Transcriptly

Why AI Can't Replicate Human Creativity | Dr. Fei-Fei Li & Dr. Andrew Huberman

Huberman Lab Clips12:13Added Aug 31, 2026

Dr. Andrew Huberman and Dr. Fei-Fei Li discuss the limitations of artificial intelligence compared to human cognition, focusing on why AI cannot replicate ab...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Brain Function & AI Limits

  2. 00:00Let's go to a really far out there
  3. 00:02aspect of brain function that we know
  4. 00:04exists in humans, which is thoughts
  5. 00:08and creativity.
  6. 00:10Now, there are probably rules for
  7. 00:14thoughts and creativity. They're a
  8. 00:16little bit harder to tack down than um
  9. 00:19examples from the visual system, like if
  10. 00:21it's a tail and it's indoors, it's
  11. 00:23likely a cat. This kind of thing, but
  12. 00:24they're there. The rules are there. If
  13. 00:26you use Apple as an example,
  14. 00:28we could have gone from low level seeing
  15. 00:31an apple
  16. 00:32to midle seeing Apple always drop not
  17. 00:36fly off
  18. 00:37at the highest level. What is the
  19. 00:40equation that governs the Apple's
  20. 00:42movement
  21. 00:43right? So that's ascending to like a
  22. 00:45higher order more reductionist analysis.
  23. 00:48What do you think about the idea that
  24. 00:50while AI is indeed intelligent, it can
  25. 00:53do things that brains can do, maybe even
  26. 00:55well certainly things that individual
  27. 00:57human brains can't do. We know this by
  28. 00:59virtue of beating humans at chess and
  29. 01:01this sort of thing.
  30. 01:03The idea right now, as I understand it,
  31. 01:05is that AI is trained on the internet,
  32. 01:08images, discussions, videos, songs, but
  33. 01:13that's not all of human cognition,
  34. 01:16right? So, are there aspects of AI that
  35. 01:19are whether or not it's chat or it's
  36. 01:21claude or even the most powerful not yet
  37. 01:24released machine learning and and AI
  38. 01:27tools that don't have access to features
  39. 01:30of human brain function yet because
  40. 01:34they've never been uploaded to the
  41. 01:36internet, at least not in a way that the
  42. 01:38AI can pull out. So, for instance, you
  43. 01:41know, you could put a symphony there and
  44. 01:43it follows certain rules of music and
  45. 01:45mathematics and sound like that. That
  46. 01:46makes sense, but you have thoughts all
  47. 01:49day long and I have thoughts all day
  48. 01:50long that don't quite mesh with language
  49. 01:53in a way that I can just type them out
  50. 01:55on the internet. Stay with me here. I
  51. 01:57know this is a long question, but I feel
  52. 01:58like this is the one thing you are
  53. 02:00perfectly poised to answer, and I've
  54. 02:02been waiting to ask you this for
  55. 02:03[laughter]
  56. 02:04a year and a half since I saw you in
  57. 02:05Utah. In the world of art, we have this
  58. 02:08thing called abstraction, right? And
  59. 02:10occasionally, somebody will come up with
  60. 02:11a painting or a drawing that it doesn't
  61. 02:13look like anything specific. This
  62. 02:15happens in music, too, where you just
  63. 02:17feel something like there's like a
  64. 02:18fundamental rule or an emotion
  65. 02:20associated with it. Like they've tapped
  66. 02:22into some aspect of brain function, but
  67. 02:24you can't say what it is.
  68. 02:27I feel like this is the sort of thing
  69. 02:29that is complicated for AI or for me to
  70. 02:32understand how AI could do because you
  71. 02:33can put that piece of art into AI and
  72. 02:36say, you know, what fundamental feature
  73. 02:38of human uh experience does this reveal
  74. 02:41and it only has access to what's on the
  75. 02:43internet? So, h how can you capture a a
  76. 02:47complex constellation of feelings and
  77. 02:49experience with AI? That seems to be the
  78. 02:52gap for me. And I'm sure we'll get there
  79. 02:53with AI, but I'm not seeing from
  80. 02:57neuroscience to AI in in any kind of
  81. 02:59direct way. The same way we could
  82. 03:00ratchet through visual motion, sadness,
  83. 03:03happiness. You could pull out a lot of
  84. 03:04things, but it's hard to get to these
  85. 03:06higher order abstract representations
  86. 03:08that can't be spoken or written down or
  87. 03:10drawn. If I just say, give me your
  88. 03:12example of whatever nostalgia for your
  89. 03:16childhood home. You could write about
  90. 03:18it, but those are just words. It's not I
  91. 03:22can't understand your experience at a
  92. 03:24first person level.
  93. 03:25Totally. Andrew, I I know you put a lot
  94. 03:28of thoughts into this question and I I
  95. 03:31think it's a very important question and
  96. 03:34let's let's peel this one step at a
  97. 03:37time. First of all, TLDDR short answer
  98. 03:40is I agree with you that we do have to
  99. 03:44be very careful recognizing what AI can
  100. 03:47do, is likely to do, not conjecturing
  101. 03:51over a 100 years or or whatever. I
  102. 03:54recognize what you just said are these
  103. 03:57extremely nuanced personalized
  104. 04:01hard to characterize or not even
  105. 04:04captured human cognitive behaviors and
  106. 04:09because they were not captured they then
  107. 04:12they were not uploaded on the internet
  108. 04:15and we don't have today's AI doesn't
  109. 04:18have a way to do that. So when you call
  110. 04:22internet
  111. 04:24which is the source of AI's data, let's
  112. 04:27be very clear what is internet.
  113. 04:30Internet is not some random thing.

  114. How the Internet Trains AI

  115. 04:33Internet is the biggest collection of
  116. 04:36human behavior
  117. 04:38in multimodal forms. Let's break it down
  118. 04:42further. Internet has the world's
  119. 04:45population typing on it for many many at
  120. 04:49this point multiple decades. That typing
  121. 04:52is a sensing mechanism that captured
  122. 04:56everything from teenager chit chitchat
  123. 04:59all the way to deep scientific articles
  124. 05:02who dig digitized and get uploaded.
  125. 05:06Right? So that capturing human language
  126. 05:09is what internet is super good at.
  127. 05:13Then internet captures
  128. 05:16images. How? Because we now have digital
  129. 05:20cameras. That's so prevalent in
  130. 05:23smartphones and digital cameras. So that
  131. 05:25humans love taking photos from you know
  132. 05:30the cat in your house to selfies to
  133. 05:33beautiful you know BBC captured photos.
  134. 05:36Those also got uploaded in our digital
  135. 05:40sphere. On top of that, there's videos.
  136. 05:44Videos now has sound, has movements that
  137. 05:48also got uploaded to our digital sphere.
  138. 05:51On top of that, there's music. We're not
  139. 05:54even getting into the legal discussion
  140. 05:56of copyrights. But let's just table that
  141. 05:59aside. I'm just talking about the forms
  142. 06:00of data. The speeches and and singing
  143. 06:04and music and orchestra that also got
  144. 06:07uploaded into the digital sphere. So now
  145. 06:10we have created this humongous library
  146. 06:14of human knowledge in words, human
  147. 06:19behavior in videos, human expressions or
  148. 06:24even nature's ex whatever in in sound
  149. 06:28and now AI gets trained on that. That is
  150. 06:31why it's so powerful. This is why
  151. 06:34especially in the words front that AI
  152. 06:37can recognize patterns can can
  153. 06:40synthesize patterns because so much of
  154. 06:43this is already there. But the thing

  155. Uncaptured Human Thoughts

  156. 06:46that you just talked about that when
  157. 06:50let's say Picasso had that incredibly
  158. 06:53profound thought about that particular
  159. 06:57way of expressing
  160. 06:59that that portrait of the of the young
  161. 07:02woman that thought has never been
  162. 07:06captured.
  163. 07:07In fact, as neuroscientists, if I ask
  164. 07:10you which brain area did that thought
  165. 07:12come from, you don't know, right? Is it
  166. 07:15broka? Is it V1? Is it motor? Is it
  167. 07:20prefrontal? We don't know. Maybe it's
  168. 07:22diffused everywhere because that thought
  169. 07:26is so personalized, so special. You can
  170. 07:28call it creativity. You can call it
  171. 07:30emotion. You can call it whatever you
  172. 07:32want. You can call it cat 231. whatever
  173. 07:36name you can give it that thought is not
  174. 07:39captured therefore it's not on the
  175. 07:42internet therefore AI has not seen it
  176. 07:46so that is where humans still remain so
  177. 07:50unique but we also need to give credit
  178. 07:53to AI because AI has learned so many
  179. 07:56things it can combine information in
  180. 08:00highly [snorts] creative way did you

  181. AlphaGo & AI Creativity

  182. 08:02remember move 37
  183. 08:04this is Alph Go. Right.
  184. 08:06Right. Move 37 has symbolized AI's
  185. 08:09creativity.
  186. 08:10I think it's both true but can be taken
  187. 08:14out of context because that was a game
  188. 08:18when Alph Go was plainly so dull and in
  189. 08:21I think it's a third game out of the
  190. 08:24five games that Alph Go as a computer
  191. 08:28algorithm made a move that the human
  192. 08:31masters of Go never thought about.
  193. 08:35And that is an incredible move, right?
  194. 08:38Because it really humans collectively,
  195. 08:40these are the masters never thought
  196. 08:43about it. But if you really go deep into
  197. 08:46what AI did there, it was because first
  198. 08:51of all, goal is a highly mathematical
  199. 08:53game. It it has very clear mathematical
  200. 08:56objective, very clear mathematical rules
  201. 08:59in terms of move. So when AI having a
  202. 09:02bigger compute and um ways to retain how
  203. 09:07many moves it can uh it can remember it
  204. 09:10was able to do things that human brains
  205. 09:12don't typically do. So is that called
  206. 09:16creativity? I think it is but we do have
  207. 09:19to recognize that's a special kind of
  208. 09:22creativity. I was talking to a

  209. AI in Mathematics

  210. 09:26incredible mathematician of our time and
  211. 09:29I was asking him about the unsolved
  212. 09:31problem of mathematics and how AI can
  213. 09:34contribute to that and he was very
  214. 09:37positive. He said there are many
  215. 09:38problems in today's mathematics.
  216. 09:42As hard as they are, even as say a field
  217. 09:46mentalist, I probably have forgotten
  218. 09:50there are known methods in math that can
  219. 09:54solve these problem because I have a
  220. 09:56human brain. I don't remember I don't
  221. 09:59know all of math's, you know, solutions
  222. 10:02in the past hundreds of years. even if I
  223. 10:05were a a field medalist. So AI can help
  224. 10:09us to solve these problems. But as a
  225. 10:12mathematician, he was also telling me he
  226. 10:15said I don't know if AI can solve all of
  227. 10:18math problems because some of these math
  228. 10:21problems require solutions that have not
  229. 10:24been invented
  230. 10:26that will push creativity to a whole
  231. 10:28different level. And this is where you
  232. 10:31you know I'm we should be curious is it
  233. 10:34going to be a human creativity or AI
  234. 10:38would go through its iterations of uh of
  235. 10:41improvement and get to a point of
  236. 10:43creativity that humans don't have or is

  237. Hybrid Creativity & Emotion

  238. 10:45it a combined creativity. My current
  239. 10:48conjecture is hybrid is that humans
  240. 10:52working alongside AI would help us to
  241. 10:55solve these problems whose solutions
  242. 10:57have yet to be invented. And then what
  243. 11:00you said especially you touched on
  244. 11:03emotion is even more personalized. This
  245. 11:06is not necessarily logic. This is not
  246. 11:08necessarily deduct deductive reasoning.
  247. 11:11This is maybe Andrew you look at this
  248. 11:14cup and say it's a great cup. What if it
  249. 11:17evoked an emotion in me, a childhood
  250. 11:22moment that a gray cup might mean
  251. 11:25something that only me and my best
  252. 11:27friend share? That is a completely
  253. 11:31inaccessible
  254. 11:33piece of information in my brain that is
  255. 11:36never uploaded on the internet and no
  256. 11:39matter how mighty AI is today cannot
  257. 11:44access that. So that my reaction to this
  258. 11:47cup and potentially what I would do with
  259. 11:50it because of that piece of memory can
  260. 11:53be completely different. You can call it
  261. 11:56creativity, you can call it expression,
  262. 11:58you can call it storytelling, you can
  263. 12:00call it in many ways, but that's where
  264. 12:03it AI cannot access.
  265. 12:06[music]
  266. 12:12[music]