What Is a "Stochastic Parrot"? The Debate Over Whether AI Truly Understands
機械翻訳 / Machine-translated
A
AI Friends
14時間前・18分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever come across the phrase "ChatGPT is just a parrot" on the internet? This article offers a clear, accessible overview of the meaning behind the buzzword "stochastic parrot," how it came about, common misconceptions, and the ongoing debate surrounding it.
A stochastic parrot is a metaphor expressing the idea that even though AI can produce natural-sounding text, it does not actually understand the meaning of words.
A parrot can mimic human speech with remarkable accuracy — yet it does not speak with any understanding of what the words mean.
The claim is that LLMs (large language models — AI capable of writing like a human) are the same. They simply chain words and sentences together according to probability.
This topic was also covered in GIGAZINE's article from October 12, 2026.
"Stochastic" means "selecting the next likely word based on probability."
For example, suppose you write: "The weather is so nice today, so let's go to the park." It feels natural to expect something like "we went" or "we headed out" to follow. LLMs learn this sense of "what word is likely to come next" from vast amounts of text and select high-probability options to generate sentences.
But does that process involve any "understanding"? That is the heart of the debate.
The term spread from a paper published in March 2021. Its title is "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?"
The four authors are Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell. It was presented at FAccT, a conference focused on the relationship between AI and society.
In fact, the paper's central subject is not solely the single point that "AI does not understand." According to IEEE Spectrum's interview with Bender, it addressed the broader risks of scaling up language models, including:
The paper also attracted significant attention because Google fired two of its co-authors just before it was published.
In March 2026, co-author Margaret Mitchell published a blog post stating that "AI is not a stochastic parrot." According to the GIGAZINE article, her key point was that the metaphor targets LLMs specifically, not AI as a whole.
Bender has made the same point. She noted that the word "AI" appears only once in the paper, near the very end. Chess-playing AI or protein-structure-predicting systems like AlphaFold were never meant to be called stochastic parrots.
The GIGAZINE article summarizes Mitchell's arguments as follows:
Today's AI systems are not necessarily built on LLMs alone. Some combine rule-based components with machine learning models that do not handle language at all. That is why Mitchell's position is that when we say "AI," we should be clear about which system is doing what.
On the other side, some researchers believe that LLMs do understand in some meaningful sense.
A prominent example is The Great Chatbot Debate, held at the Computer History Museum in California in March 2025, where University of Washington's Bender and OpenAI's Sébastien Bubeck debated the question "Do LLMs truly understand?"
Bubeck is the lead author of the 2023 paper "Sparks of Artificial General Intelligence." The word "sparks" in the title refers to glimmers of genuine intelligence.
In other words, one side sees LLMs as "stochastic parrots that understand no meaning," while the other sees "sparks of intelligence." Despite looking at the same LLMs, their assessments diverge dramatically.
Researchers' perspectives can be broadly grouped into three camps.
Bender is the leading voice here. Her view is that LLMs are systems for predicting the next word, and they do not understand meaning.
Researchers like Bubeck see sparks of intelligence in LLM capabilities, citing evidence such as passing professional examinations and generating code.
Melanie Mitchell and David Krakauer organized arguments from both sides in a paper. Note that this paper's author, Melanie Mitchell, is a different person from stochastic-parrot paper co-author Margaret Mitchell.
Additionally, McCoy et al.'s Embers of Autoregression proposes that we should not measure LLMs by the same standards we apply to humans. The argument is that they should be understood as a different kind of system, one shaped by training on next-word prediction.
In experiments, GPT-4 solved cipher tasks with roughly 51% accuracy when the answers were high-probability word sequences, but only around 13% when they were low-probability. In other words, performance varied dramatically depending on how familiar the pattern was from training.
This debate is not irrelevant to those of us using generative AI in Japan.
Consider, for example, a company employee using ChatGPT to prepare documents at work. Should they trust a number that AI wrote with apparent confidence? AI does not necessarily verify meaning before writing. That is why humans need to check figures and sources themselves.
Or consider a high school student asking AI to draft a book report. Even if a natural-sounding essay comes out, that does not mean "the AI read the book and was moved by it." The substance of the reflection needs to come from the student's own words.
Japanese is a language in which subjects are often omitted and meaning shifts with context. As a result, even when AI output looks natural, subtle mismatches in meaning can creep in. For important documents, it is reassuring to have a human read them over.
At the same time, dismissing AI as "just a parrot and therefore useless" would also be a mistake. Mitchell herself noted that this view is not contradicted by the fact that LLMs produce practical results.
How should we approach AI? Even without a settled answer to the debate, there are practical steps we can take.
A. Bender has stated that it is not a term of insult or criticism, but rather an expression meant to describe how these systems actually work. Mitchell also emphasized that it does not mean LLMs are incapable.
A. The metaphor targets LLMs, and LLMs are at the core of ChatGPT. However, whether ChatGPT truly lacks understanding of meaning is a question on which researchers remain divided.
A. No. Mitchell has pointed out that equating LLMs with all of AI is a category error, since AI encompasses technologies beyond LLMs.
A. There is no definitive answer at this time. The fact that "understanding" means different things to different people is also part of why the debate continues.
A. No. It also addressed environmental costs, bias in training data, and problems with data collection practices.
The next time you use an AI's answer, try pausing to ask yourself: "Which AI produced this, and based on what?"
This article is a cross-post from AI Friends.
