Why 'Stochastic Parrots' Are More Powerful Than Critics Admit

Why the 'stochastic parrot' critique misses the practical significance of generative AI for knowledge work.

Original publication, January 2023. The title predates the widespread use of “is all you need”; I would choose a different title now.

CIOs either already consider AI an urgent priority, will soon recognize the need to do so, or will be out of work in two years.

The June 2017 paper “Attention Is All You Need” changed how natural language processing models were trained. ChatGPT launched on November 30, 2022 and immediately became the fastest-growing platform in Internet history. Its success came from making the latest language tools usable by anyone, rather than from an enormous technological advance.

Generative AI will introduce greater change, faster, than the Internet did.

A Word About Knowledge

Knowledge can be broadly categorized as: explicit (documented), implicit (stored in someone’s head), or tacit (“we know more than we can tell”). ChatGPT is trained on the public Internet, making it a repository of all explicit knowledge available publicly.

For the first time, a tool can retrieve, synthesize, and articulate the sum of publicly documented human knowledge in response to natural language questions.

Tacitness is agent-relative: knowledge that humans cannot articulate may become explicit with AI assistance. Defining knowledge as capability to produce outcomes, rather than justified true belief, also changes how we assess these tools. The parrot’s “regurgitation” qualifies operationally as knowledge if it reliably produces the required outcome.

The Stochastic Parrot

ChatGPT has been criticized as a dangerous “stochastic parrot” that simply regurgitates without insight. The term comes from a 2021 paper warning about the risks of large language models: systems that generate plausible-sounding text by predicting likely word sequences without genuine understanding.

The criticism is valid. These models don’t truly comprehend what they’re saying. They pattern-match and recombine. They hallucinate confidently. They lack the contextual judgment that humans bring to complex decisions.

Generative AI can nevertheless change business because most business activity deliberately repeats established practice.

Established Practice in Knowledge Work

Most knowledge work applies someone else’s prior work. We avoid much creative effort because novel solutions carry more risk than established ones.

Common tasks include applying established frameworks, following documented procedures, synthesizing research, and reformulating existing ideas.

Most knowledge work aims to minimize risk by using explicit, established solutions. Reports follow standard formats, presentations use proven structures, communications use templates, and new problems are addressed by adapting solutions that worked elsewhere.

Organizations rationally prefer established solutions in most business contexts. They value consistency and predictability, and clients expect proven approaches.

But it also means that explicit solutions are accessible to the pre-trained generative “stochastic parrot.”

Competition from AI-Assisted Workers

There’s a phrase circulating in discussions about AI and employment:

“AI will not replace you. A person using AI will.”

A person using AI may be able to do your job faster, cheaper, or better than you can without it.

The model needs to be fast, inexpensive, and good enough; creativity and understanding are unnecessary for these tasks. For work based primarily on documented procedures, established practices, and publicly available information, adequate output is increasingly sufficient.

Skills and Adaptation

Work supported by documentation and established practice is most susceptible to acceleration and eventual automation by AI. Assess how much of your work depends on those resources and how much requires original contributions.

The more your work depends on synthesizing explicit knowledge, the more vulnerable it is to AI augmentation or replacement. The more it requires tacit knowledge, contextual judgment, real creativity, or deep relationship management, the more insulated it remains.

Learn to use these tools, understand their capabilities and limits, and develop skills that complement them.

Generative AI might provide much of what is needed for the vast majority of knowledge work that deliberately relies on established, documented, proven approaches. Those who recognize this will learn to work with it. Those who do not will compete against it and lose.