YouTube2Text

LA IA YA TIENE MARCA DE AGUA — Transcript

by DotTech-ES · 5,990 words · 920 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Open your editor and look at the last
  2. 0:01thing artificial intelligence wrote for
  3. 0:03you. That code carries a signature
  4. 0:04inside that you cannot see. Since
  5. 0:06Tuesday, August 11, it signs what comes
  6. 0:08out of the cloud, both text and code.
  7. 0:10Also via API, and also in Cloud Code.
  8. 0:13Models released from August 2nd onwards
  9. 0:15come signed from the factory. Those
  10. 0:17already out have until December 2nd,
  11. 0:19and no option to remove it has been
  12. 0:21announced. They operate at the model
  13. 0:23level, so there is no product or
  14. 0:25interface where you can escape it. It's
  15. 0:28doing this under Article 50 of the
  16. 0:30European AI Act, which mandates marking
  17. 0:32from August 2nd, with fines reaching 15
  18. 0:35million euros or 3%of global revenue,
  19. 0:37but it will be applied worldwide. And
  20. 0:40here the problem begins, because code
  21. 0:42is not prose; Seoul National University
  22. 0:44measured it, and with the mark applied,
  23. 0:47the success rate in code generation
  24. 0:49drops from 33%to 25%. Anthropic has not
  25. 0:52published a single piece of data—no
  26. 0:54detection rate, no false positives, no
  27. 0:56impact on code, not a single figure in
  28. 0:58the entire document—and neither has
  29. 1:00Decor. No one can read that mark—not
  30. 1:02you, not your company, and not the
  31. 1:04person reviewing your pull request.
  32. 1:05Today I am going to show you how text
  33. 1:07is marked from the inside, word by word
  34. 1:09. And why does something that goes
  35. 1:11unnoticed in a novel break your code in
  36. 1:13a function? Furthermore, OpenAI has had
  37. 1:16it finished and working with 99.9%
  38. 1:18effectiveness for over two years and
  39. 1:20decided not to release it. Today you
  40. 1:23will hear why, and it is not the reason
  41. 1:25they gave in public. The universities
  42. 1:27of Maryland and Harvard have
  43. 1:28demonstrated that you can take a text
  44. 1:30written by a person and trick a
  45. 1:31detector into flagging it as
  46. 1:33AI-generated. The reliability plummets
  47. 1:35to 1.3%. Those detectors don't just let
  48. 1:39culprits escape; they create them. And
  49. 1:41Stanford used English language exam
  50. 1:43essays written by people who do not
  51. 1:45have English as their native language.
  52. 1:4861%were marked as artificial
  53. 1:49intelligence, compared to only 5%for
  54. 1:51native speakers. Meanwhile, Twitch has
  55. 1:55signed up all its users to be part of
  56. 1:57Amazon's AI training. Livestreams,
  57. 1:59clips, and chats are enabled by default
  58. 2:02. At the 2026 Google Made by event, the
  59. 2:04Pixel 11s were presented. The Pro XL
  60. 2:06goes up by € 100, the foldable isn't
  61. 2:08coming to Spain, and the interesting
  62. 2:10part was in the watch. It's all about
  63. 2:13labels: who puts them on, who can read
  64. 2:14them, and who doesn't bother looking at
  65. 2:16them. And today we will answer one
  66. 2:18question. If the watermark is deleted
  67. 2:20by translating the text and nobody can
  68. 2:22read it anyway, why are they even
  69. 2:24putting it there? I'm Carlos Caballero
  70. 2:26and this is Top 10 News. Let’s get
  71. 2:28down to business. Let’s get to the
  72. 2:35story. On Tuesday, August 11, 2026,
  73. 2:38Anthropic published an article in their
  74. 2:41help center titled "How Claude marks
  75. 2:43AI-generated content" and it says
  76. 2:45Claude is going to mark what is written
  77. 2:48. There are two different systems and
  78. 2:50it's best to separate them. For files,
  79. 2:52they use C2PA. It is an open standard
  80. 2:55that embeds a history inside the
  81. 2:57cryptographically signed file itself:
  82. 2:59what program generated it, with which
  83. 3:01model, when, and what was done to it
  84. 3:02afterward. If you later crop it in
  85. 3:05Photoshop, that crop is also recorded;
  86. 3:07it’s not new and it’s not from
  87. 3:09Anthropic. Behind it are Adobe,
  88. 3:11Microsoft, Intel, Sony, the BBC, Amazon
  89. 3:13, Google, Meta, and OpenAI itself. And
  90. 3:16it’s already out there. The world's
  91. 3:18first camera to have it from the
  92. 3:20factory was the Leica M11-P at the end
  93. 3:22of 2023. Today, nine Sony models have
  94. 3:24it, several from Canon, and the Pixel
  95. 3:2610 has it enabled by default. Nikon
  96. 3:29enabled it on the Z6 III and had to
  97. 3:31suspend the service when
  98. 3:33vulnerabilities were found. And that is
  99. 3:35a fundamental problem. It's metadata,
  100. 3:37it's attached to the outside; take a
  101. 3:40screenshot and 100%of the trail
  102. 3:42disappears. That's why text goes down a
  103. 3:45different path. In text, the mark goes
  104. 3:47inside the text itself. They describe
  105. 3:49it this way. An imperceptible watermark
  106. 3:51within the text itself; you won't see
  107. 3:53it, and it doesn't change the meaning,
  108. 3:56quality, or readability. Three things
  109. 3:58that almost no one is saying right now.
  110. 4:01First, it’s not optional. There’s
  111. 4:03no button, no setting, no checkbox.
  112. 4:06It’s at the model level, not the
  113. 4:07product level. So there is no surface
  114. 4:10from which you can escape. Second,
  115. 4:12it’s not just for Europe. The
  116. 4:14watermark will be applied to the
  117. 4:16outputs of compatible models wherever
  118. 4:18Claude is offered, that is, worldwide.
  119. 4:21Anthropic models released starting
  120. 4:23August 2nd come marked from day one.
  121. 4:27The ones already out there, no; those
  122. 4:28are in progress, and the company says
  123. 4:30it is taking advantage of the
  124. 4:32transition period, which in principle
  125. 4:33will be starting December 2nd. The
  126. 4:36motive is European, but the
  127. 4:37implementation is global, because
  128. 4:39setting up two systems—one marked for
  129. 4:41Europe and one clean for the rest—is
  130. 4:43more expensive than setting up just one
  131. 4:45. So if you are in Bogotá, Mexico City
  132. 4:47, or Málaga, you have the same
  133. 4:49watermark as someone in Germany. And
  134. 4:52the third one, which goes directly to
  135. 4:54you, enters the API and enters Clockco,
  136. 4:57the complete list, the API platform,
  137. 4:59Cloud, Clockco, Cloud Cowork and Cloud
  138. 5:01Tag, plus the deployments on Amazon Web
  139. 5:04Services, Google Cloud and Microsoft
  140. 5:06Foundry. If you call the API thinking
  141. 5:09it returns raw text, it actually
  142. 5:11returns signed text, and if you ask
  143. 5:13Clock Code to write a function for you,
  144. 5:15that function will come out watermarked
  145. 5:17. And let's look at this with a
  146. 5:19step-by-step example. A language model
  147. 5:21doesn't write sentences; it writes the
  148. 5:24next word, again and again. And at each
  149. 5:26step, what it has in front of it isn't
  150. 5:28a word, it's a list of candidates, each
  151. 5:30with its own probability. Suppose, "My
  152. 5:33favorite tropical fruits are mango and
  153. 5:35..." What do you have in front of you?
  154. 5:37Banana with a 20%probability, papaya
  155. 5:40with 15%, pineapple with 12%, and then
  156. 5:42all the others sharing the rest. Now
  157. 5:45the watermark comes in. A secret key
  158. 5:47splits that list in two. Half of the
  159. 5:49words become "awarded" and the other
  160. 5:51half do not. And whenever the model can
  161. 5:53do so without ruining the sentence, it
  162. 5:55chooses one of the awarded ones. Notice
  163. 5:58that nothing is lost here. If instead
  164. 6:00of banana it says papaya, the sentence
  165. 6:02remains correct and still means the
  166. 6:04same thing. You don't notice a thing.
  167. 6:07But whoever has the key does, because
  168. 6:09they count how many times an awarded
  169. 6:10word appeared throughout 1,000
  170. 6:12decisions in a row. And if the number
  171. 6:14is well above what chance would dictate
  172. 6:16, they know that text was written by a
  173. 6:18machine. That is the watermark. It is
  174. 6:21not a hidden stamp; it is a statistical
  175. 6:23deviation. Now we have to explain what
  176. 6:26logits are. When the model finishes
  177. 6:29processing your sentence, the last
  178. 6:30thing it produces isn't words, it's a
  179. 6:32list of numbers, one for each word in
  180. 6:34its vocabulary. And that vocabulary
  181. 6:37consists of tens of thousands of
  182. 6:38entries. Those numbers are the logits,
  183. 6:41raw scores. And I say raw because they
  184. 6:45aren't probabilities as such yet; they
  185. 6:47can even be negative, they don't add up
  186. 6:49to one, and they mean nothing
  187. 6:50separately, only mattering when
  188. 6:52compared to each other. To convert them
  189. 6:55into probabilities, a function often
  190. 6:57called softmax is applied, which does
  191. 6:59two things: it makes them all positive
  192. 7:01and normalizes them so they add up to
  193. 7:03one. And that is when banana becomes 20
  194. 7:06%. Besides, think about it, it could be
  195. 7:10like an election count. Logits are raw
  196. 7:13votes, probability is the percentage,
  197. 7:16and the written word is the actual seat
  198. 7:19. Well, the classic watermark method
  199. 7:22acts before the scrutiny. They add a
  200. 7:25few votes to the winning words, not
  201. 7:27many. Just enough for them to win when
  202. 7:30it's a close call. And that changes the
  203. 7:32outcome. And that is exactly the
  204. 7:35problem with the code. When there is
  205. 7:38only one word that compiles, you are
  206. 7:39giving those extra votes to what
  207. 7:40doesn't compile. Or worse, you are
  208. 7:44flagging code as truly watermarked when
  209. 7:47it isn't. And now, the mathematical
  210. 7:50concept that explains why this works in
  211. 7:52prose and will fail in code. Entropy is
  212. 7:55a concept that appears in physics,
  213. 7:57mathematics, and information theory.
  214. 7:59And here we are going to use it in the
  215. 8:01Shannon sense. Entropy measures the
  216. 8:03uncertainty present in a probability
  217. 8:05distribution. If all probability is
  218. 8:08concentrated on a single option, the
  219. 8:10entropy is zero. There is nothing to
  220. 8:12decide. The more spread out it is among
  221. 8:15many similar options, the higher it is,
  222. 8:17measured in bits. One bit of entropy is
  223. 8:21equivalent to choosing between two
  224. 8:23equally likely options—that is, how
  225. 8:25many different paths the sentence can
  226. 8:27take in the next step. If I say "happy,
  227. 8:30" then "birthday," you already know the
  228. 8:32word that follows; entropy is zero. But
  229. 8:35if I say "yesterday I bought a," there
  230. 8:37are 1,000 things that could fit;
  231. 8:39entropy is high. And watermarking only
  232. 8:41has room when there is entropy. It
  233. 8:44needs several valid options to choose
  234. 8:46from. If there is only one, there is
  235. 8:48nothing to bias, we cannot
  236. 8:50differentiate. And code, code is low
  237. 8:52entropy, and not by accident. It is low
  238. 8:56entropy because we make it that way on
  239. 8:58purpose. We have built it that way for
  240. 9:00years. Code thrives on repetition.
  241. 9:03Design patterns, naming conventions,
  242. 9:06templates, boilerplate code, frameworks
  243. 9:09. A "for" loop is written the same way
  244. 9:12everywhere, a getter is written the
  245. 9:14same, a singleton is written the same
  246. 9:16in Java as in your project, and we do
  247. 9:18it precisely so another human can
  248. 9:19recognize it at a glance. That
  249. 9:23repetition, which in engineering is a
  250. 9:25virtue, in information theory is low
  251. 9:26entropy; and furthermore, there are
  252. 9:28positions where it's not that there are
  253. 9:29few options, it's that there is only
  254. 9:31one. A closing parenthesis, a type that
  255. 9:35has to match, a variable name you
  256. 9:37declared three lines earlier. So, two
  257. 9:40things happen here. And the second is
  258. 9:41more interesting than the first, but
  259. 9:43let's start with the first. The first
  260. 9:45is that it breaks. If you push the
  261. 9:47model toward a winning word when only
  262. 9:50one compiles, you aren't adding a
  263. 9:52signature or a watermark; you are
  264. 9:54injecting an error. And the other point
  265. 9:56, which is the truly interesting one,
  266. 9:58is that it becomes homogenized because
  267. 10:00the mark doesn't choose at random; it
  268. 10:02always skews toward the same options.
  269. 10:05And if at every point there are three
  270. 10:07valid names for a variable and the
  271. 10:08model always pulls the winning one, you
  272. 10:10end up with a model that always writes
  273. 10:12the same code. That, stated like that,
  274. 10:15doesn't have to be bad; on the contrary
  275. 10:17, a team with their style guides and
  276. 10:19linter have spent years trying to do
  277. 10:22exactly that: get everyone to write the
  278. 10:24same way. A model that always solves
  279. 10:27the same thing in the same way is a
  280. 10:28more predictable, more deterministic
  281. 10:30model, and in production, that is worth
  282. 10:32money. The issue is that nobody decided
  283. 10:35this. It happens as a rebound effect
  284. 10:38from a mechanism installed for
  285. 10:40something else, and nobody has measured
  286. 10:42if what comes out of your style guide
  287. 10:44is a quirk of the random number
  288. 10:45generator, a person's habit, or where
  289. 10:47it comes from and how much it breaks,
  290. 10:49yet it is measured. The paper is called
  291. 10:52"Who wrote this code?" Watermarks for
  292. 10:55code generation. And the important part
  293. 10:58: they didn't test it with just any
  294. 10:59model; they tested it with StarCoder,
  295. 11:01which is an open model specialized in
  296. 11:02generating code—in other words, the
  297. 11:04favorable scenario. They measured it
  298. 11:06against three benchmarks. The first is
  299. 11:08called HumanEval, created by OpenAI,
  300. 11:10consisting of 164 hand-written
  301. 11:12programming problems at a technical
  302. 11:14interview level. The second is called
  303. 11:17MBPP, Mostly Basic Programming Problems
  304. 11:19. It was created by Google Research in
  305. 11:232021 with 974 beginner-level exercises
  306. 11:25—that is, string manipulation,
  307. 11:26iterating through lists, mathematical
  308. 11:28operations—and what they measure is
  309. 11:30exactly this: out of every 100 problems
  310. 11:32, how many does the model write correct
  311. 11:34code for on the first try without
  312. 11:36second chances. That model without a
  313. 11:39watermark solved 33 problems out of 100
  314. 11:42in HumanEval, and with the mark on, 29.
  315. 11:45In MBPP, it drops from 38 to 34. That
  316. 11:5029 is with the watermark calibrated
  317. 11:52loosely, on purpose, so that the
  318. 11:54quality drops little, and a loose mark
  319. 11:56is barely detected. In MBPP, at that
  320. 11:59strength, the detector performs little
  321. 12:01better than a coin toss. When they set
  322. 12:03it to a strength that can actually be
  323. 12:06read, HumanEval drops from 33 to 25 and
  324. 12:08MBPP from 38 to 24. One out of every
  325. 12:11three problems that previously turned
  326. 12:13out well now turns out wrong just
  327. 12:15because of the marking, and it has a
  328. 12:17known solution. The same paper proposes
  329. 12:19marking only where there is entropy and
  330. 12:21skipping positions where there is only
  331. 12:23one valid option. With that, they
  332. 12:25recover almost everything. From 29 up
  333. 12:27to 32 out of 33, so it can be done well
  334. 12:30and has been known for at least 2 years
  335. 12:33. Those figures were not provided by
  336. 12:35Anropy. They come from the classic
  337. 12:38method, splitting vocabulary into
  338. 12:40greenlisted and non-greenlisted and
  339. 12:42shifting logits. That method distorts
  340. 12:45the model's distribution by design,
  341. 12:47which is why it breaks the code. But
  342. 12:50there is another way to do it,
  343. 12:52published by Google in Nature in
  344. 12:53October 2024. The paper is called
  345. 12:56Scalable Watermarking for Identifying
  346. 12:58Large Language Model Outputs. SynthID
  347. 13:01does not touch the logits. It leaves
  348. 13:04the scoring intact; instead, it
  349. 13:06organizes a tournament, and it works
  350. 13:07like this. With a secret key and the
  351. 13:10four previous words of the text, it
  352. 13:12generates a function that assigns each
  353. 13:14candidate word a zero or a one. It is
  354. 13:18random, but random in a way that is
  355. 13:19reproducible with the same key and
  356. 13:21context. The result is always the same.
  357. 13:24And that is what allows it to be
  358. 13:25verified later. And that is where the
  359. 13:27tournament begins. The candidates are
  360. 13:29paired off two by two. The one with the
  361. 13:31one wins. If they both have the same,
  362. 13:33it’s a coin toss. The winners are
  363. 13:35paired again, but now with a different
  364. 13:37function and a different key. And again
  365. 13:39, and again, and again. 30 rounds. 30
  366. 13:42distinct functions. The one that
  367. 13:44survives is the word that gets written,
  368. 13:46and that is the mathematical
  369. 13:47description, not what the machine
  370. 13:49executes. A 30-round tournament would
  371. 13:52require over a billion participants for
  372. 13:54each word. But what the code does is
  373. 13:57solve the entire tournament with a
  374. 13:58single formula. Instead of drawing a
  375. 14:01billion candidates to compete, it
  376. 14:03directly readjusts the probabilities of
  377. 14:06the entire vocabulary 30 times in a row
  378. 14:08and picks one word—same result, same
  379. 14:11exact distribution, but instead of
  380. 14:13exponential cost, it costs 30
  381. 14:15multiplications. Google measures it.
  382. 14:17The speed impact is half a percentage
  383. 14:19point per token. And how is it detected
  384. 14:21? Well, this is what is truly elegant.
  385. 14:24The detector doesn't need the model,
  386. 14:26your prompt, or anything else; just the
  387. 14:29text and the keys. It recalculates all
  388. 14:31those zeros and ones by looking at the
  389. 14:33average. If the text is written by a
  390. 14:35person or any other model, the average
  391. 14:37comes out to 0.5. Pure chance. If it
  392. 14:40went through the tournament, the
  393. 14:41winners carry extra ones, and the
  394. 14:43average goes up. With that alone, it
  395. 14:45works. And with the Bayesian score,
  396. 14:49which knows how much signal fit into
  397. 14:51each round, it detects 85%of texts with
  398. 14:54a 1%false positive rate, compared to 73
  399. 14:56%for the previous best method. Now,
  400. 15:00about that claim that the watermark
  401. 15:02doesn't distort the text, it's worth
  402. 15:04understanding what exactly that means
  403. 15:06because it's less than it seems. What
  404. 15:08it is demonstrating is this. If you
  405. 15:11repeat the process with every possible
  406. 15:13secret key, one after another, and you
  407. 15:15took the average of all those times,
  408. 15:16each word would come out with the exact
  409. 15:18probability it was supposed to have.
  410. 15:21Not one more, not one less. The problem
  411. 15:23is that you don't use all the keys, you
  412. 15:25use one. It’s like a rigged die where
  413. 15:28if you test every possible trick and
  414. 15:30calculate the average, it gives
  415. 15:32perfectly balanced results. But you
  416. 15:34aren’t playing with every trick,
  417. 15:36you’re playing with one, and with
  418. 15:38that one, it isn’t balanced. With a
  419. 15:40fixed key, each word comes out biased,
  420. 15:42and the next one too. And they are all
  421. 15:44biased in the same direction because
  422. 15:46they share the same key. So the text
  423. 15:49you receive isn't the one you would
  424. 15:51have received without the watermark;
  425. 15:53it’s another, probably just as good,
  426. 15:54but another—and I'm not saying this,
  427. 15:56the authors themselves say it. They
  428. 15:59call what they’ve demonstrated the
  429. 16:01weakest version of the property. The
  430. 16:04strong version would be to guarantee
  431. 16:06that the entire text, not each word
  432. 16:08separately, has the same probability of
  433. 16:10appearing. And they don't demonstrate
  434. 16:12that. The theorem doesn't prove that
  435. 16:15quality doesn't drop; 20 million Gemini
  436. 16:17responses compared in production prove
  437. 16:19it, with a difference in user
  438. 16:21satisfaction of one hundredth of a
  439. 16:22percentage point. So in summary, there
  440. 16:26is a way to watermark that breaks the
  441. 16:28code and a way that doesn't. The first
  442. 16:31is easy; the second is published,
  443. 16:33peer-reviewed, and running on Gemini
  444. 16:36since 2024. And which one does
  445. 16:38Anthropic use? We don't know; they
  446. 16:40haven't said. I’m not saying their
  447. 16:43watermark degrades your code; I’m
  448. 16:45saying both ways exist, both have been
  449. 16:47measured, and they haven't shown which
  450. 16:49one they chose. Second thing, and this
  451. 16:52affects everyone, whether they code or
  452. 16:54not: the watermark doesn't distinguish
  453. 16:56between writing text and fixing a comma
  454. 16:58. People often use Claude to proofread,
  455. 17:01translate, summarize, or convert files.
  456. 17:04And the output can carry a Claude
  457. 17:06watermark, even if the ideas, text, or
  458. 17:09data came from another source. You
  459. 17:12write something that is entirely yours,
  460. 17:14you put it through to fix typos, and it
  461. 17:16returns it marked as processed by
  462. 17:18Claude—the same signal as if the
  463. 17:19machine had written it from scratch.
  464. 17:22And the third, the big one. For
  465. 17:25Anthropic, no detector exists, not
  466. 17:27public, not restricted, not even on a
  467. 17:29waiting list. The only thing that
  468. 17:31exists is a sentence in the document.
  469. 17:34We will share details about the
  470. 17:35detection mechanisms in upcoming
  471. 17:37technical documentation. Without a date
  472. 17:39, Zaris and Hipar from the Claude team
  473. 17:42expanded on this week on X. This is
  474. 17:44part of working with the European AI
  475. 17:46Act. Other labs are adding similar
  476. 17:49labeling. It is difficult to identify
  477. 17:52AI-generated text and this gives people
  478. 17:54better tools to do so. We are also
  479. 17:57going to release a text detection API
  480. 17:58that you will be able to use yourself.
  481. 18:01And this sounds good, but it does not
  482. 18:02exist. They have deployed the
  483. 18:04transmitter without the receiver. There
  484. 18:06is a mark and it will be inside the
  485. 18:08text, being there. And nobody outside
  486. 18:10of Anthropic can read it, not you, not
  487. 18:12your company, not the person reviewing
  488. 18:14your code, and the others, because the
  489. 18:16regulation binds them equally. And
  490. 18:18let's go one by one. On July 31st, the
  491. 18:21European Commission published the list
  492. 18:23of signatories to the Transparency Code
  493. 18:25, some 190 organizations, 82 signed the
  494. 18:27section for providers of generative
  495. 18:29artificial intelligence systems.
  496. 18:32Anthropic signed and is the one that
  497. 18:33has gone the furthest. It marks text
  498. 18:36and code in all its products, worldwide
  499. 18:38. Google signed, it has SynthID, the
  500. 18:40one from the tournament we just saw. It
  501. 18:43does mark text, but only in the Gemini
  502. 18:45app and website, not in its API. And
  503. 18:48its detector remains restricted with a
  504. 18:50waiting list. For now, only journalists
  505. 18:52and media professionals are testing it.
  506. 18:55OpenAI signed, it has Content
  507. 18:56Credentials for images and a public
  508. 18:58verifier that works today. You upload a
  509. 19:01photo or audio and it tells you if it
  510. 19:03carries a signature. Meta signed. Its
  511. 19:06public work is on C2PA for images and
  512. 19:08on labeling, but regarding text
  513. 19:09labeling, nothing has been announced.
  514. 19:12And X.AI does not appear on the list.
  515. 19:15They neither signed nor is there record
  516. 19:16that they rejected it. They are simply
  517. 19:18not there. In total, six sign, two mark
  518. 19:21text, and none let you verify it. A
  519. 19:25public text detector for the entire
  520. 19:27industry, but with that on the table,
  521. 19:29the question is inevitable. If they
  522. 19:32don't provide any figures and no one
  523. 19:34can verify it, what is this for? Well,
  524. 19:37to answer that, we have to go back 2
  525. 19:39years in time. On August 4, 2024, The
  526. 19:42Wall Street Journal published a report
  527. 19:44titled: There is a tool to catch
  528. 19:46students cheating with ChatGPT. OpenAI
  529. 19:50has not released it, and based on the
  530. 19:52company's internal documents, OpenAI
  531. 19:55had its text watermark finished and
  532. 19:57working with 99.9%effectiveness when
  533. 19:59there is enough generated text. They
  534. 20:03had it ready for about a year and had
  535. 20:05been debating it internally for about 2
  536. 20:07years and decided not to release it.
  537. 20:09And from there, two phrases arise. One
  538. 20:11is written in an internal document. The
  539. 20:13other is leaked to a newspaper by
  540. 20:15someone supposedly familiar with the
  541. 20:17project. The first says the company can
  542. 20:20no longer claim it can't watermark text
  543. 20:22, now that they know doing so doesn't
  544. 20:24worsen the responses. According to them
  545. 20:27, everything is fine. In other words,
  546. 20:29we know it works, we know it doesn't
  547. 20:31break the product, and we no longer
  548. 20:32have an excuse. They wrote that for
  549. 20:35themselves, never thinking it would
  550. 20:37leave the office as an internal memo.
  551. 20:40And the other, when asked what stage
  552. 20:42the project is at, the answer is it's
  553. 20:44just a matter of pushing a button, 2
  554. 20:46years without pushing it. And now a
  555. 20:48European regulation is coming. And why?
  556. 20:50Well, this is very interesting because
  557. 20:52the leak wasn't even necessary. OpenAI
  558. 20:54published it themselves that same day
  559. 20:56on their website. The article is called
  560. 20:58"Understanding the origin of what we
  561. 21:00see and hear on the internet." It's
  562. 21:03from May 7, 2024, and on August 4, they
  563. 21:05added an update with a section titled "
  564. 21:08Text Watermarking," and the link is
  565. 21:10still in the description. They give two
  566. 21:13reasons. The first is technical. They
  567. 21:15say the watermark holds up well against
  568. 21:18local manipulation, such as
  569. 21:19paraphrasing, but struggles with what
  570. 21:21they call global manipulation. And they
  571. 21:24give three examples. Running the text
  572. 21:26through a translator, rewriting it with
  573. 21:28another generative model, or asking the
  574. 21:30model to insert a special character
  575. 21:31between every two words and then
  576. 21:33deleting it. And the second reason, and
  577. 21:36this is a direct quote, is that it
  578. 21:38could stigmatize the use of artificial
  579. 21:40intelligence as a useful writing tool
  580. 21:41for non-native English speakers. The
  581. 21:45real reason is in two surveys OpenAI
  582. 21:47conducted in April 2023; the first was
  583. 21:50for the general public in several
  584. 21:52countries to see if people wanted such
  585. 21:54a tool, and it came back "yes" by a 4-
  586. 21:57to-1 margin. The second was for their
  587. 22:00own ChatGPT users, and the numbers
  588. 22:02changed radically. 69%thought that
  589. 22:05technology to detect copies would end
  590. 22:08up causing false accusations, and
  591. 22:10nearly 30%said they would use ChatGPT
  592. 22:12less. If OpenAI implemented watermarks
  593. 22:15and a competitor didn't. And that is
  594. 22:17the calculation. The public wants it 4
  595. 22:20to 1. 30%of customers would leave, and
  596. 22:23the one who decides is the one who
  597. 22:25bills, the one who pays. There is no
  598. 22:28report, methodology, or number of
  599. 22:30respondents for these two surveys. They
  600. 22:33are internal documents that the
  601. 22:34newspaper saw. On August 2, Article 50
  602. 22:37of the European Artificial Intelligence
  603. 22:39Act came into effect. It mandates
  604. 22:41watermarking text in a machine-readable
  605. 22:43format. Fines reach up to 15 million
  606. 22:46euros or 3%of global turnover,
  607. 22:48whichever is higher. Systems that were
  608. 22:50already on the market before August 2nd
  609. 22:52have until December 2nd, 2026, to
  610. 22:54comply. But there are two more things.
  611. 22:57First, Article 50 itself states that
  612. 22:59the marking obligation does not apply
  613. 23:02when systems perform, and I quote, a
  614. 23:04standard editing support function or
  615. 23:07when they do not substantially alter
  616. 23:09the input data provided by the deployer
  617. 23:12or its semantics. Where is the line?
  618. 23:15Because that is key. The European
  619. 23:18Commission published the final
  620. 23:20guidelines on July 20th of this year,
  621. 23:22and the test isn't how much text the
  622. 23:24machine touched; the test is whether
  623. 23:26the output meaning is the input meaning
  624. 23:29, and let's look at concrete cases. A
  625. 23:32spell and grammar checker like
  626. 23:33Grammarly doesn't need to be marked;
  627. 23:35it's the textbook example of what's
  628. 23:37excluded. Correcting errors in a text
  629. 23:40you wrote yourself doesn't either—
  630. 23:42same case—but rewriting an email of
  631. 23:44yours to sound more formal is a gray
  632. 23:46area, and probably not, because you're
  633. 23:49changing the register, not the message.
  634. 23:52But if, from "more formal," you start
  635. 23:54adding a paragraph with arguments you
  636. 23:56hadn't included, then yes. Writing that
  637. 23:59entire email based on three bullet
  638. 24:01points you provide must be marked.
  639. 24:03There is no editing there; there is
  640. 24:05generation. The system doesn't alter
  641. 24:07your input data; it replaces it.
  642. 24:09Restructuring a paper from top to
  643. 24:11bottom or summarizing it must be marked
  644. 24:13. The guidelines state this expressly:
  645. 24:16substantive rewrites and summaries do
  646. 24:18not fall under the exception. In other
  647. 24:21words, correcting a typo doesn't change
  648. 24:23what you said, but summarizing does.
  649. 24:25And writing from scratch has nothing
  650. 24:28you said to compare it to. So, in the
  651. 24:31case we saw earlier, you pass your text
  652. 24:34to Claude to remove typos and it
  653. 24:36returns it marked. It’s exactly the
  654. 24:39case Europe excludes, yet Anthropic
  655. 24:42marks it anyway. It marks more than
  656. 24:44what Europe is demanding and justifies
  657. 24:47it by invoking Europe. And the second
  658. 24:49thing, which is good, is to look at
  659. 24:51what the law itself says about whether
  660. 24:52this is technically possible. It says
  661. 24:55that providers shall ensure their
  662. 24:57technical solutions are effective,
  663. 25:00interoperable, robust, and reliable,
  664. 25:02and I quote verbatim, "insofar as this
  665. 25:05is technically feasible, taking into
  666. 25:07account generally acknowledged
  667. 25:09state-of-the-art." Brussels forces you
  668. 25:12to mark and, in the same sentence,
  669. 25:13admits that maybe the technology isn't
  670. 25:15there, that it's not possible. And
  671. 25:17listen, one quick thing, all this I'm
  672. 25:19telling you, the dates, the figures,
  673. 25:21the direct quotes from the document,
  674. 25:22take me many hours of verification, one
  675. 25:24by one. If it helps you and you want
  676. 25:27this to continue, you can become a
  677. 25:28member of the vault here on YouD. In
  678. 25:31exchange, I give you the full dossier
  679. 25:32each week, but don't see it that way.
  680. 25:34See it more as patronage so that we can
  681. 25:36keep going here week after week. If you
  682. 25:38can't or don't want to, it's no problem
  683. 25:40. So let's continue. Because while
  684. 25:43people debate if the watermark lasts,
  685. 25:45if it gets erased, and if anyone will
  686. 25:46be able to read it someday, something
  687. 25:48is already happening, and it's not a
  688. 25:49laboratory problem. Since there are no
  689. 25:52watermarks and the one that exists
  690. 25:54can't be read by anyone, what's being
  691. 25:55used instead are detectors, programs
  692. 25:57that look at a text and guess from the
  693. 25:59outside. And the day one of those
  694. 26:01programs makes a mistake, it stops
  695. 26:03being a technical debate and becomes
  696. 26:05the problem of a real person with a
  697. 26:06name and surname. On January 28, 2026,
  698. 26:10a New York court issued a ruling that
  699. 26:12has barely had any echo, and I think it
  700. 26:14should. A student from Adelphi
  701. 26:17University on Long Island turned in a
  702. 26:19history paper. Turnitin, the world's
  703. 26:22most widespread anti-plagiarism program
  704. 26:24, marked it as 100%AI-generated text.
  705. 26:26100%. Come on. The university declared
  706. 26:30him guilty based simply on that number.
  707. 26:33He ran the paper through two other
  708. 26:35detectors, GPTZero and GPT-4. Both said
  709. 26:38zero, and so he sued. The judge
  710. 26:42annulled the offense, ordered the
  711. 26:43record cleared and the sanction
  712. 26:45rescinded, and wrote that the
  713. 26:47accusation and its denial lacked a
  714. 26:48valid basis and were, quote, devoid of
  715. 26:50reason. Now, why did this happen to him
  716. 26:53? Well, let's go to Stanford, 2023. The
  717. 26:57paper is called "GPT detectors are
  718. 26:59biased against non-native English
  719. 27:00writers." They take seven commercial
  720. 27:04detectors used by universities and give
  721. 27:06them 91 essays from the TOEFL exam, the
  722. 27:09English proficiency test, published
  723. 27:11before 2020, meaning before GPT existed
  724. 27:13, written by people. Well, 61.22%false
  725. 27:17positives: six out of 10 human essays
  726. 27:19flagged as written by a machine. And
  727. 27:22there's more. 18 of those 91 were
  728. 27:24marked by all seven detectors at once,
  729. 27:27unanimously, and 89 out of 91 by at
  730. 27:29least one. And what happens with essays
  731. 27:32by Americans? For native speakers, 5.19
  732. 27:35%, meaning from 61 down to 5. The best
  733. 27:38part is what they did next. They took
  734. 27:40those same TOEFL essays and asked
  735. 27:42ChatGPT to literally improve the word
  736. 27:44choice to make it sound more native.
  737. 27:47False positives dropped from 61 to 11,
  738. 27:49and then they did the reverse. They
  739. 27:52took the native essays and asked
  740. 27:54ChatGPT to simplify the vocabulary as
  741. 27:56if written by a non-native. False
  742. 27:59positives rose from 5 to 56. And what
  743. 28:02does that mean? That those detectors
  744. 28:04aren't looking for whether a machine
  745. 28:05wrote it; they look at whether the text
  746. 28:07is predictable. The technique is called
  747. 28:09perplexity, and the authors explain it
  748. 28:11like this. How surprised a language
  749. 28:13model is when trying to guess the next
  750. 28:15word in a sentence. And a person who is
  751. 28:18learning a language uses fewer words.
  752. 28:21Simpler constructions, more repetitive
  753. 28:23structures; they write in a more
  754. 28:25mechanically predictable way, so the
  755. 28:27detector confuses them with a machine.
  756. 28:30But would the same thing happen today,
  757. 28:32because everything I just told you is
  758. 28:34from 2023, and things have changed. On
  759. 28:37June 29 of this year, 2026, the Vrije
  760. 28:39Universiteit Brussel published an
  761. 28:41evaluation of four current detectors:
  762. 28:44GPTZero, Paraphrasing, CopyLeaks, and
  763. 28:46Turnitin. 160 documents, and the human
  764. 28:50texts are graduate student papers from
  765. 28:52non-native English speakers written
  766. 28:54before 2019, so there is no possible
  767. 28:56contamination. Zero false positives.
  768. 28:59Across all four, not a single one. And
  769. 29:01there is a second paper from Charles
  770. 29:03University in Prague and the University
  771. 29:04of Oslo, presented this year at the
  772. 29:05European Chapter of the Association for
  773. 29:07Computational Linguistics. They ran a
  774. 29:09current commercial detector through
  775. 29:10those same 91 essays from the Stanford
  776. 29:12study. False positives dropped from an
  777. 29:15average of 61 to 23.1, so the bias has
  778. 29:17been greatly reduced. The tool works
  779. 29:20better, that is true, but things must
  780. 29:22be stated as they are. In that same
  781. 29:25study, that same detector failed zero
  782. 29:27times with the essays from native
  783. 29:28speakers. Zero versus 23. The bias has
  784. 29:32been reduced, not disappeared. The
  785. 29:34problem has been turned on its head
  786. 29:36because in the Brussels study, they
  787. 29:37also included pure machine text
  788. 29:38generated with the most advanced tools
  789. 29:40available when they set up the
  790. 29:41experiment. In that mode, it was GPT-4o
  791. 29:44, and it didn't get a single one right.
  792. 29:46It gave 100%of the machine texts a
  793. 29:48score between 0 and 20%. GPTZero and
  794. 29:51CopyLeaks didn't detect a single text
  795. 29:53with confidence either; at most, they
  796. 29:55reached the doubt range of 20 to 40%in
  797. 29:57three out of every 10 and one out of
  798. 29:59every four. Only one worked,
  799. 30:02Paraphrasing, with a 65%strict accuracy
  800. 30:05rate. In other words, the detectors
  801. 30:07have stopped accusing the innocent
  802. 30:09because they have stopped detecting.
  803. 30:11They’ve gone from unfair to useless,
  804. 30:13and no one has told the universities
  805. 30:15that are still paying for the licenses.
  806. 30:18But there is one more piece that will
  807. 30:20affect us directly, which closes the
  808. 30:23loop on what it means to go through an
  809. 30:26AI detector. There’s a study
  810. 30:28presented last year at the AI ethics
  811. 30:30conference by the AAAI and the ACM.
  812. 30:33It’s called AI self-preference in
  813. 30:36algorithmic hiring. 2,245 real resumes
  814. 30:40written by people, nine models acting
  815. 30:43as recruiters. And what it shows is
  816. 30:46that when a language model writes the
  817. 30:49resume, that model prefers texts
  818. 30:51written by itself between two-thirds
  819. 30:53and over 80%of the time. A candidate
  820. 30:57using the same model as the evaluator
  821. 30:59has between a 23 and 60%higher chance
  822. 31:02of passing the filter compared to an
  823. 31:04equally qualified person who wrote it
  824. 31:06by hand. In other words, you get
  825. 31:10flagged for using AI, and the filter
  826. 31:11that decides if you get an interview
  827. 31:13discards you for not using it. And here
  828. 31:17I pass the ball to you because I think
  829. 31:19this is open for much debate and there
  830. 31:20are many studies. What do you use it
  831. 31:23for? For editing, for writing, for
  832. 31:26programming. And if it were proven
  833. 31:28tomorrow that you use it, would you
  834. 31:29stop? Do you think it’s unfair? And
  835. 31:32now that we know anyone can apply the
  836. 31:34label and anyone can read it, let’s
  837. 31:36go straight to the bursts. The National
  838. 31:39Police arrested a man in Murcia on
  839. 31:40August 11th who was obtaining digital
  840. 31:42certificates in other people's names.
  841. 31:44The method: a real-time deepfake to
  842. 31:46pass video identification, the kind
  843. 31:48that asks you to move your head and
  844. 31:50show your ID. They caught him because
  845. 31:52his real face appeared on screen for a
  846. 31:54second. And it wasn't just the face; he
  847. 31:56had set up a homemade lighting system
  848. 31:58to simulate the document's holograms in
  849. 32:00front of his face. With that, he
  850. 32:02tricked the check ensuring the ID
  851. 32:03you're showing is physical and not a
  852. 32:05photocopy. He made 38 attempts,
  853. 32:07affecting more than 30 people. The
  854. 32:09generated video lagged slightly behind
  855. 32:11the actual movement. At one point
  856. 32:13during the process, it failed, and
  857. 32:14that’s when his real face appeared.
  858. 32:16Twitch announced that content from the
  859. 32:18platform is used to train artificial
  860. 32:19intelligence models. Amazon’s move on
  861. 32:22Wednesday included livestreams, clips,
  862. 32:24highlights, and the text and images
  863. 32:26from your channel, even the chat. There
  864. 32:29is a box to opt out. It’s in settings
  865. 32:31, under security and privacy at the
  866. 32:32very bottom; it comes enabled by
  867. 32:34default. If it were optional, no one
  868. 32:35would turn it on. That sentence was
  869. 32:37said by the chief product officer of
  870. 32:38Twitch. Google unveiled the Pixel 11 on
  871. 32:40August 12th in New York, and the entire
  872. 32:41lineup is seeing a € 100 price
  873. 32:43increase in Europe. In the United
  874. 32:44States, the price remains unchanged.
  875. 32:46The foldable, the Pixel 11 Pro Fold,
  876. 32:47will only be sold in France within
  877. 32:49Europe. As for the rest, there's a new
  878. 32:51chip and better cameras. Those who have
  879. 32:53tested it say it performs just like
  880. 32:54last year's model, but what really
  881. 32:56matters was in the watch. It's called
  882. 32:58Health Guardian, and it estimates your
  883. 33:00blood pressure cuffless by passively
  884. 33:02measuring your pulse and movement
  885. 33:04patterns all month long, without you
  886. 33:05doing a thing. And there is a second
  887. 33:08feature that analyzes weeks of
  888. 33:09physiological data to track your
  889. 33:11insulin resistance. This week marks 30
  890. 33:13years since the browser that won the
  891. 33:15war was released. Internet Explorer 3
  892. 33:17was released on August 13, 1996, and I
  893. 33:20bring it up for one reason: it was the
  894. 33:22first commercial browser to support CSS
  895. 33:25stylesheets, which today shapes
  896. 33:27absolutely every page you see in
  897. 33:29wonderful detail. Actually, CSS wasn't
  898. 33:31even a standard yet. The first
  899. 33:33specification from the W3C was
  900. 33:35published on December 17th of that same
  901. 33:37year, 4 months later. In other words,
  902. 33:40Microsoft implemented a standard that
  903. 33:41didn't exist yet, and they only got it
  904. 33:43half right. Colors and fonts worked,
  905. 33:45but internal margins didn't, and
  906. 33:47neither did borders. The people who
  907. 33:49wrote CSS summarized it like this. It
  908. 33:52supports almost everything related to
  909. 33:53color and text well, but doesn't
  910. 33:54implement much of the box model. 30
  911. 33:56years later, that box model is the
  912. 33:58first thing anyone learns when they
  913. 34:00start building websites. And CSS is the
  914. 34:02yardstick by which all browsers in the
  915. 34:04world are measured. If you enjoyed the
  916. 34:07video, you know what to do: like,
  917. 34:08subscribe, and share it with anyone you
  918. 34:10think needs to hear this. I'm Carlos
  919. 34:12Caballero, and this has been Top 10
  920. 34:14News. See you later.

About this transcript

This page contains the full transcript of LA IA YA TIENE MARCA DE AGUA by DotTech-ES, generated from the public captions YouTube serves with the video. The transcript has 5,990 words across 920 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.