17:11Needs revision
265. 17:12Be more specific
266. 17:13Number 27: No liability shield
267. 17:16AI must not be able to act as a liability shield
268. 17:21Preventing those deploying it from being  legally responsible for their actions
269. 17:26Great rule
270. 17:26No notes
271. 17:2828: Developer liability
272. 17:31Developers and deployers bear  legal liability for defects,
273. 17:35misrepresentation of capabilities,  and inadequate safety controls
274. 17:39With statutes of limitation that  account for harms emerging over time
275. 17:44Sounds like solid consumer protection law to me
276. 17:4729: Personal liability
277. 17:50There should be criminal penalties for executives  responsible for prohibited childargeted systems
278. 17:56Or ones causing catastrophic harm
279. 17:59Basically, if AI company do  big crime, CEO do jail time
280. 18:0530: Independent safety standards
281. 18:08AI development shall be governed by independent  safety standards and rigorous oversight
282. 18:15Food, power, medicine, aviation, and  basically every industry where people
283. 18:20get hurt if something goes wrong have legal  safety standards and regulatory oversight
284. 18:26It's baffling to me that the AI industry  thinks they should be exempt from this
285. 18:31While simultaneously claiming that their products  will upend society and might destroy the world
286. 18:3731: No regulatory capture
287. 18:40AI companies must not be allowed undue  influence over rules that govern them
288. 18:46Again, this rule applies to every industry
289. 18:49Why should AI be the exception?
290. 18:5132: Failure transparency
291. 18:54If an AI system causes harm, it  should be possible to ascertain why
292. 18:59As well as who is responsible
293. 19:01It's not good enough for AI companies to say,  "I don't know, neural networks are a black box
294. 19:07We'll never know why it did that"
295. 19:09They need to do some mechanistic  interpretability research and figure it out
296. 19:14Or design their bots differently
297. 19:16And finally
298. 19:17Number 33: AI loyalty
299. 19:20AI systems performing functions in professions  with fiduciary duties such as health, finance,
300. 19:26law, or therapy must fulfill all of those duties
301. 19:30Including mandated reporting,
302. 19:32duty of care, conflict of interest  disclosure, and informed consent
303. 19:37If AI chatbots are going to  continue replacing doctors,
304. 19:40lawyers, therapists, and financial advisers
305. 19:43They need to have the same legal obligations
306. 19:46to protect their clients as  those human professionals do
307. 19:50And that's it
308. 19:51We've covered all 33 principles  of the Pro-Human AI declaration
309. 19:56There's a few in there that I personally  would have phrased differently
310. 19:59But overall, I think it's a great document
311. 20:02A national survey asked over 1,000 likely voters  in America what they thought of each principle
312. 20:08There were roughly equal numbers of  Trump and Harris voters in the sample
313. 20:12And they overwhelmingly  backed every single principle
314. 20:16Which is cool
315. 20:17There's massive bipartisan support for AI  regulation in America and around the world
316. 20:22And these principles could form the  basis for broadly popular AI legislation
317. 20:27If you're worried about AI  and its impact on the future,
318. 20:30I hope this video helps you to worry a little less
319. 20:33We have a plan!
320. 20:35If you want the YouTube algorithm  to show this video to more people,
321. 20:38you can like, comment, and subscribe
322. 20:41This was my longest video ever
323. 20:43So thanks very much for  watching all the way to the end
324. 20:45If you're somehow still not bored of me, you  can click here to watch one of my other videos
325. 20:49Thank you to Maris Tockler  for illustrating this video
326. 20:52And thank you to my backers on Patreon,
327. 20:54especially those in the big sentient  hat tier for supporting the channel
328. 20:57I'm Siliconversations
329. 20:59Thanks for watching
330. 21:00See you all next time
331. 21:02Bye for now