Attempts to explain generative artificial intelligence (AI) have led to a variety of metaphors and analogies. From being described as a “black box” to “autocomplete on steroids,” a “parrot,” and even a pair of “sneakers,” these comparisons aim to make a complex technology more accessible by linking it to everyday experiences. However, these analogies can often be oversimplified or misleading.
One increasingly popular analogy describes generative AI as a “calculator for words.” This comparison, partly popularized by OpenAI’s CEO Sam Altman, suggests that much like the calculators used in math class, generative AI tools are designed to process large amounts of linguistic data. While this analogy has been criticized for potentially obscuring the more troubling aspects of AI, it also highlights the core function of these tools as word calculators.
The Complexity Behind the Calculator Analogy
Critics of the calculator analogy argue that it fails to capture the complexities and ethical dilemmas posed by generative AI. Unlike calculators, AI systems can have built-in biases, make mistakes, and raise ethical questions. Despite these issues, dismissing the analogy entirely overlooks the fundamental nature of generative AI tools as systems designed to calculate linguistic patterns.
The focus should not solely be on the object itself but on the practice of calculating. Generative AI tools are designed to mimic the statistical calculations that underpin human language use. This involves understanding the hidden statistics of language, which most users are only indirectly aware of.
The Hidden Statistics of Language
Consider the discomfort of hearing someone say “pepper and salt” instead of “salt and pepper,” or ordering “powerful tea” rather than “strong tea.” These examples illustrate how language users rely on statistical calculations based on the frequency of social encounters with certain word sequences. The more often something is heard in a particular way, the less viable any alternative will sound.
In linguistics, these sequences are known as “collocations.” They demonstrate how humans calculate multiword patterns based on whether they “feel right”—whether they sound appropriate, natural, and human. One of the achievements of large language models (LLMs) and chatbots is their ability to formalize this “feel right” factor, deceiving human intuition with their outputs.
Why Chatbot Output ‘Feels Right’
LLMs, such as GPT-5 and Gemini, are some of the most powerful collocation systems in the world. They calculate statistical dependencies between tokens—words, symbols, or even dots of color—inside an abstract space that maps their meanings and relations. This allows AI to produce sequences that not only pass the Turing test but also engage users on an emotional level.
The development of these models has roots in the Cold War era, with machine translation tools designed to translate Russian into English. Over time, the focus shifted from simple translation to understanding the principles of natural language processing, influenced by figures like Noam Chomsky. This evolution involved stages from mechanizing grammatical rules to statistical approaches and eventually to neural networks capable of generating fluid language.
Generative AI tools are as much a product of computer science as they are of different branches of linguistics.
The Role of Linguistics in AI Development
The process of LLM development has evolved, but the underlying practice of calculating probabilities remains the same. Despite changes in scale and form, contemporary AI tools are still statistical systems of pattern recognition. They are designed to calculate how we use language to express knowledge, behavior, or emotions, without direct access to these phenomena.
Yet, the language used by companies to describe AI practices often obscures this calculating nature. Terms like “thinking,” “reasoning,” “searching,” or even “dreaming” suggest that AI has gained access to human values transmitted through language. However, AI is still fundamentally about calculation, not understanding or emotion.
The Future of Generative AI
As AI technology continues to advance, it is crucial to recognize its limitations and the nature of its calculations. Generative AI can predict that “I” and “you” are likely to collocate with “love,” but it does not understand these concepts. It is neither a person nor capable of comprehending the emotions or intentions behind language.
Understanding generative AI as a “calculator for words” offers a useful perspective, but it is important not to mistake this analogy for more than it is. The future of AI will likely involve further developments in its ability to mimic human language, but its fundamental nature as a calculating tool remains unchanged.
This material, originally published by The Conversation, has been edited for clarity and length. Mirage.News does not take institutional positions or sides, and all views, positions, and conclusions expressed herein are solely those of the author(s).
