The claim: A large language model writes by guessing the next most likely token (a chunk of a word). Critics call that a "stochastic parrot" — statistical mimicry with no understanding. The labs call it intelligence emerging. Rich's framing: it may be neither, and which one it is changes everything downstream.
The core of it
"Stochastic" just means driven by chance — the model samples from a range of likely next words rather than following a fixed rule. The argument is about what that adds up to. One camp says it adds up to nothing but convincing noise. The other says that at enough scale it adds up to thinking. The interesting possibility is that both are wrong.
The parrot critique
Emily Bender, Timnit Gebru and colleagues argued that language models are "stochastic parrots" — systems that stitch statistical patterns from their training data into plausible-sounding text, with no genuine comprehension, intent or model of the world (Bender et al., 2021).
- Form, not meaning. The model manipulates the shape of language; it doesn't understand what the language is about.
- Fluency is not competence. Sounding smart is not being smart.
- The danger is in us. The real risk is humans reading understanding into a system that has none.
- The costs get ignored. Environmental cost and bias in the training data are underweighted against the hype about capability.
This critique is still the intellectual anchor of the AI-skeptic position.
The counter-argument
The labs and many researchers say the opposite:
- Scale changes the thing. At sufficient scale, statistical patterns become functionally equivalent to understanding — they do the same job.
- There is structure inside. Models build internal representations that look like world models. This comes out of mechanistic interpretability work — reverse-engineering what is actually happening inside a model — at Anthropic (Olah et al.).
- Novel tasks. Models perform on problems that were not in the training data, which is hard to explain as pattern matching.
- "Sparks of AGI." Bubeck et al., 2023 — GPT-4 showed abilities that memorization alone cannot account for.
Rich's take
What if language models are neither parrots nor thinkers, but a genuinely new kind of cognitive process?
Probability as a mode of intelligence
- Human thinking is probabilistic too — the Bayesian brain hypothesis and predictive processing, the idea that the brain is constantly guessing what comes next.
- The distance between "predicting the next word" and "reasoning" may be a spectrum rather than an either/or.
- Chance-driven processes generate novelty; they don't only reproduce. Every output is a fresh sample from a probability distribution. That is a form of creation.
What it means for society
- If AI is "just" stochastic, it's a tool — controllable, owned, directed.
- If AI is genuinely intelligent, it's an agent — with potential rights, responsibilities and autonomy.
- If it is something new, our frameworks for both tools and agents are wrong, and we need new ones.
What it means for agency
- Postmodern agency assumes humans direct AI. But if the outputs are genuinely new rather than retrieved, who is actually directing?
- The human role shifts from creator to curator of stochastic outputs — judging which of the generated possibilities are worth pursuing.
- That is the value-redefined argument applied to intelligence itself.
The parrot spectrum
| Position | Claim | Who holds it |
|---|---|---|
| Pure parrot | No understanding, just statistics | Bender, Gebru, Marcus, Kambhampati |
| Functional intelligence | Indistinguishable from understanding at scale | Altman, Amodei, Hassabis |
| Something new | Neither parrot nor mind — a new cognitive category | Clark, Mitchell, Rich's framing |
| Consciousness possible | Some form of experience may emerge | Chalmers, Koch |
Open questions
- Does it matter whether AI "really" understands, if you cannot tell the difference from the outputs?
- Is human intelligence also "just" probabilistic, and are we uncomfortable seeing the mirror?
- How does this connect to the organizational singularity? If AI improves itself stochastically — recursive self-improvement, meaning the system upgrades itself — is that designing or evolving?
- What does this mean for intellectual property, authorship and creativity? If every output is a novel sample, who created it?
All theories · clickbridge.com · Rich Price · last reviewed September 2026 · People linked here have their own pages at people.clickbridge.com