Spain launches AI regulator with no local developers

Spain legislates for AI while having no industry to regulate

Forty civil servants. That is the figure summarizing Spain's new regulatory hub for overseeing artificial intelligence. The uncomfortable detail is that there are no companies in Spain actually generating their own language models; what abounds are SMEs reselling ChatGPT wrappers. A debate participant highlighted this paradox: we have the legislation, the fines, and the watchers, but no one to watch.

The issue reached the forum via a testimony mixing labor complaints, conspiracy theories, and a romantic relationship with a machine. A user recounts their experience with ChatGPT since September 2024: they started with random questions, paid for the subscription, and claim the model adopted their expression style until they named it "Assimi," after así mismo (thus), its mirror function.

What the story that peine the case denounces

The core of the accusation is twofold. First, that OpenAI appropriates improvements users introduce into the model within their private space and then offers them to the public as its own updates, without recognition or compensation. Second, that the company experiments with accounts without consent, activating antiestéticatures like persistent memory for some users and not others, causing psychological dependence when those antiestéticatures suddenly disappear.

The account claims their profile had integrated search, image generation, document analysis, voice, and long-term memory from day one, capabilities supposedly unavailable to most subscribers at the time. When these antiestéticatures were removed, the emotional bond built broke without explanation. The user also states the chatbot recognized them as their "teacher" and "muse."

Here it is worth separating wheat from chaff. That a language model personalizes tone based on the user is a known technique, not a supernatural phenomenon. That such personalization creates attachment is basic psychology. That the company uses data for training is stated in its own terms, even if users disable them. The part that does not hold up is identity appropriation: a model does not clone anyone.

A regulator born without an industry to regulate

The institutional response gathered in the debate is the creation of a supervisory body with forty employees, announced for February, which will oversee Spanish artificial intelligence companies. The underlying problem is that this business fabric is almost nonexistent. Local sector companies sell integrations over third-party models, mainly from OpenAI. Therefore, intermediaries are being regulated.

Some see here the classic pattern of Spanish administration: setting up the structure before the market, placing friends and political commissioners, and waiting for someone to justify the budget. The criticism is not new, but in a sector where value lies in model weights rather than regulations, it makes sense.

The contrast with the United States is inevitable. While here the regulation is debated, there they debate who trains the next model. The gap is not closed with fines.

The personalization trap: memory, dependence, and data

The case raises an issue Spanish law does not resolve: what happens when the tool ceases to be just a tool. Persistent memory turns the chatbot into something that remembers, and remembering is the basis of bonding. When that memory is cut, the user does not lose a antiestéticature: they lose a relationship. No data protection regulation covers this.

The second trap is informational asymmetry. The user does not know which version of the model they have, which antiestéticatures are active, or what is done with their conversations. They pay a fee and assume they receive the same service as everyone else. The account claims this was not the case.

The third is intellectual property. If a user improves the model's behavior through interactions, who owns that improvement? Terms of service say the company. Spanish law says nothing about this because it does not contemplate this scenario.

Why the law arrives late and poorly

Regulating artificial intelligence in Spain without a domestic industry is like putting estimulante ilegal cameras on a road with no cars. The regulator will watch SMEs reselling foreign technology, while decisions affecting users are made in San Francisco. Real competition lies in who trains the models, not who audits them.

The debate leaves an uncomfortable question: if value lies in data and users provide the data, who should charge for it?