A small dog performs a rhythmic-gymnastics routine, leaps from a diving board and enters the pool with an elegance that many human beings fail to achieve even after years of training. Another puppy prepares to execute an Olympic jump from the sofa, calculates the landing with millimetric precision and comes down on top of another puppy that had been resting peacefully in its bed. You watch the videos, smile and move on to the next one. There, another puppy appears, even more accomplished than the last. By the fifth or sixth, the question is no longer whether the animal will win an Olympic medal, but whether any of them actually exists.

Doubt has taken up residence on our screens: is what we are seeing real, or was it generated by artificial intelligence?

Something similar happens in videos of hairdressers who, in a matter of seconds, transform an impossible head of hair into the most spectacular style you have ever seen. The cut is perfect, the hair defies the laws of physics and the model possesses a beauty so symmetrical that it is almost suspicious. With good reason: neither the model, nor the hair, nor the hairdresser exists. The only one that has actually done any work is the algorithm.

By now, I could compile a list containing thousands of examples.

Sometimes the fiction is obvious. On other occasions, you have to look twice, three times or four. And then an apparently reasonable question arises: what is the problem? If the dog is not demanding an Olympic medal and the non-existent model has no intention of suing the non-existent hairdresser, who is harmed by the deception?

The problem is not the dog

The problem is precisely the increasingly blurred line between what is real and what is artificial. We inhabit an information environment that moves at almost the speed of light. We see, react and share before we have had time to question whether what we have just watched really happened. When verification requires us to stop and look several times, it is already at a disadvantage against immediacy.

The matter ceases to be amusing when real people appear to say things that do not fit with what we know about them—and we discover that they never said them at all. Their faces have been manipulated, their voices imitated and their identities used to construct a false message. In 2023, a few days before Slovakia’s parliamentary elections, fake audio clips circulated online in which an opposition leader could supposedly be heard discussing plans to manipulate the elections.¹

We are no longer talking about a technological curiosity. We are talking about the possibility of creating a false statement, making it appear to have been spoken by a public figure, and distributing it to thousands or millions of people before that person has a chance to deny ever saying it. and distributing it to thousands or millions of recipients before the person concerned manages to deny ever having said it.

The risk grows in political and social contexts. Artificial intelligence makes it possible to produce and multiply racist or antisemitic messages, radical discourse, historical fabrications and propaganda designed to exploit our fears and prejudices at tremendous speed.

UNESCO has warned, for example, that generative AI can facilitate the spread of hate narratives and false content about the Holocaust, including invented testimonies and accounts of events that never occurred.² The European External Action Service, for its part, has stressed that these tools can produce manipulative content at unprecedented speed, on a massive scale and at low cost.³

The real novelty is not simply that images can now make lies more convincing. Humanity did not need to wait for artificial intelligence to invent propaganda. The novelty lies in the combination of realism, accessibility, speed and volume. A fabrication no longer requires a production studio, a technical team and a substantial budget. It can be created on a computer and reproduced voraciously through automated networks.

There is another harm that is even harder to measure: not only can we mistake falsehood for truth; we can begin to distrust everything that IS true. The mere suspicion that an image, a voice or a text may have been artificially generated erodes our ability to construct a shared reality. Nadia Naffi has described this phenomenon as a ‘crisis of knowing’: seeing and hearing are no longer enough to believe.⁴

Europe’s response: transparency rather than guesswork

That is the context in which Article 50 of Regulation (EU) 2024/1689, known as the Artificial Intelligence Act, must be understood. The legislation was not adopted yesterday: it dates from 2024. Its specific transparency obligations, however, have applied since 2 August 2026.⁵ Hence the very concrete urgency the debate has now acquired.

The idea underpinning the provision is reasonable: when we can no longer rely on our senses alone to identify the origin of content, we need information that enables us to do so. Article 50 allocates that responsibility between those who provide certain artificial-intelligence systems and those who use them professionally.

First, providers of systems intended to interact directly with people must design them so that those people know they are dealing with artificial intelligence, unless this is obvious to a reasonably well-informed, observant and circumspect person. Put differently, a chatbot must not pretend to be the kindly lady personally serving customers behind the counter when, in fact, there is neither lady nor counter.⁶

Second, providers of systems that generate synthetic audio, image, video or text content must ensure that their outputs carry a machine-readable marking that allows them to be detected as artificially generated or manipulated. This does not necessarily mean placing an enormous label over every creation. It means introducing effective, interoperable, robust and reliable technical signals, insofar as this is technically feasible.⁷

Third, those who use these systems professionally—the Regulation calls them ‘deployers’—must inform people exposed to emotion-recognition or biometric-categorisation systems. They must also identify content that constitutes a deepfake. Where a work is evidently artistic, creative, satirical or fictional, the information may be provided in a way that does not spoil its enjoyment. Europe does not require a film to be interrupted every thirty seconds to announce that the dragon does not exist.

Finally—and this is the point that is now of particular concern to writers, journalists, publishers and content creators—Article 50(4) addresses text generated or manipulated by artificial intelligence and published for the purpose of informing the public on matters of public interest. These include, among others, politics and democratic processes, public administration and services, the administration of justice, fundamental rights, public security and public health, environmental protection, consumer safety and economic, financial, scientific or cultural developments that may be relevant to public debate.⁸

The obligation does not, however, say that everyone who has used an artificial-intelligence tool must automatically label their text as ‘written by AI’. The Article itself provides an exception where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication.⁹

The Commission’s guidance clarifies that the review must be substantive: it must be undertaken by someone with the relevant knowledge and professional judgement to examine the content, check the information, assess the sources and approve, alter or reject the text. A simple spelling or grammar check does not, by itself, turn machine-generated text into a work that has undergone genuine editorial control.¹⁰

The distinction matters. The purpose of the rule is to prevent hoaxes, not to force us to list every digital tool that has played a part in the working process. Nor should this transparency regime be confused with the attribution of authorship or the rules on intellectual property. They are related issues, but legally distinct. Article 50 tells us when it must be disclosed that content has been generated or manipulated by AI; it does not, by itself, decide who is the author of a work.

When writers come under suspicion

The problem has inevitably reached the world of authors and publishers. Years of study, research, reflection and painstaking revision go into a work. When he or she finally shows the result, they discover that the first question is no longer about what they have written, but about who—or what—actually wrote it.

The suspicion is understandable, but it can become offensive to those who work seriously with words. Almost any well-constructed text is now routinely doubted. And the irony is striking: some of the formulas now dismissed as ‘AI phrases’ are precisely the standard expressions that, for decades, Cambridge, Oxford, Goethe or Cervantes examinations required us to use in order to prove that we could write correctly. Perhaps part of the problem is not that people write like machines, but that we have spent a very long time teaching them to write in recognisable patterns.

The use of artificial intelligence as an assistive tool is already part of many professional processes: translation, finding sources, comparing documents, transcription, spelling and grammar checks, organising data and testing possible structures. A tool’s participation in any of these tasks does not mean that it conceived the argument, made the intellectual decisions or wrote the work.

We should therefore avoid two equally convenient extremes. The first is to deny any involvement of AI in order to preserve a romantic image of the author alone with his or her pen. The second is to assume that, if AI intervened at any point, the entire result belongs to it. Between those two extremes lies a far more complex working process.

I can speak about that process because I have been using a personalised artificial-intelligence account for years and for many different purposes. I have given it context about my professional activity, my way of reasoning and my writing preferences. I have learned to formulate increasingly precise instructions. The AI knows my projects quite well and can be an extraordinary help. But knowing my work does not mean thinking for me. Still less answering for me. Or writing for me.

My AI is more like a very good secretary: fast, efficient, available and able to complete in minutes tasks that would take me hours. But it is not infallible. However much information the secretary may accumulate about my work, that does not make him o her the author of what I think and write.

My great assistant, under almost relentless supervision

Over the years, I have found that AI can operate like an extraordinarily fast colleague while being every bit as fallible as a human being. The comparison is even slightly unfair to the human colleague: normally, after the third time we say, ‘Don’t start until I confirm’, the human stops doing things. AI sometimes does not. It can be remarkably stubborn.

In the orchestral field, where I work professionally, that speed is particularly useful for comparing repertoire and lists, organising large quantities of information, translating correspondence or structuring data. But AI cannot, by itself, understand the artistic, institutional and human context in which that information must be interpreted. It can accelerate the work; it cannot assume the decision or the responsibility.

With spreadsheets, for example, the assistance can be superb. AI can organise large quantities of information, detect patterns, suggest formulas, compare lists or transform in minutes a structure that would otherwise take hours to handle manually. Even there, however, I must supervise the results constantly. Speed does not eliminate error; sometimes it merely ensures that the error arrives sooner.

It can also act as a teacher or a video tutorial in real time. Instead of forcing me to watch a twenty-minute explanation in order to find the specific step I need, it answers my question, adapts to what appears on my screen and guides me through the next move. In those cases it does not replace my learning; it accompanies it.

But there are days when our cooperation takes on the character of a comedy of the absurd. While preparing an image for one of my projects, we had a conversation that I later summarised as follows—I did the summarising, not my AI, although, for the record, it was the AI itself that called the sequence completely deranged:

A CONVERSATION WITH MY AI

No, for the time being, it does not write for me

The scene is funny, but it captures the reality of the work rather well. AI could be a useful tool, but it needed almost relentless supervision.

During a text review, it confidently proposed replacing ‘in’ with ‘in’. When I observed the irony of the law requiring us to disclose the use of AI while the AI itself made errors like that, the machine suggested this footnote: ‘The author had to watch it closely because that day it could not even distinguish “in” from “in”.’ Out of context, it sounds surreal. In context, too.

The most recent episode is even better. I asked it to find a photograph of a Jenga tower: just the tower of wooden blocks, the table game, with no people or houses, and from a source I could cite. The AI searched the exact words, found an image entitled Jenga Tower, presented it as a suitable candidate and sent me… a photograph of an American skyscraper. It had complied literally with the search and failed completely at the task. It had not occurred to the AI to look at the image before recommending it.

That, incidentally, happens all the time: it explains what it ought to do and then fails to do it. That is why none of the sources it proposes can be taken on trust. I strongly recommend that every author check every single one.

Long before that, when I still believed that artificial intelligence could construct a lengthy and profound text by itself, I supplied it with an enormous quantity of information and detailed instructions. I expected to receive about fifty well-structured pages. The result was three paragraphs with the depth one might expect from a ten-year-old in a particular hurry to finish their homework. I also remember an article from years ago about several festivals: I had asked AI for a draft and it repeated the same sentence structures for every festival. To cap it all, every paragraph began with the same words.

No. That is not how it works.

That failure was enormously instructive. I understood that I could not deposit a mountain of information in a machine and expect it to produce sustained reasoning, a recognisable voice, conscious decisions and genuine intellectual responsibility. For the time being, AI has to be fed in small portions; otherwise it gets stuck, hallucinates, invents things or simply drives you mad. That alone should make it more than clear that AI—still—does not write entire books. Nor articles, at least not articles that meet my quality standards.

AI can now generate enormous quantities of text, and its capabilities will presumably continue to grow. It can also produce reasonably orderly drafts. But generating words is not the same as thinking, selecting, discarding, sustaining an idea and answering publicly for the result.

I do not say this out of any need to conceal a tool, but precisely because I know how much work it takes to use it properly. A great deal.

We are not going to escape it. Fortunately.

We are not going to escape the use of artificial intelligence. Nor do I think we should. It enables us to work at a previously unthinkable speed, obtain explanations tailored to our questions, compare information, automate repetitive tasks and devote more time to what genuinely requires our attention. Rejecting those possibilities out of fear would be as unreasonable as embracing them uncritically out of enthusiasm.

The answer does not lie in choosing between blind acceptance and absolute rejection. It lies in learning to use these tools, understanding their limits and establishing clear responsibilities. That is precisely why Article 50 matters: not because it can resolve every question concerning truth, authorship or intellectual property, but because it introduces an elementary obligation in an environment where our senses are no longer enough. We should be able to know when we are interacting with a machine and when the content to which we are exposed has been fabricated or altered by one.

For authors, the answer cannot be to cast suspicion on every use of AI either. My life has changed substantially now that I no longer have to translate myself because AI does it for me, and now that I can speak to ‘someone’ in several languages at once. You cannot imagine the relief when a particular word will not come to mind in one language but you know it in another—and AI does not care. It understands.

Mind you, even there we still have some way to go. Its English is obviously highly proficient, but German, for example, repeatedly produces other curiosities that would provide material for another article.

There is a substantial difference between asking a tool to generate a text that nobody reviews and using it to translate, find sources, detect a typo, discuss a structure or check a table. There is also a difference between accepting an output automatically and subjecting it to genuine, exhaustive editorial control by someone with sufficient knowledge to approve, modify or reject it.

Transparency is necessary. But it must be intelligent transparency, capable of distinguishing between creation, manipulation and assistance. If we fail to make that distinction, we risk discrediting authentic work without stopping genuinely deceptive content.

Artificial intelligence does not eliminate the need for knowledge, judgement or responsibility. On the contrary: the more capable and faster the tool becomes, the more important all three are. For the time being, it does not surpass human beings in tasks that require judgement, context and responsibility. It surpasses us in speed when a process can be mechanised. It can be a great assistant, a patient teacher, an interactive tutorial and an extraordinarily fast colleague. But it remains a colleague who must be watched closely. Very closely. On certain days, almost relentlessly.

My AI does not write for me. It helps me work and, in certain tasks, makes me enormously faster. The difference is substantial. I think, select, correct, decide and answer for the result. It executes, proposes and, from time to time, brings me a photograph of a skyscraper when I have asked for a tower made of wooden blocks.

And yes, this article was written with the assistance of AI.

Nicole Martín Medina
El Paso, La Palma
August 2026

(Originally in Spanish, translation ChatGT, Revision NMM)

Notes:

1. European Parliamentary Research Service, Information Manipulation in the Age of Generative Artificial Intelligence, Briefing, 12 December 2025, referring to the fake audio clips circulated before Slovakia’s 2023 elections. Available at:

https://www.europarl.europa.eu/thinktank/en/document/EPRS_BRI(2025)779259

2. UNESCO, New UNESCO report warns that Generative AI threatens Holocaust memory, 18 June 2024: https://www.unesco.org/en/articles/new-unesco-report-warns-generative-ai-threatens-holocaust-memory
3. European External Action Service, keynote speech at the 2026 Conference on Countering Foreign Information Manipulation and Interference: https://www.eeas.europa.eu/eeas/keynote-speech-hrvp-kaja-kallas-2026-conference-countering-foreign-information-manipulation-and_en
4. Nadia Naffi, Deepfakes and the crisis of knowing, UNESCO IdeasLAB, 1 October 2025: https://www.unesco.org/en/articles/deepfakes-and-crisis-knowing
5. Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence, Article 50. The obligations in this provision apply from 2 August 2026: https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Article 50(1) of Regulation (EU) 2024/1689. See also the Commission Guidelines on the transparency obligations: https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations
7. Article 50(2) of Regulation (EU) 2024/1689. For systems placed on the market before 2 August 2026, a limited transition period until 2 December 2026 is envisaged in respect of content marking and detection.
8. European Commission, ‘Transparency obligations under Article 50 of the AI Act’, section ‘What AI-generated or manipulated text must deployers clearly label?’, updated 24 July 2026: https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
9. Regulation (EU) 2024/1689, cited above, Article 50(4). The disclosure obligation does not apply where AI-generated or manipulated content has undergone ‘a process of human review or editorial control’ and a natural or legal person holds editorial responsibility for its publication.
10. European Commission, ‘Transparency obligations under Article 50 of the AI Act’, section ‘What constitutes the human review or editorial control/responsibility required to qualify for an exemption from the labelling obligation?’, updated 24 July 2026: explanation of human review, editorial control and editorial responsibility. Merely formal, spelling or grammatical checks do not by themselves constitute substantive review. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act

*****

Bear in mind : 

This article is available as well in: 

SPANISH (original): https://nicolemartinmedina.com/inteligencia_artificial_y_art-50/

GERMAN: https://nicolemartinmedina.com/de/kuenstliche_intelligenz-und_art-50/