I have said this out loud more than once, usually with some confidence: the chatbots everyone is excited about are basically glorified T9. You remember T9, the predictive text on old phones, where you mashed the number keys and it guessed the word. The argument goes that a language model is the same trick, just enormous: it has read a huge pile of text, and all it really does is guess the next word, over and over, very convincingly. No understanding, no thought, just a very expensive autocomplete.
It is a satisfying thing to believe. It cuts a hyped technology down to size, and it has a kernel of something true in it. But "satisfying" and "true" are not the same word, and I did not want to keep repeating a line I had not actually checked. So I went and read the research, looking specifically for the evidence that would prove me wrong, because that is the only honest way to test an opinion you are fond of.
First, what T9 actually was, because the comparison only means something if we are fair about both sides. Tegic’s 1990s system disambiguated number-key sequences against a vocabulary and ranked matching words. The patent also describes user vocabulary and adaptive ranking, so calling it a permanently fixed dictionary was too strong. Its core task was word selection from ambiguous key presses, not open-ended text generation. [1]
So when someone says a language model is "just T9," the strict version of that claim is: it is a lookup table that regurgitates stored words. And that part is simply not how a language model works. There is no dictionary of sentences inside it to look things up in. But there is a softer, more interesting version of the claim, and that one deserves a serious hearing.
The steelman, the strongest version of my own opinion, goes like this. A decoder language model’s pretraining objective is to predict the next token, a chunk of text. The GPT-2 paper describes learning these probabilities from the preceding context. [2] That makes autocomplete a useful description of the training task. It is not a complete description of a modern assistant: instruction tuning, preference training, tools, and the surrounding application can add other objectives and behaviour.
There is a serious academic version of my skepticism too. Bender, Gebru, McMillan-Major, and Shmitchell’s 2021 “Stochastic Parrots” paper argues that producing statistically plausible language does not establish meaning grounded in communicative intent. [3] Readers can supply coherence and intent that the system itself has not demonstrated. That objection deserves more than a dismissive joke about old phones.
The reliability objection is stronger. Language models can produce fluent falsehoods. Farquhar and colleagues’ 2024 Nature paper studies one subset, confabulations that vary in meaning across repeated answers. Their semantic-entropy method detects this instability; it does not catch every error, especially a false answer repeated consistently. [4] Vectara’s public leaderboard separately measures whether generated summaries contain unsupported information. Its percentages concern summaries in that benchmark, not the proportion of all claims a model gets wrong. The live leaderboard changes, so I have removed the unpinned May 2026 snapshot figures. [5]
And it has bitten real people. In Mata v. Avianca, lawyers submitted fabricated authorities generated by ChatGPT and continued defending them after their existence was questioned. The court’s 22 June 2023 order imposed a $5,000 penalty jointly on the two lawyers and their firm and identified six fake opinions in its remedial instructions. [6] Fluent legal language had supplied the appearance of authority without the underlying cases.
Now the part that made me put the brakes on. The "just T9" line quietly assumes that because the training task is simple, the result must be simple too. That assumption is where it falls apart, and the clearest evidence is an experiment about, of all things, the board game Othello.
A group of researchers trained a model to predict legal Othello moves from move sequences, without giving it the board state explicitly. They found internal representations from which board state could be recovered, and interventions on those representations changed its move predictions. [7] That is evidence that a prediction task can produce useful internal structure beyond a dictionary lookup. It does not settle whether the model understands a game in the human sense, or prove that every language model builds equally faithful models of the world.
The GPT-3 paper provides another example: performance on several tasks improved when examples were placed in the prompt, without updating the model’s weights. [8] This is in-context learning. It does not establish that the task or similar examples were absent from pretraining; the paper also examines test contamination. The result is still more flexible than T9’s word-disambiguation interface, without needing the stronger claim that the model had never encountered the task.
Any honest answer has to admit the science is still arguing with itself. I do not want to swap one tidy story for another. There is a genuinely open debate about these surprising new skills, often called emergent abilities. One influential 2022 paper argued they appear suddenly at scale and cannot be predicted from smaller models. [9] A 2023 paper pushed back hard, arguing that a lot of that apparent suddenness is an illusion created by how we score the tests: measure the same skill on a gentler scale and the jump smooths out into a steady, predictable climb. [10] Both papers are from serious researchers, and the matter is not settled. So if anyone tells you, in either direction, that the question of machine "understanding" is closed, they are ahead of the evidence.
So where does that leave my favourite line? Partly vindicated, partly retired. Here is the verdict I would now defend, and it is a split decision.
My instinct was right about the objective. Next-token prediction really is central to these models’ pretraining, and they really do produce confident nonsense when that prediction goes wrong, in ways a thoughtful person would not. Anyone who talks about AI as if it knows things the way a human expert does has skipped over something real, and the stochastic-parrot crowd are right to keep pointing at it.
But my instinct was wrong about the mechanism and the consequences. "Glorified T9" smuggles in the idea of a lookup table, and that is not what is inside. A system trained only to predict can, at sufficient scale, build internal structure that looks a lot like a model of its little world, and can adapt its responses to examples supplied in a prompt. That is categorically more than T9 ever did, and pretending otherwise is its own kind of dishonesty. The truthful sentence is longer and less quotable than my old one: they are next-word predictors whose predictions got good enough to build surprising machinery underneath, and we are still arguing about how much that machinery amounts to.
I landed somewhere less tidy than either slogan. Language models are genuinely shallow in several ways their critics describe and genuinely surprising in several ways those critics miss. Both are worth understanding. I will keep the T9 joke and stop pretending it is the whole story.