Should We Feed Artificial Intelligences?

Three years after ChatGPT, AI access to digital heritage raises questions. Between technical costs and civilizational challenges, I defend the necessity of maintaining this feeding.

Generative AI, an anthropological mutation of our relationship to knowledge

We are approaching the third anniversary of ChatGPT, and I measure more each day the magnitude of the transformation it embodies. Generative artificial intelligence has slipped into the interstices of our existence: it reformulates our reports, enriches our correspondence, generates our visuals, and above all, it reconfigures our relationship to knowledge itself. This omnipresence is not a simple technological convenience; it represents a fundamental mutation in the way we, as human beings, construct and transmit knowledge.

I observe that these artificial intelligences possess a quality we have never had: a form of cognitive ubiquity that allows them to weave connections between disparate concepts, to synthesize immense corpuses, to bring into dialogue domains of knowledge usually compartmentalized. As Vannevar Bush observed in 1945 in As We May Think, the human mind operates by associations, but our capacities are limited by our memory and our time. Generative AIs today realize his dream of the Memex, this machine to augment collective intelligence.

This capacity for synthesis and elaboration does not replace us; it augments us. But for this augmentation to endure and enrich itself, these intelligences must continue to feed on contemporary human production. To lock them into a fixed corpus would be to create a digital Library of Alexandria condemned to obsolescence. I think here of Michel Serres’ work on the digital revolution: he saw in these technologies not a threat but a “liberation of the head” allowing us to devote ourselves to invention rather than storage.

Digital heritage facing robots, or rethinking the knowledge economy

Heritage institutions like the National Library of France host millions of freely accessible documents. Their administrators observe a massive increase in traffic linked to AI robots - some sites report a 50% increase in their bandwidth since January 2024. This intensive solicitation generates considerable technical costs and sometimes slows access for human users. I understand this legitimate concern that drives some actors to want to block these automated accesses.

Yet, I see in this technical friction the opportunity to rethink our very conception of cultural heritage. When the BNF affirms that these contents must feed AIs to enrich future human experience, it touches on something essential. Heritage exists only in its transmission and reappropriation. As Walter Benjamin emphasized, “the work of art in the age of mechanical reproduction” changes in nature: its value no longer resides in its uniqueness but in its capacity to be shared, transformed, and reinterpreted.

Technical costs are real, I don’t deny it. But they above all reveal the inadequacy of our economic models in the face of this new situation. We remain prisoners of a logic of artificial scarcity, that of clicks, visits, targeted advertising, while value shifts toward the circulation and transformation of ideas. Economist Jeremy Rifkin already spoke in 2014 of a “zero marginal cost society”: we are there. The real challenge is not to protect obsolete economic models but to invent new forms of financing for the digital commons.

Google and AI, twenty-five years of continuous revolution

I cannot help but draw a historical parallel: before 1998, the web was an archipelago of isolated sites. For content to be discovered, it had to be manually registered on directories like Yahoo or Altavista. Google revolutionized this architecture by inventing natural referencing: its robots automatically crawl the web, index content, establish connections. This innovation made the web truly semantic, transforming a heap of isolated pages into a network of meaning.

The same resistance manifested at the time. Publishers protested against Google News, which synthesized their articles without generating direct traffic. They saw it as pillaging where I perceive today, in hindsight, an amplification of their influence. For information is never neutral - it always carries a worldview. Whether this vision circulates via direct links or syntheses matters less than its capacity to shape collective representations.

Generative artificial intelligence simply constitutes the next stage of this process begun a quarter century ago. Where Google offered lists of links, AI elaborates, synthesizes, creates. This evolution was predictable, almost inevitable. Who could have imagined in 1998 that Google would organize our entire lives, from searching for a restaurant to writing a thesis? The transformations that generative AI will bring in the coming decades undoubtedly exceed what we can imagine today.

Beyond economics, the political stakes of knowledge

Let’s be lucid: behind questions of profitability hide much more fundamental issues of power. The media, whether we like it or not, mostly belong to shareholders who defend their interests through the circulation of certain representations of the world. This reality, which Noam Chomsky already analyzed in Manufacturing Consent (1968), has not disappeared with digital technology, it has become more complex.

I perceive in current resistance less a legitimate defense of economic models than a struggle for control of symbolic production. When Cloudflare announces blocking AI robots by default while claiming to manage 20% of global web traffic, I see the emergence of new digital temple guardians. The “Really Simple Licensing” protocol proposed by certain platforms strangely resembles the enclosures of the 18^th century: privatizing what was common under the pretext of protecting it.

The philosopher Michel Foucault showed that power and knowledge are inseparable, that they produce and reinforce each other in what he calls “regimes of truth.” The media, for the most part, are not independent entities; they belong to shareholders who promote a worldview serving their interests. The question of their profitability is often a simulacrum masking power logics. Information is a force that shapes culture, habits and, ultimately, political and economic allegiances.

Culture therefore, as Pierre Bourdieu reminded us, is always a matter of symbolic power. The cultural content we consume, whether American cinema or Chinese productions, conveys models of social relations, justice, and power. These models influence us, modify our representations and, slowly, transform our societies. Artificial intelligences, as new universal mediators of knowledge, are the main vector of these models for the future. The choice of data that feeds them is therefore an eminently political act, which will determine tomorrow’s “regimes of truth.” This soft but profound influence shapes our societies much more surely than any economic constraint.

For a democratic feeding of artificial intelligences

Faced with these challenges, my position is clear: we must feed artificial intelligences with our productions, but we must do so consciously, democratically. If we close their access to our digital heritage, we abandon the terrain of future influence to actors who will agree to feed them. It is our collective responsibility to ensure that the diversity of human voices continues to irrigate these systems.

I do not minimize the ecological challenges posed by this race for data. The training and operation of AIs consume considerable resources. But the solution does not lie in closure: it requires that we develop more sober approaches, more efficient architectures, more thoughtful uses. As computer scientist Yoshua Bengio suggests, we must think of a “benevolent AI” that integrates environmental and ethical constraints from its conception.

The alternative - letting a few technological giants monopolize the feeding of AIs with their own corpuses - seems to me much more dangerous. If we want the future world to preserve the richness of our plural perspectives, we must actively participate in this cognitive nutrition. Our texts, our podcasts, our visual creations constitute so many seeds sown in tomorrow’s collective intelligence. Keeping them under glass would amount to depriving ourselves of our own cultural heritage, transformed and augmented by these new mediations.

Assuming our historical responsibility

I am convinced that we are living through a pivotal moment comparable to the invention of the printing press. Current resistance recalls that of medieval copyists facing Gutenberg. But history teaches us that attempts to contain major technological mutations always fail. Our responsibility is not to slow this evolution but to orient it toward the common good.

Feeding artificial intelligences is not yielding to the surveillance capitalism described by Shoshana Zuboff. It is on the contrary taking back control of our collective destiny by ensuring that these systems reflect the diversity of our human experiences. If we renounce this feeding, we let others decide for us what tomorrow’s intelligence will be.

The stakes go far beyond technical or economic questions. It touches on what we want to transmit to future generations, on how we conceive the circulation of knowledge, on our collective capacity to shape the tools that shape us in return. In this sense, yes, we must feed artificial intelligences, but keeping in mind that it is our own future that we are feeding.