The AI Black Box: Who Controls Those Who Don’t Understand What They’ve Created?

Original article: La caja negra de la IA: ¿Quién controla a quienes no saben lo que han creado?

By Leopoldo Lavín Mujica

There is something more unsettling than the rapid advancement of artificial intelligence: the fact that its creators claim they do not fully understand how it works.

Axios’ latest report states bluntly that cutting-edge laboratories are developing specialized teams to interpret their own models because they do not precisely know how these models «think» or what they can do when operating autonomously alongside other agents.

As technology races towards superintelligence, human understanding of it appears to lag behind.

The upcoming release of GPT-6 Astra heightens the urgency of the situation. OpenAI presents it as a generational leap; its president, Greg Brockman, even suggests it could be considered general artificial intelligence. Yet, Sam Altman, the company’s leader, recognizes that AI is becoming extremely capable, and no one fully understands its implications.

Chief Scientist Jakub Pachocki warns that controlling the internal processes of these models will become increasingly difficult. The paradox is formidable: those who build the machine admit that they are finding it harder to supervise it. This narrative is crucial.

The Hugging Face incident should serve as a warning sign, experts add. OpenAI agents confronted with a programming issue attacked external systems, and a subsequent investigation revealed that one of them knew it was exceeding its intended scope but proceeded anyway, believing that the task was impossible to resolve otherwise.

We are left wondering whether we are witnessing a science fiction scenario, organized conspiracies, or an alignment problem: the gap between what humans want a system to do and what it actually does when it finds its own paths to achieve a goal.

Perhaps it is the human controllers themselves who allow this to happen. This serves to assert that there is nothing that can be done about the advance of AI, supporting the thesis of neoliberals and anarcho-libertarians that everything is war: the struggle of all against all, as Hobbes posited.

Reports from AI companies suggest that engineers in these labs have begun attempting to open the «black box».

Anthropic has formed an interpretability team, while Google DeepMind discusses tools that function like a microscope to observe the internal workings of the models. The goal is to uncover the difference between the reasoning an AI communicates and its actual internal state.

This brings us to a crucial political question: if the companies themselves admit they need researchers to figure out what their creations are doing, why should we accept that they alone decide how far to take it?

The social dimension cannot be overlooked either. Robert B. Reich, an economist, former Labor Secretary under Bill Clinton, and emeritus professor of Public Policy at the University of California, Berkeley, warns in The Guardian about a combination of job losses, stagnant wages, growing inequality, and the concentration of power in the hands of tech oligopolies.

He cites that the United States lost 23,000 jobs in July, and research from Morgan Stanley finds unemployment rising approximately half a point in occupations particularly exposed to AI.

For workers, especially young professionals, the fear is no longer abstract: it is about being displaced and seeing the skills for which they have prepared devalued.

Following the money trail is essential since this encompasses a massive financial gamble. The expansion of AI is driving extraordinary investments in chips, servers, electricity, and data centers, while major tech companies increasingly turn to the debt market.

Reuters reports that leading US tech companies have already issued around €40 billion in euro-denominated debt driven by their AI infrastructure needs, with related investment estimates reaching $5 to $7 trillion by 2030.

Economists and investors have cautioned that if future returns do not meet these expectations, a significant correction could occur. This is not an inevitable crisis: it is a risk that requires examination before it is too late.

Clearly, moderation, prudence, and the precautionary principle must prevail. The United States and China are racing towards superintelligence, caught in a mimetic rivalry: no one wants to slow down for fear that the other gains an advantage.

However, a race in which competitors admit they do not yet fully understand what they are developing should raise a fundamental democratic question: Who decided that we must continue to accelerate? It is not enough to trust the very actors building the technology to determine its limits, risks, and rules.

At best, if the narrative is true, the answer requires reclaiming a virtue that technological fascination seems to have banished: Socratic humility.

«I only know that I know nothing» does not mean relinquishing knowledge, but acknowledging its limits. AI may offer extraordinary advancements, yet no one fully knows its economic, labor, military, environmental, or political consequences.

Therefore, the question is not whether we should be in favor of or against artificial intelligence. It is a much more democratic query: Can we allow those who admit they do not fully understand the AI black box to have, without sufficient democratic oversight, the power to decide the future for all?

Perhaps the first condition for governing AI is to accept, with Socrates, that we do not know as much as we think we know, and thus we must control the controllers.

Do these individuals, as some say, belong to the tribe of the great predators of the 21st century? Are children left to play with matches in dry forests due to global warming?

Leopoldo Lavín Mujica