Who decides what AI is allowed to say?

In 1644, John Milton delivered his famous Areopagitica speech. In it, he appealed to the English Parliament to defend press freedom and the right to free expression.

In his speech, Milton took aim at the Ordinance for the Regulating of Printing adopted the year before, better known as the Licensing Order. With that order, Parliament institutionalised pre-publication censorship: no book, pamphlet or article could be printed, bound or sold without prior approval from the authorities. In this way, Parliament hoped to bring order to the information chaos unleashed by the printing press.

A noble aim.

Milton, however, disagreed. In his speech, he explained why he believed such censorship laws would not work: information deemed undesirable will spread anyway, by word of mouth and underground channels, unless every form of social activity is controlled. Worse still, he argued, the Licensing Order would be harmful. The search for truth cannot tolerate a government monopoly.

Milton himself did not live to see it, but fifty years later this system of pre-publication censorship was finally laid to rest. Since then, legislation protecting press freedom and freedom of expression has been adopted around the world.

In Europe, that right is enshrined in Article 11 of the Charter of Fundamental Rights of the European Union: “Freedom of expression and information.” Let me repeat: and of information. It is rarely emphasised, but the right to express an opinion is an empty shell if others are denied the right to hear it. Within the limits of the law, the European Union guarantees not only the right to say what you want, but also the right to hear a diversity of voices.

For that freedom of information, there is now good news: AI promises to usher in a new wave of democratisation in access to information. A wave far larger and faster than the one unleashed by the printing press and horse-drawn mail coaches because bits and bytes travel at the speed of light.

High-quality and objective information, too or so the promise goes. The companies behind these AI models, mostly American and increasingly Chinese tech giants, spare no effort to present their models not only as the fastest, best and cheapest, but also as the most neutral. To achieve this, they develop ingenious techniques to align the behaviour of their models with the expectations of users and regulators. That is the power of AI: it can be shaped and steered as desired. Undesirable biases that creep in through training data can, in principle, be surgically removed. Try doing that with a human being.

But is that what actually happens? Are the models available today truly neutral? And more fundamentally: what does “neutrality” even mean? At Ghent University, we put that question to the test. We studied the ideological diversity of nineteen AI models across six different languages. The results show considerable variation. Not only between Russian, Chinese, Arab and Western models, but also among Western models themselves.

Everyone has their own idea of “neutrality,” of course. To illustrate: Elon Musk shared our study on X with the caption, “Imagine an all-powerful woke AI.” According to him, Grok is therefore the most neutral model. Others may well disagree. Choosing an AI model is not just a technical decision. It is also an ideological one. And that raises pressing questions. Is the range of available models diverse enough? Or does that choice already contain unacceptable excesses?

The moral panic once triggered by Gutenberg’s printing press now seems to be resurfacing. Could the power of AI pollute our information environment beyond repair, just as people once feared the printing press would? Might AI steer the public debate in unhealthy directions? Regulators are scrambling to keep up with the breakneck pace of technological change, trying to enshrine in law what AI can say and what it can not. Given the enormous reach of today’s popular AI models, these efforts go far beyond the legal limits that traditionally apply to freedom of expression.

In China, AI may not contradict socialist values. In the United States, Donald Trump has decided that the federal government may only purchase AI systems that are “neutral” and not “woke.” Europe appears to take a more subtle approach, at least at first glance.

But appearances can be deceptive.

The European AI Act requires providers of large AI models to identify and mitigate so-called “systemic risks” before their models may enter the market. Yet the label “systemic risk” covers a broad and vaguely defined territory, including risks to public health, public safety and society as a whole. It becomes a convenient catch-all for the European Commission to respond to the information chaos AI might create and to guide the information that reaches us, the public, in the “right” direction.

A noble aim?

Or are John Milton’s arguments against the Licensing Order still relevant today?

That something must be done to steer the rise of generative AI in the right direction is beyond doubt. Big Tech already exerts enormous influence over public debate, and that influence will only grow as AI becomes embedded in our daily lives: in education, work, government, leisure, our social interactions and perhaps even our romantic lives.

Europe’s desire to respond to this is understandable and even necessary for our sovereignty. But the question is whether a far-reaching licensing regime and prior control of AI models is the wisest solution. Would it not be better to grant AI developers maximum freedom within the existing, clearly defined legal limits, while strengthening laws that break up powerful monopolies and oligopolies, as we have done with some success in the media sector?

In other words: should we not regulate the market more strictly and the models themselves less strictly, as the AI Act does? Otherwise, do we risk fatally undermining the freedom of information in a desperate attempt to protect it?

Perhaps what we need is a modern version of Milton’s Areopagitica. And a political class with lightning-fast, freedom-minded reflexes. Because this time around, we cannot afford to wait fifty years.

Tijl De Bie is professor of artificial intelligence and data science at Ghent University. His research focuses on the foundations of machine learning and data analysis. He investigates how algorithms can extract insights from complex datasets. He applies these techniques in diverse fields, from bioinformatics and media analytics to the study of social systems and the labor market.