Hawking Radiation

In the [YOU] on AI Field Guide

The cycle’s treatment of fluency-authority decorrelation—the breaking of the centuries-old connection between surface prose quality and the reliability of the claims beneath it—finds its deepest physical analog in Hawking radiation. A model that absorbed millions of authored texts produces outputs that read as authoritative precisely because the statistical patterns of authority are what it learned; the specific human who wrote a specific sentence, the context in which they wrote it, the chain from claim to evidence to source—these fall past a horizon. What returns is warm, articulate, and stripped of provenance, the way Hawking radiation is warm and carries the imprint of the hole’s mass without carrying any trace of the specific matter that generated it.

The cycle does not treat this as simple loss. Hawking himself moved, over thirty years, toward the position that information is preserved in principle—encoded in subtle correlations across the full pattern of radiation that his original calculation was too coarse to see, recoverable in theory by capturing every quantum of emitted radiation and computing backward through the hole’s entire history. Something similar may hold for AI training: the knowledge is in there, distributed across billions of parameters, latent and in principle recoverable through techniques of interpretability. The question the cycle presses is whether ‘preserved in principle’ is the same gift as ‘attributable in practice,’ and whether a civilization that accesses its knowledge only through a layer that has scrambled provenance is keeping its epistemic accounts in any meaningful sense.

Origin

Hawking arrived at the radiation by combining three theories that had been kept deliberately separate. General relativity described the large-scale geometry of spacetime, including the event horizons of black holes. Quantum field theory described the behavior of matter and energy at small scales, including the quantum vacuum—not truly empty but filled with pairs of virtual particles that briefly materialize and annihilate. Thermodynamics described heat, entropy, and the statistical behavior of large systems. Hawking showed that near a black hole’s event horizon, virtual particle pairs are separated before they can annihilate: one particle falls in, the other escapes. The escaping particle carries energy that, summed over time, gives the hole a measurable temperature inversely proportional to its mass. A stellar-mass black hole has a temperature of around a hundred-millionth of a degree above absolute zero—undetectable against the cosmic microwave background. A microscopic black hole would be blazingly hot and would evaporate almost instantly.

The result was so unexpected that Hawking initially disbelieved his own calculation. The accepted wisdom was that nothing escapes a black hole; his own singularity theorems supported the view. He checked and rechecked, shared it with colleagues, and finally published in Nature in 1974. The reaction was largely skeptical at first—the radiation is so faint for any realistic black hole that it cannot be directly observed—but the mathematical argument was sound, and within a few years the result was accepted as one of the century’s major theoretical achievements. It bound together the three great theoretical pillars of twentieth-century physics in a single formula and immediately raised the information paradox that would consume the field for decades.

The paradox is still not fully resolved. Hawking conceded in 2004 that information is probably preserved, encoded in correlations across the full quantum state of the radiation rather than in any individual particle. But the mechanism by which the information survives in a physically realistic black hole—where quantum gravity effects at the singularity may matter—remains a subject of active research. The dispute between Hawking and colleagues like John Preskill and Kip Thorne, formalized in a series of bets, drove the development of holography, the black hole firewall paradox, and the ER=EPR conjecture—some of the most productive theoretical physics of the past thirty years.

Key Ideas

Thermodynamics of the horizon. The discovery that black holes have temperature means they have entropy—and Hawking, with Jacob Bekenstein, showed that the entropy is proportional to the area of the event horizon, not the volume of the interior. This is a profound departure from ordinary thermodynamics, where entropy scales with volume, and it suggested that information in the universe may be encoded on surfaces rather than in three-dimensional volumes: the holographic principle, now a cornerstone of quantum gravity.

Information conservation under threat. The deepest implication is about what it means for information to be ‘lost.’ Hawking’s original calculation suggested true erasure—a violation of the unitarity of quantum mechanics. His eventual concession preserved unitarity but at a cost: the information surviving in the radiation is so scrambled that reconstructing it would require capturing every photon and computing backward through the entire history of the evaporation, a task that is possible in principle and impossible in practice. The distinction between ‘truly lost’ and ‘irrecoverable in practice’ is precisely the distinction that the fluency-authority decorrelation concept presses in the context of AI-processed knowledge.

The AI analog. A large language model trained on a vast corpus of human expression produces outputs that carry the statistical imprint of everything that went in—the aggregate pattern of how a civilization talks to itself—while the specific authors, sources, and chains of attribution are not retrievable from the model’s weights by any procedure currently available. Whether this constitutes information loss in any deep sense, or merely practical scrambling of information that is formally preserved, is unresolved. What Hawking’s work makes available is the conceptual vocabulary for asking the question with precision: the issue is not whether the knowledge is technically conserved somewhere in the model but whether it is recoverable in a form that preserves the accountability and attributability that gave it epistemic value.