LA IA YA TIENE MARCA DE AGUA — Transcript
Full transcript
- 0:00Open your editor and look at the last
- 0:01thing artificial intelligence wrote for
- 0:03you. That code carries a signature
- 0:04inside that you cannot see. Since
- 0:06Tuesday, August 11, it signs what comes
- 0:08out of the cloud, both text and code.
- 0:10Also via API, and also in Cloud Code.
- 0:13Models released from August 2nd onwards
- 0:15come signed from the factory. Those
- 0:17already out have until December 2nd,
- 0:19and no option to remove it has been
- 0:21announced. They operate at the model
- 0:23level, so there is no product or
- 0:25interface where you can escape it. It's
- 0:28doing this under Article 50 of the
- 0:30European AI Act, which mandates marking
- 0:32from August 2nd, with fines reaching 15
- 0:35million euros or 3%of global revenue,
- 0:37but it will be applied worldwide. And
- 0:40here the problem begins, because code
- 0:42is not prose; Seoul National University
- 0:44measured it, and with the mark applied,
- 0:47the success rate in code generation
- 0:49drops from 33%to 25%. Anthropic has not
- 0:52published a single piece of data—no
- 0:54detection rate, no false positives, no
- 0:56impact on code, not a single figure in
- 0:58the entire document—and neither has
- 1:00Decor. No one can read that mark—not
- 1:02you, not your company, and not the
- 1:04person reviewing your pull request.
- 1:05Today I am going to show you how text
- 1:07is marked from the inside, word by word
- 1:09. And why does something that goes
- 1:11unnoticed in a novel break your code in
- 1:13a function? Furthermore, OpenAI has had
- 1:16it finished and working with 99.9%
- 1:18effectiveness for over two years and
- 1:20decided not to release it. Today you
- 1:23will hear why, and it is not the reason
- 1:25they gave in public. The universities
- 1:27of Maryland and Harvard have
- 1:28demonstrated that you can take a text
- 1:30written by a person and trick a
- 1:31detector into flagging it as
- 1:33AI-generated. The reliability plummets
- 1:35to 1.3%. Those detectors don't just let
- 1:39culprits escape; they create them. And
- 1:41Stanford used English language exam
- 1:43essays written by people who do not
- 1:45have English as their native language.
- 1:4861%were marked as artificial
- 1:49intelligence, compared to only 5%for
- 1:51native speakers. Meanwhile, Twitch has
- 1:55signed up all its users to be part of
- 1:57Amazon's AI training. Livestreams,
- 1:59clips, and chats are enabled by default
- 2:02. At the 2026 Google Made by event, the
- 2:04Pixel 11s were presented. The Pro XL
- 2:06goes up by € 100, the foldable isn't
- 2:08coming to Spain, and the interesting
- 2:10part was in the watch. It's all about
- 2:13labels: who puts them on, who can read
- 2:14them, and who doesn't bother looking at
- 2:16them. And today we will answer one
- 2:18question. If the watermark is deleted
- 2:20by translating the text and nobody can
- 2:22read it anyway, why are they even
- 2:24putting it there? I'm Carlos Caballero
- 2:26and this is Top 10 News. Let’s get
- 2:28down to business. Let’s get to the
- 2:35story. On Tuesday, August 11, 2026,
- 2:38Anthropic published an article in their
- 2:41help center titled "How Claude marks
- 2:43AI-generated content" and it says
- 2:45Claude is going to mark what is written
- 2:48. There are two different systems and
- 2:50it's best to separate them. For files,
- 2:52they use C2PA. It is an open standard
- 2:55that embeds a history inside the
- 2:57cryptographically signed file itself:
- 2:59what program generated it, with which
- 3:01model, when, and what was done to it
- 3:02afterward. If you later crop it in
- 3:05Photoshop, that crop is also recorded;
- 3:07it’s not new and it’s not from
- 3:09Anthropic. Behind it are Adobe,
- 3:11Microsoft, Intel, Sony, the BBC, Amazon
- 3:13, Google, Meta, and OpenAI itself. And
- 3:16it’s already out there. The world's
- 3:18first camera to have it from the
- 3:20factory was the Leica M11-P at the end
- 3:22of 2023. Today, nine Sony models have
- 3:24it, several from Canon, and the Pixel
- 3:2610 has it enabled by default. Nikon
- 3:29enabled it on the Z6 III and had to
- 3:31suspend the service when
- 3:33vulnerabilities were found. And that is
- 3:35a fundamental problem. It's metadata,
- 3:37it's attached to the outside; take a
- 3:40screenshot and 100%of the trail
- 3:42disappears. That's why text goes down a
- 3:45different path. In text, the mark goes
- 3:47inside the text itself. They describe
- 3:49it this way. An imperceptible watermark
- 3:51within the text itself; you won't see
- 3:53it, and it doesn't change the meaning,
- 3:56quality, or readability. Three things
- 3:58that almost no one is saying right now.
- 4:01First, it’s not optional. There’s
- 4:03no button, no setting, no checkbox.
- 4:06It’s at the model level, not the
- 4:07product level. So there is no surface
- 4:10from which you can escape. Second,
- 4:12it’s not just for Europe. The
- 4:14watermark will be applied to the
- 4:16outputs of compatible models wherever
- 4:18Claude is offered, that is, worldwide.
- 4:21Anthropic models released starting
- 4:23August 2nd come marked from day one.
- 4:27The ones already out there, no; those
- 4:28are in progress, and the company says
- 4:30it is taking advantage of the
- 4:32transition period, which in principle
- 4:33will be starting December 2nd. The
- 4:36motive is European, but the
- 4:37implementation is global, because
- 4:39setting up two systems—one marked for
- 4:41Europe and one clean for the rest—is
- 4:43more expensive than setting up just one
- 4:45. So if you are in Bogotá, Mexico City
- 4:47, or Málaga, you have the same
- 4:49watermark as someone in Germany. And
- 4:52the third one, which goes directly to
- 4:54you, enters the API and enters Clockco,
- 4:57the complete list, the API platform,
- 4:59Cloud, Clockco, Cloud Cowork and Cloud
- 5:01Tag, plus the deployments on Amazon Web
- 5:04Services, Google Cloud and Microsoft
- 5:06Foundry. If you call the API thinking
- 5:09it returns raw text, it actually
- 5:11returns signed text, and if you ask
- 5:13Clock Code to write a function for you,
- 5:15that function will come out watermarked
- 5:17. And let's look at this with a
- 5:19step-by-step example. A language model
- 5:21doesn't write sentences; it writes the
- 5:24next word, again and again. And at each
- 5:26step, what it has in front of it isn't
- 5:28a word, it's a list of candidates, each
- 5:30with its own probability. Suppose, "My
- 5:33favorite tropical fruits are mango and
- 5:35..." What do you have in front of you?
- 5:37Banana with a 20%probability, papaya
- 5:40with 15%, pineapple with 12%, and then
- 5:42all the others sharing the rest. Now
- 5:45the watermark comes in. A secret key
- 5:47splits that list in two. Half of the
- 5:49words become "awarded" and the other
- 5:51half do not. And whenever the model can
- 5:53do so without ruining the sentence, it
- 5:55chooses one of the awarded ones. Notice
- 5:58that nothing is lost here. If instead
- 6:00of banana it says papaya, the sentence
- 6:02remains correct and still means the
- 6:04same thing. You don't notice a thing.
- 6:07But whoever has the key does, because
- 6:09they count how many times an awarded
- 6:10word appeared throughout 1,000
- 6:12decisions in a row. And if the number
- 6:14is well above what chance would dictate
- 6:16, they know that text was written by a
- 6:18machine. That is the watermark. It is
- 6:21not a hidden stamp; it is a statistical
- 6:23deviation. Now we have to explain what
- 6:26logits are. When the model finishes
- 6:29processing your sentence, the last
- 6:30thing it produces isn't words, it's a
- 6:32list of numbers, one for each word in
- 6:34its vocabulary. And that vocabulary
- 6:37consists of tens of thousands of
- 6:38entries. Those numbers are the logits,
- 6:41raw scores. And I say raw because they
- 6:45aren't probabilities as such yet; they
- 6:47can even be negative, they don't add up
- 6:49to one, and they mean nothing
- 6:50separately, only mattering when
- 6:52compared to each other. To convert them
- 6:55into probabilities, a function often
- 6:57called softmax is applied, which does
- 6:59two things: it makes them all positive
- 7:01and normalizes them so they add up to
- 7:03one. And that is when banana becomes 20
- 7:06%. Besides, think about it, it could be
- 7:10like an election count. Logits are raw
- 7:13votes, probability is the percentage,
- 7:16and the written word is the actual seat
- 7:19. Well, the classic watermark method
- 7:22acts before the scrutiny. They add a
- 7:25few votes to the winning words, not
- 7:27many. Just enough for them to win when
- 7:30it's a close call. And that changes the
- 7:32outcome. And that is exactly the
- 7:35problem with the code. When there is
- 7:38only one word that compiles, you are
- 7:39giving those extra votes to what
- 7:40doesn't compile. Or worse, you are
- 7:44flagging code as truly watermarked when
- 7:47it isn't. And now, the mathematical
- 7:50concept that explains why this works in
- 7:52prose and will fail in code. Entropy is
- 7:55a concept that appears in physics,
- 7:57mathematics, and information theory.
- 7:59And here we are going to use it in the
- 8:01Shannon sense. Entropy measures the
- 8:03uncertainty present in a probability
- 8:05distribution. If all probability is
- 8:08concentrated on a single option, the
- 8:10entropy is zero. There is nothing to
- 8:12decide. The more spread out it is among
- 8:15many similar options, the higher it is,
- 8:17measured in bits. One bit of entropy is
- 8:21equivalent to choosing between two
- 8:23equally likely options—that is, how
- 8:25many different paths the sentence can
- 8:27take in the next step. If I say "happy,
- 8:30" then "birthday," you already know the
- 8:32word that follows; entropy is zero. But
- 8:35if I say "yesterday I bought a," there
- 8:37are 1,000 things that could fit;
- 8:39entropy is high. And watermarking only
- 8:41has room when there is entropy. It
- 8:44needs several valid options to choose
- 8:46from. If there is only one, there is
- 8:48nothing to bias, we cannot
- 8:50differentiate. And code, code is low
- 8:52entropy, and not by accident. It is low
- 8:56entropy because we make it that way on
- 8:58purpose. We have built it that way for
- 9:00years. Code thrives on repetition.
- 9:03Design patterns, naming conventions,
- 9:06templates, boilerplate code, frameworks
- 9:09. A "for" loop is written the same way
- 9:12everywhere, a getter is written the
- 9:14same, a singleton is written the same
- 9:16in Java as in your project, and we do
- 9:18it precisely so another human can
- 9:19recognize it at a glance. That
- 9:23repetition, which in engineering is a
- 9:25virtue, in information theory is low
- 9:26entropy; and furthermore, there are
- 9:28positions where it's not that there are
- 9:29few options, it's that there is only
- 9:31one. A closing parenthesis, a type that
- 9:35has to match, a variable name you
- 9:37declared three lines earlier. So, two
- 9:40things happen here. And the second is
- 9:41more interesting than the first, but
- 9:43let's start with the first. The first
- 9:45is that it breaks. If you push the
- 9:47model toward a winning word when only
- 9:50one compiles, you aren't adding a
- 9:52signature or a watermark; you are
- 9:54injecting an error. And the other point
- 9:56, which is the truly interesting one,
- 9:58is that it becomes homogenized because
- 10:00the mark doesn't choose at random; it
- 10:02always skews toward the same options.
- 10:05And if at every point there are three
- 10:07valid names for a variable and the
- 10:08model always pulls the winning one, you
- 10:10end up with a model that always writes
- 10:12the same code. That, stated like that,
- 10:15doesn't have to be bad; on the contrary
- 10:17, a team with their style guides and
- 10:19linter have spent years trying to do
- 10:22exactly that: get everyone to write the
- 10:24same way. A model that always solves
- 10:27the same thing in the same way is a
- 10:28more predictable, more deterministic
- 10:30model, and in production, that is worth
- 10:32money. The issue is that nobody decided
- 10:35this. It happens as a rebound effect
- 10:38from a mechanism installed for
- 10:40something else, and nobody has measured
- 10:42if what comes out of your style guide
- 10:44is a quirk of the random number
- 10:45generator, a person's habit, or where
- 10:47it comes from and how much it breaks,
- 10:49yet it is measured. The paper is called
- 10:52"Who wrote this code?" Watermarks for
- 10:55code generation. And the important part
- 10:58: they didn't test it with just any
- 10:59model; they tested it with StarCoder,
- 11:01which is an open model specialized in
- 11:02generating code—in other words, the
- 11:04favorable scenario. They measured it
- 11:06against three benchmarks. The first is
- 11:08called HumanEval, created by OpenAI,
- 11:10consisting of 164 hand-written
- 11:12programming problems at a technical
- 11:14interview level. The second is called
- 11:17MBPP, Mostly Basic Programming Problems
- 11:19. It was created by Google Research in
- 11:232021 with 974 beginner-level exercises
- 11:25—that is, string manipulation,
- 11:26iterating through lists, mathematical
- 11:28operations—and what they measure is
- 11:30exactly this: out of every 100 problems
- 11:32, how many does the model write correct
- 11:34code for on the first try without
- 11:36second chances. That model without a
- 11:39watermark solved 33 problems out of 100
- 11:42in HumanEval, and with the mark on, 29.
- 11:45In MBPP, it drops from 38 to 34. That
- 11:5029 is with the watermark calibrated
- 11:52loosely, on purpose, so that the
- 11:54quality drops little, and a loose mark
- 11:56is barely detected. In MBPP, at that
- 11:59strength, the detector performs little
- 12:01better than a coin toss. When they set
- 12:03it to a strength that can actually be
- 12:06read, HumanEval drops from 33 to 25 and
- 12:08MBPP from 38 to 24. One out of every
- 12:11three problems that previously turned
- 12:13out well now turns out wrong just
- 12:15because of the marking, and it has a
- 12:17known solution. The same paper proposes
- 12:19marking only where there is entropy and
- 12:21skipping positions where there is only
- 12:23one valid option. With that, they
- 12:25recover almost everything. From 29 up
- 12:27to 32 out of 33, so it can be done well
- 12:30and has been known for at least 2 years
- 12:33. Those figures were not provided by
- 12:35Anropy. They come from the classic
- 12:38method, splitting vocabulary into
- 12:40greenlisted and non-greenlisted and
- 12:42shifting logits. That method distorts
- 12:45the model's distribution by design,
- 12:47which is why it breaks the code. But
- 12:50there is another way to do it,
- 12:52published by Google in Nature in
- 12:53October 2024. The paper is called
- 12:56Scalable Watermarking for Identifying
- 12:58Large Language Model Outputs. SynthID
- 13:01does not touch the logits. It leaves
- 13:04the scoring intact; instead, it
- 13:06organizes a tournament, and it works
- 13:07like this. With a secret key and the
- 13:10four previous words of the text, it
- 13:12generates a function that assigns each
- 13:14candidate word a zero or a one. It is
- 13:18random, but random in a way that is
- 13:19reproducible with the same key and
- 13:21context. The result is always the same.
- 13:24And that is what allows it to be
- 13:25verified later. And that is where the
- 13:27tournament begins. The candidates are
- 13:29paired off two by two. The one with the
- 13:31one wins. If they both have the same,
- 13:33it’s a coin toss. The winners are
- 13:35paired again, but now with a different
- 13:37function and a different key. And again
- 13:39, and again, and again. 30 rounds. 30
- 13:42distinct functions. The one that
- 13:44survives is the word that gets written,
- 13:46and that is the mathematical
- 13:47description, not what the machine
- 13:49executes. A 30-round tournament would
- 13:52require over a billion participants for
- 13:54each word. But what the code does is
- 13:57solve the entire tournament with a
- 13:58single formula. Instead of drawing a
- 14:01billion candidates to compete, it
- 14:03directly readjusts the probabilities of
- 14:06the entire vocabulary 30 times in a row
- 14:08and picks one word—same result, same
- 14:11exact distribution, but instead of
- 14:13exponential cost, it costs 30
- 14:15multiplications. Google measures it.
- 14:17The speed impact is half a percentage
- 14:19point per token. And how is it detected
- 14:21? Well, this is what is truly elegant.
- 14:24The detector doesn't need the model,
- 14:26your prompt, or anything else; just the
- 14:29text and the keys. It recalculates all
- 14:31those zeros and ones by looking at the
- 14:33average. If the text is written by a
- 14:35person or any other model, the average
- 14:37comes out to 0.5. Pure chance. If it
- 14:40went through the tournament, the
- 14:41winners carry extra ones, and the
- 14:43average goes up. With that alone, it
- 14:45works. And with the Bayesian score,
- 14:49which knows how much signal fit into
- 14:51each round, it detects 85%of texts with
- 14:54a 1%false positive rate, compared to 73
- 14:56%for the previous best method. Now,
- 15:00about that claim that the watermark
- 15:02doesn't distort the text, it's worth
- 15:04understanding what exactly that means
- 15:06because it's less than it seems. What
- 15:08it is demonstrating is this. If you
- 15:11repeat the process with every possible
- 15:13secret key, one after another, and you
- 15:15took the average of all those times,
- 15:16each word would come out with the exact
- 15:18probability it was supposed to have.
- 15:21Not one more, not one less. The problem
- 15:23is that you don't use all the keys, you
- 15:25use one. It’s like a rigged die where
- 15:28if you test every possible trick and
- 15:30calculate the average, it gives
- 15:32perfectly balanced results. But you
- 15:34aren’t playing with every trick,
- 15:36you’re playing with one, and with
- 15:38that one, it isn’t balanced. With a
- 15:40fixed key, each word comes out biased,
- 15:42and the next one too. And they are all
- 15:44biased in the same direction because
- 15:46they share the same key. So the text
- 15:49you receive isn't the one you would
- 15:51have received without the watermark;
- 15:53it’s another, probably just as good,
- 15:54but another—and I'm not saying this,
- 15:56the authors themselves say it. They
- 15:59call what they’ve demonstrated the
- 16:01weakest version of the property. The
- 16:04strong version would be to guarantee
- 16:06that the entire text, not each word
- 16:08separately, has the same probability of
- 16:10appearing. And they don't demonstrate
- 16:12that. The theorem doesn't prove that
- 16:15quality doesn't drop; 20 million Gemini
- 16:17responses compared in production prove
- 16:19it, with a difference in user
- 16:21satisfaction of one hundredth of a
- 16:22percentage point. So in summary, there
- 16:26is a way to watermark that breaks the
- 16:28code and a way that doesn't. The first
- 16:31is easy; the second is published,
- 16:33peer-reviewed, and running on Gemini
- 16:36since 2024. And which one does
- 16:38Anthropic use? We don't know; they
- 16:40haven't said. I’m not saying their
- 16:43watermark degrades your code; I’m
- 16:45saying both ways exist, both have been
- 16:47measured, and they haven't shown which
- 16:49one they chose. Second thing, and this
- 16:52affects everyone, whether they code or
- 16:54not: the watermark doesn't distinguish
- 16:56between writing text and fixing a comma
- 16:58. People often use Claude to proofread,
- 17:01translate, summarize, or convert files.
- 17:04And the output can carry a Claude
- 17:06watermark, even if the ideas, text, or
- 17:09data came from another source. You
- 17:12write something that is entirely yours,
- 17:14you put it through to fix typos, and it
- 17:16returns it marked as processed by
- 17:18Claude—the same signal as if the
- 17:19machine had written it from scratch.
- 17:22And the third, the big one. For
- 17:25Anthropic, no detector exists, not
- 17:27public, not restricted, not even on a
- 17:29waiting list. The only thing that
- 17:31exists is a sentence in the document.
- 17:34We will share details about the
- 17:35detection mechanisms in upcoming
- 17:37technical documentation. Without a date
- 17:39, Zaris and Hipar from the Claude team
- 17:42expanded on this week on X. This is
- 17:44part of working with the European AI
- 17:46Act. Other labs are adding similar
- 17:49labeling. It is difficult to identify
- 17:52AI-generated text and this gives people
- 17:54better tools to do so. We are also
- 17:57going to release a text detection API
- 17:58that you will be able to use yourself.
- 18:01And this sounds good, but it does not
- 18:02exist. They have deployed the
- 18:04transmitter without the receiver. There
- 18:06is a mark and it will be inside the
- 18:08text, being there. And nobody outside
- 18:10of Anthropic can read it, not you, not
- 18:12your company, not the person reviewing
- 18:14your code, and the others, because the
- 18:16regulation binds them equally. And
- 18:18let's go one by one. On July 31st, the
- 18:21European Commission published the list
- 18:23of signatories to the Transparency Code
- 18:25, some 190 organizations, 82 signed the
- 18:27section for providers of generative
- 18:29artificial intelligence systems.
- 18:32Anthropic signed and is the one that
- 18:33has gone the furthest. It marks text
- 18:36and code in all its products, worldwide
- 18:38. Google signed, it has SynthID, the
- 18:40one from the tournament we just saw. It
- 18:43does mark text, but only in the Gemini
- 18:45app and website, not in its API. And
- 18:48its detector remains restricted with a
- 18:50waiting list. For now, only journalists
- 18:52and media professionals are testing it.
- 18:55OpenAI signed, it has Content
- 18:56Credentials for images and a public
- 18:58verifier that works today. You upload a
- 19:01photo or audio and it tells you if it
- 19:03carries a signature. Meta signed. Its
- 19:06public work is on C2PA for images and
- 19:08on labeling, but regarding text
- 19:09labeling, nothing has been announced.
- 19:12And X.AI does not appear on the list.
- 19:15They neither signed nor is there record
- 19:16that they rejected it. They are simply
- 19:18not there. In total, six sign, two mark
- 19:21text, and none let you verify it. A
- 19:25public text detector for the entire
- 19:27industry, but with that on the table,
- 19:29the question is inevitable. If they
- 19:32don't provide any figures and no one
- 19:34can verify it, what is this for? Well,
- 19:37to answer that, we have to go back 2
- 19:39years in time. On August 4, 2024, The
- 19:42Wall Street Journal published a report
- 19:44titled: There is a tool to catch
- 19:46students cheating with ChatGPT. OpenAI
- 19:50has not released it, and based on the
- 19:52company's internal documents, OpenAI
- 19:55had its text watermark finished and
- 19:57working with 99.9%effectiveness when
- 19:59there is enough generated text. They
- 20:03had it ready for about a year and had
- 20:05been debating it internally for about 2
- 20:07years and decided not to release it.
- 20:09And from there, two phrases arise. One
- 20:11is written in an internal document. The
- 20:13other is leaked to a newspaper by
- 20:15someone supposedly familiar with the
- 20:17project. The first says the company can
- 20:20no longer claim it can't watermark text
- 20:22, now that they know doing so doesn't
- 20:24worsen the responses. According to them
- 20:27, everything is fine. In other words,
- 20:29we know it works, we know it doesn't
- 20:31break the product, and we no longer
- 20:32have an excuse. They wrote that for
- 20:35themselves, never thinking it would
- 20:37leave the office as an internal memo.
- 20:40And the other, when asked what stage
- 20:42the project is at, the answer is it's
- 20:44just a matter of pushing a button, 2
- 20:46years without pushing it. And now a
- 20:48European regulation is coming. And why?
- 20:50Well, this is very interesting because
- 20:52the leak wasn't even necessary. OpenAI
- 20:54published it themselves that same day
- 20:56on their website. The article is called
- 20:58"Understanding the origin of what we
- 21:00see and hear on the internet." It's
- 21:03from May 7, 2024, and on August 4, they
- 21:05added an update with a section titled "
- 21:08Text Watermarking," and the link is
- 21:10still in the description. They give two
- 21:13reasons. The first is technical. They
- 21:15say the watermark holds up well against
- 21:18local manipulation, such as
- 21:19paraphrasing, but struggles with what
- 21:21they call global manipulation. And they
- 21:24give three examples. Running the text
- 21:26through a translator, rewriting it with
- 21:28another generative model, or asking the
- 21:30model to insert a special character
- 21:31between every two words and then
- 21:33deleting it. And the second reason, and
- 21:36this is a direct quote, is that it
- 21:38could stigmatize the use of artificial
- 21:40intelligence as a useful writing tool
- 21:41for non-native English speakers. The
- 21:45real reason is in two surveys OpenAI
- 21:47conducted in April 2023; the first was
- 21:50for the general public in several
- 21:52countries to see if people wanted such
- 21:54a tool, and it came back "yes" by a 4-
- 21:57to-1 margin. The second was for their
- 22:00own ChatGPT users, and the numbers
- 22:02changed radically. 69%thought that
- 22:05technology to detect copies would end
- 22:08up causing false accusations, and
- 22:10nearly 30%said they would use ChatGPT
- 22:12less. If OpenAI implemented watermarks
- 22:15and a competitor didn't. And that is
- 22:17the calculation. The public wants it 4
- 22:20to 1. 30%of customers would leave, and
- 22:23the one who decides is the one who
- 22:25bills, the one who pays. There is no
- 22:28report, methodology, or number of
- 22:30respondents for these two surveys. They
- 22:33are internal documents that the
- 22:34newspaper saw. On August 2, Article 50
- 22:37of the European Artificial Intelligence
- 22:39Act came into effect. It mandates
- 22:41watermarking text in a machine-readable
- 22:43format. Fines reach up to 15 million
- 22:46euros or 3%of global turnover,
- 22:48whichever is higher. Systems that were
- 22:50already on the market before August 2nd
- 22:52have until December 2nd, 2026, to
- 22:54comply. But there are two more things.
- 22:57First, Article 50 itself states that
- 22:59the marking obligation does not apply
- 23:02when systems perform, and I quote, a
- 23:04standard editing support function or
- 23:07when they do not substantially alter
- 23:09the input data provided by the deployer
- 23:12or its semantics. Where is the line?
- 23:15Because that is key. The European
- 23:18Commission published the final
- 23:20guidelines on July 20th of this year,
- 23:22and the test isn't how much text the
- 23:24machine touched; the test is whether
- 23:26the output meaning is the input meaning
- 23:29, and let's look at concrete cases. A
- 23:32spell and grammar checker like
- 23:33Grammarly doesn't need to be marked;
- 23:35it's the textbook example of what's
- 23:37excluded. Correcting errors in a text
- 23:40you wrote yourself doesn't either—
- 23:42same case—but rewriting an email of
- 23:44yours to sound more formal is a gray
- 23:46area, and probably not, because you're
- 23:49changing the register, not the message.
- 23:52But if, from "more formal," you start
- 23:54adding a paragraph with arguments you
- 23:56hadn't included, then yes. Writing that
- 23:59entire email based on three bullet
- 24:01points you provide must be marked.
- 24:03There is no editing there; there is
- 24:05generation. The system doesn't alter
- 24:07your input data; it replaces it.
- 24:09Restructuring a paper from top to
- 24:11bottom or summarizing it must be marked
- 24:13. The guidelines state this expressly:
- 24:16substantive rewrites and summaries do
- 24:18not fall under the exception. In other
- 24:21words, correcting a typo doesn't change
- 24:23what you said, but summarizing does.
- 24:25And writing from scratch has nothing
- 24:28you said to compare it to. So, in the
- 24:31case we saw earlier, you pass your text
- 24:34to Claude to remove typos and it
- 24:36returns it marked. It’s exactly the
- 24:39case Europe excludes, yet Anthropic
- 24:42marks it anyway. It marks more than
- 24:44what Europe is demanding and justifies
- 24:47it by invoking Europe. And the second
- 24:49thing, which is good, is to look at
- 24:51what the law itself says about whether
- 24:52this is technically possible. It says
- 24:55that providers shall ensure their
- 24:57technical solutions are effective,
- 25:00interoperable, robust, and reliable,
- 25:02and I quote verbatim, "insofar as this
- 25:05is technically feasible, taking into
- 25:07account generally acknowledged
- 25:09state-of-the-art." Brussels forces you
- 25:12to mark and, in the same sentence,
- 25:13admits that maybe the technology isn't
- 25:15there, that it's not possible. And
- 25:17listen, one quick thing, all this I'm
- 25:19telling you, the dates, the figures,
- 25:21the direct quotes from the document,
- 25:22take me many hours of verification, one
- 25:24by one. If it helps you and you want
- 25:27this to continue, you can become a
- 25:28member of the vault here on YouD. In
- 25:31exchange, I give you the full dossier
- 25:32each week, but don't see it that way.
- 25:34See it more as patronage so that we can
- 25:36keep going here week after week. If you
- 25:38can't or don't want to, it's no problem
- 25:40. So let's continue. Because while
- 25:43people debate if the watermark lasts,
- 25:45if it gets erased, and if anyone will
- 25:46be able to read it someday, something
- 25:48is already happening, and it's not a
- 25:49laboratory problem. Since there are no
- 25:52watermarks and the one that exists
- 25:54can't be read by anyone, what's being
- 25:55used instead are detectors, programs
- 25:57that look at a text and guess from the
- 25:59outside. And the day one of those
- 26:01programs makes a mistake, it stops
- 26:03being a technical debate and becomes
- 26:05the problem of a real person with a
- 26:06name and surname. On January 28, 2026,
- 26:10a New York court issued a ruling that
- 26:12has barely had any echo, and I think it
- 26:14should. A student from Adelphi
- 26:17University on Long Island turned in a
- 26:19history paper. Turnitin, the world's
- 26:22most widespread anti-plagiarism program
- 26:24, marked it as 100%AI-generated text.
- 26:26100%. Come on. The university declared
- 26:30him guilty based simply on that number.
- 26:33He ran the paper through two other
- 26:35detectors, GPTZero and GPT-4. Both said
- 26:38zero, and so he sued. The judge
- 26:42annulled the offense, ordered the
- 26:43record cleared and the sanction
- 26:45rescinded, and wrote that the
- 26:47accusation and its denial lacked a
- 26:48valid basis and were, quote, devoid of
- 26:50reason. Now, why did this happen to him
- 26:53? Well, let's go to Stanford, 2023. The
- 26:57paper is called "GPT detectors are
- 26:59biased against non-native English
- 27:00writers." They take seven commercial
- 27:04detectors used by universities and give
- 27:06them 91 essays from the TOEFL exam, the
- 27:09English proficiency test, published
- 27:11before 2020, meaning before GPT existed
- 27:13, written by people. Well, 61.22%false
- 27:17positives: six out of 10 human essays
- 27:19flagged as written by a machine. And
- 27:22there's more. 18 of those 91 were
- 27:24marked by all seven detectors at once,
- 27:27unanimously, and 89 out of 91 by at
- 27:29least one. And what happens with essays
- 27:32by Americans? For native speakers, 5.19
- 27:35%, meaning from 61 down to 5. The best
- 27:38part is what they did next. They took
- 27:40those same TOEFL essays and asked
- 27:42ChatGPT to literally improve the word
- 27:44choice to make it sound more native.
- 27:47False positives dropped from 61 to 11,
- 27:49and then they did the reverse. They
- 27:52took the native essays and asked
- 27:54ChatGPT to simplify the vocabulary as
- 27:56if written by a non-native. False
- 27:59positives rose from 5 to 56. And what
- 28:02does that mean? That those detectors
- 28:04aren't looking for whether a machine
- 28:05wrote it; they look at whether the text
- 28:07is predictable. The technique is called
- 28:09perplexity, and the authors explain it
- 28:11like this. How surprised a language
- 28:13model is when trying to guess the next
- 28:15word in a sentence. And a person who is
- 28:18learning a language uses fewer words.
- 28:21Simpler constructions, more repetitive
- 28:23structures; they write in a more
- 28:25mechanically predictable way, so the
- 28:27detector confuses them with a machine.
- 28:30But would the same thing happen today,
- 28:32because everything I just told you is
- 28:34from 2023, and things have changed. On
- 28:37June 29 of this year, 2026, the Vrije
- 28:39Universiteit Brussel published an
- 28:41evaluation of four current detectors:
- 28:44GPTZero, Paraphrasing, CopyLeaks, and
- 28:46Turnitin. 160 documents, and the human
- 28:50texts are graduate student papers from
- 28:52non-native English speakers written
- 28:54before 2019, so there is no possible
- 28:56contamination. Zero false positives.
- 28:59Across all four, not a single one. And
- 29:01there is a second paper from Charles
- 29:03University in Prague and the University
- 29:04of Oslo, presented this year at the
- 29:05European Chapter of the Association for
- 29:07Computational Linguistics. They ran a
- 29:09current commercial detector through
- 29:10those same 91 essays from the Stanford
- 29:12study. False positives dropped from an
- 29:15average of 61 to 23.1, so the bias has
- 29:17been greatly reduced. The tool works
- 29:20better, that is true, but things must
- 29:22be stated as they are. In that same
- 29:25study, that same detector failed zero
- 29:27times with the essays from native
- 29:28speakers. Zero versus 23. The bias has
- 29:32been reduced, not disappeared. The
- 29:34problem has been turned on its head
- 29:36because in the Brussels study, they
- 29:37also included pure machine text
- 29:38generated with the most advanced tools
- 29:40available when they set up the
- 29:41experiment. In that mode, it was GPT-4o
- 29:44, and it didn't get a single one right.
- 29:46It gave 100%of the machine texts a
- 29:48score between 0 and 20%. GPTZero and
- 29:51CopyLeaks didn't detect a single text
- 29:53with confidence either; at most, they
- 29:55reached the doubt range of 20 to 40%in
- 29:57three out of every 10 and one out of
- 29:59every four. Only one worked,
- 30:02Paraphrasing, with a 65%strict accuracy
- 30:05rate. In other words, the detectors
- 30:07have stopped accusing the innocent
- 30:09because they have stopped detecting.
- 30:11They’ve gone from unfair to useless,
- 30:13and no one has told the universities
- 30:15that are still paying for the licenses.
- 30:18But there is one more piece that will
- 30:20affect us directly, which closes the
- 30:23loop on what it means to go through an
- 30:26AI detector. There’s a study
- 30:28presented last year at the AI ethics
- 30:30conference by the AAAI and the ACM.
- 30:33It’s called AI self-preference in
- 30:36algorithmic hiring. 2,245 real resumes
- 30:40written by people, nine models acting
- 30:43as recruiters. And what it shows is
- 30:46that when a language model writes the
- 30:49resume, that model prefers texts
- 30:51written by itself between two-thirds
- 30:53and over 80%of the time. A candidate
- 30:57using the same model as the evaluator
- 30:59has between a 23 and 60%higher chance
- 31:02of passing the filter compared to an
- 31:04equally qualified person who wrote it
- 31:06by hand. In other words, you get
- 31:10flagged for using AI, and the filter
- 31:11that decides if you get an interview
- 31:13discards you for not using it. And here
- 31:17I pass the ball to you because I think
- 31:19this is open for much debate and there
- 31:20are many studies. What do you use it
- 31:23for? For editing, for writing, for
- 31:26programming. And if it were proven
- 31:28tomorrow that you use it, would you
- 31:29stop? Do you think it’s unfair? And
- 31:32now that we know anyone can apply the
- 31:34label and anyone can read it, let’s
- 31:36go straight to the bursts. The National
- 31:39Police arrested a man in Murcia on
- 31:40August 11th who was obtaining digital
- 31:42certificates in other people's names.
- 31:44The method: a real-time deepfake to
- 31:46pass video identification, the kind
- 31:48that asks you to move your head and
- 31:50show your ID. They caught him because
- 31:52his real face appeared on screen for a
- 31:54second. And it wasn't just the face; he
- 31:56had set up a homemade lighting system
- 31:58to simulate the document's holograms in
- 32:00front of his face. With that, he
- 32:02tricked the check ensuring the ID
- 32:03you're showing is physical and not a
- 32:05photocopy. He made 38 attempts,
- 32:07affecting more than 30 people. The
- 32:09generated video lagged slightly behind
- 32:11the actual movement. At one point
- 32:13during the process, it failed, and
- 32:14that’s when his real face appeared.
- 32:16Twitch announced that content from the
- 32:18platform is used to train artificial
- 32:19intelligence models. Amazon’s move on
- 32:22Wednesday included livestreams, clips,
- 32:24highlights, and the text and images
- 32:26from your channel, even the chat. There
- 32:29is a box to opt out. It’s in settings
- 32:31, under security and privacy at the
- 32:32very bottom; it comes enabled by
- 32:34default. If it were optional, no one
- 32:35would turn it on. That sentence was
- 32:37said by the chief product officer of
- 32:38Twitch. Google unveiled the Pixel 11 on
- 32:40August 12th in New York, and the entire
- 32:41lineup is seeing a € 100 price
- 32:43increase in Europe. In the United
- 32:44States, the price remains unchanged.
- 32:46The foldable, the Pixel 11 Pro Fold,
- 32:47will only be sold in France within
- 32:49Europe. As for the rest, there's a new
- 32:51chip and better cameras. Those who have
- 32:53tested it say it performs just like
- 32:54last year's model, but what really
- 32:56matters was in the watch. It's called
- 32:58Health Guardian, and it estimates your
- 33:00blood pressure cuffless by passively
- 33:02measuring your pulse and movement
- 33:04patterns all month long, without you
- 33:05doing a thing. And there is a second
- 33:08feature that analyzes weeks of
- 33:09physiological data to track your
- 33:11insulin resistance. This week marks 30
- 33:13years since the browser that won the
- 33:15war was released. Internet Explorer 3
- 33:17was released on August 13, 1996, and I
- 33:20bring it up for one reason: it was the
- 33:22first commercial browser to support CSS
- 33:25stylesheets, which today shapes
- 33:27absolutely every page you see in
- 33:29wonderful detail. Actually, CSS wasn't
- 33:31even a standard yet. The first
- 33:33specification from the W3C was
- 33:35published on December 17th of that same
- 33:37year, 4 months later. In other words,
- 33:40Microsoft implemented a standard that
- 33:41didn't exist yet, and they only got it
- 33:43half right. Colors and fonts worked,
- 33:45but internal margins didn't, and
- 33:47neither did borders. The people who
- 33:49wrote CSS summarized it like this. It
- 33:52supports almost everything related to
- 33:53color and text well, but doesn't
- 33:54implement much of the box model. 30
- 33:56years later, that box model is the
- 33:58first thing anyone learns when they
- 34:00start building websites. And CSS is the
- 34:02yardstick by which all browsers in the
- 34:04world are measured. If you enjoyed the
- 34:07video, you know what to do: like,
- 34:08subscribe, and share it with anyone you
- 34:10think needs to hear this. I'm Carlos
- 34:12Caballero, and this has been Top 10
- 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.