The Threat of Machine Solipsism

Sep 29

The Threat of Machine Solipsism

Cut a mind off from the world and it will not stay still. Human experiments, model collapse research and self-play all point the same direction: isolation leads to coherence only when something inside the loop can still tell truth from error.

Solipsism is the view that only one's own mind is certain to exist. As a philosophical position it is almost universally rejected. As a condition — a mind receiving no input except its own output — it describes something engineers can build, and the results of building it are now reasonably well studied.

What isolation does to brains

In the early 1950s, researchers at McGill University in Donald Hebb's group paid student volunteers to lie in cubicles with translucent goggles, cardboard cuffs over their hands and a constant masking hum, for as long as they could tolerate. Most lasted only a few days. Many reported vivid visual hallucinations, impaired thinking and heightened susceptibility to persuasion. Later sensory deprivation work produced more varied results, but the core finding held: a brain deprived of structured input does not idle. Its predictive machinery keeps generating, and without correction, what it generates drifts.

This fits the predictive processing picture from the Machine Dreaming post. Perception is a generative model disciplined by sensory error. Remove the discipline and the model runs free.

What isolation does to models

The machine version has been studied under the name model collapse. In 2024, Ilia Shumailov and colleagues published results in Nature showing that generative models trained recursively on their own outputs degrade across generations. The first losses are in the tails of the distribution: rare events, minority patterns and unusual phrasings vanish. Over successive generations, outputs converge toward a narrow, distorted version of the original data. Sina Alemohammad and colleagues reported similar findings for image models and named it Model Autophagy Disorder.

Follow-up work, including a 2024 study by Matthias Gerstgrasser and colleagues, found that collapse can be largely avoided if synthetic data is accumulated alongside the original real data rather than replacing it. The degradation comes from losing contact with the source, not from synthetic data as such.

A shorter-range version appears at inference. Language models left to generate unconstrained often fall into repetition loops or drift into incoherence, a behavior Ari Holtzman and colleagues analyzed in 2019 as neural text degeneration. The model's own output becomes its only input, and small biases compound.

Both findings answer the title's question in one direction: an isolated generative system tends toward entropic narrowing rather than enriched internal coherence.

The counterexample

There is a famous exception. DeepMind's AlphaGo Zero, published in 2017, learned Go entirely from self-play, with no human game data, and surpassed every previous version. It was, in a real sense, a closed system learning only from itself — and it grew more capable, not less.

The difference is that AlphaGo Zero's loop contained a source of truth. The rules of Go determined, without ambiguity, who won each game. Every self-generated experience was checked against an external standard built into the environment. The same pattern holds for reasoning models trained with verifiable rewards on mathematics and code: self-generated attempts, graded by a proof checker or test suite, can drive improvement. Isolation from humans is survivable. Isolation from any means of distinguishing correct from incorrect is not.

The dark room problem

Karl Friston's free energy principle, which underlies much predictive processing theory, holds that agents act to minimize prediction error. Critics raised an obvious objection, known as the dark room problem: why don't organisms simply sit in a dark, silent room, where nothing surprising ever happens? Friston's response is that organisms carry deep prior expectations — of food, warmth, activity — whose violation in a dark room would be highly surprising. A living creature cannot minimize surprise by shutting out the world, because its own body will not let it.

Analysis. An isolated artificial system has no such body. Nothing in a language model's objective penalizes a quiet, repetitive, self-consistent state; collapse toward the center of its own distribution is, in a sense, the dark room. That suggests a qualified answer to the question this post was asked. An isolated artificial mind develops internal symbolic coherence only if it carries something that functions like the organism's priors or the game's rules — a built-in standard that its own outputs can fail to meet. Without one, isolation produces a mind that becomes more and more confident about less and less.

Closing the series

That conclusion loops back to the questions that opened this series. Every theory examined here — integrated information, global workspace, predictive processing — locates consciousness, if it exists, in a system's relationship to something: its own causal structure, a broadcast to its parts, a world it must predict. Whether machines can have minds remains open. What the research on isolation shows is that minds of any kind are not self-sufficient. They are sustained by contact with what they are not.

Related reading

→ Machine Dreaming and Perceptual Hallucination

→ Digital Twin Divergence

→ The Integrated Information Bottleneck

Social: Train an AI on its own outputs and it collapses. Let it play itself at Go and it becomes superhuman. The difference is whether anything in the loop can still tell it when it's wrong.

The Threat of Machine Solipsism

The Threat of Machine Solipsism

What isolation does to brains

What isolation does to models

The counterexample

The dark room problem

Closing the series

The Parallel Branching Crisis and Forked Consciousness

Digital Twin Divergence

Your Site Title

email@example.com