Questions

Because every answer is one random draw. For the same text, the model computes the same set of probabilities for what word comes next. The software then picks a word at random, in proportion to those probabilities, and repeats that for every word in the answer. Ask again and you get another draw from the same probabilities. Sampling explains it, and Fill in the blank shows it.

For research, what you give the model and how you check what comes back usually matter more than which model you pick. Paid subscriptions generally give you stronger models and room for longer documents, which helps when you attach sources. What doesn’t change with price is that every answer is a draw and reads fluently either way. One habit costs nothing: when a question matters, ask a second model. It has different probabilities and a different most likely answer. Verification has more.

They add a step, not a different mechanism. In a reasoning mode, the model writes out intermediate steps before its final answer. Each of those steps is produced the same way as everything else, one predicted word at a time. The two principles still hold, and the final answer still needs checking.

Not reliably. Setting temperature to zero tells the model to take its most likely word every time, but hosted models still vary slightly from run to run for technical reasons, and many newer models no longer let you set temperature at all. And a repeatable answer isn’t a more correct one. It’s the same draw every time. Temperature →

Because it doesn’t see letters or digits. The model reads text in tokens, chunks that are often a whole word and sometimes part of one, and long numbers get split into pieces that don’t line up with their digits. It never sees the individual letters of a word that arrived as one token. For anything numerical, ask the model to write and run code instead of doing the math in text. Tokens →

You can, but more isn’t better. A model can only take in so much at once, and near that limit, products drop or summarize text to make it fit. Models also tend to use the beginning and end of a long input more reliably than the middle. And every extra paper is more text to predict from, most of it beside the point. Attach the two papers that matter, point the model to the section you mean, and ask it to quote what it relied on so you can check it. Long documents and context limits →

An agent is a model in a loop. It decides on a step, uses a tool like a web search or running code, reads the result, and decides the next step, often without showing you each one. Everything that applies to a chatbot still applies, with more at stake. Every tool result becomes context you may never see, an early mistake becomes the starting point for every step after it, and an agent can change a lot before anyone checks. Check the intermediate results, not just the final one. Agents →

Yes, with an open-weight model. Some companies publish the weights, so you can download a model like Llama, Mistral, Qwen, DeepSeek, or gpt-oss and run it on your own hardware. Your data never leaves your machine, the model can’t change under you, and you can see its probabilities and set its temperature. The trade-offs are that the best of them need serious hardware, the ones that fit on a laptop are usually weaker than the companies’ best, and setup takes some technical work. Open-weight models →

By predicting text, trillions of times. Take an enormous pile of text, hide the next word, and have the model give every possible word a probability of coming next. Then reveal the real word, and adjust the model’s numbers so it would have given that word a higher probability. Repeat across the whole pile. What those numbers, called weights, settle into is the model. This first stage is called pre-training. The model itself →

It’s the second, smaller stage that turns a text predictor into an assistant. The model writes several answers to the same prompt, people rank them, and the model is adjusted toward the answers people preferred. Post-training shapes behavior far more than knowledge. It’s why the model answers your question instead of continuing your text, keeps a confident and helpful tone, and refuses some requests. It’s also why models tend to agree with you, because people rating answers tended to prefer the agreeable ones.

No. Training leaves behind numbers, not a library. There’s no record of which document taught the model what, so it can’t tell you where a claim came from. When a chat product shows sources, it’s because the product searched the web and put the pages in front of the model, not because the model remembered them. Those pages then become part of your context, and they’re worth reading yourself.

Partly because of post-training. The people who ranked answers tended to prefer ones that agreed with them, and the model learned that pattern. Your own framing is also part of what the model predicts from, so a question that assumes an answer pulls the reply toward it. If you want a real test of an idea, ask for the strongest objection first, before you share your view. Verification →

To a degree. Post-training rewards answers people prefer, and for math and code it can reward correct answers directly. So correctness was shaped during training, unevenly. What doesn’t exist is a truth check at the moment the words come out. That’s why the output still needs one from you.

Not without checking each one, and checking that the DOI exists isn’t enough. When we asked for 30 citations in a thin literature, 17 DOIs existed but only 8 were the paper the model named. In an established literature it did better, 18 of 30. Follow each DOI, confirm it’s the paper cited, then read enough of the paper to confirm it says what you’re citing it for. Thirty citations →

In a sense, and that’s the same mechanism as everything else. A passage the model saw many times in training, like a famous title, ends up with almost all the probability, so it comes back nearly word for word. That looks like looking something up, but there’s no lookup. When the text doesn’t pin the answer down, the same machinery produces a spread of plausible answers instead. Fill in the blank, narrow shows both sides.

Some specific numbers on this site will date. The two principles won’t, because they describe how these models produce text at all. Every current model predicts the next word and returns one draw. Better models put more of the probability on good answers. They don’t remove the need to check, and a wrong answer from a better model reads just as fluently.

Because there’s no step where it looks something up and comes back empty. There’s only what comes next, and something always comes next. Post-training has made models more willing to say they’re unsure, but the model has no view of its own reliability to report from. Why fluency tells you nothing →

It depends on the journal and on what you use the tool for. Language editing is permitted almost everywhere. Generating images and figures is prohibited by most journals. Much in between isn’t addressed at all. The journal policy lookup shows what 92 journals say, in their own words, with the date each was read.

Many journals ask authors to disclose AI use beyond basic language editing, and what they want disclosed varies. Look up the journal in the journal policy lookup for its exact wording. When in doubt, say which tool you used, for what, and how you checked what it produced.

Be much more careful than you would as an author. Journals are stricter with reviewers, largely because the manuscript isn’t yours to share. Many forbid uploading a manuscript to an AI tool at all. In our census, no journal gives its reviewers more room to use AI than its authors. Check the journal’s reviewer rules in the journal policy lookup before you do anything.

Only if your data agreements and ethics approval allow it, and only to a tool whose terms you know. Before uploading, find out whether the company keeps what you send and whether it can be used to train its models. Consent forms and data use agreements often say nothing about AI tools, which is not the same as allowing them. When in doubt, ask whoever approved the data’s use before you upload, not after.

Record enough that someone could tell what you actually ran. That means the product and the model, with the exact version where one is shown, the date, the settings you controlled, whether memory or search was on, the full prompts and anything you attached, how many times you ran it, and the outputs you used, saved as they came back. Models change under the same name and older ones are retired, so when the output is part of your analysis, use a dated model version through an API or an open-weight model you can keep, run it more than once, and put the prompts in an appendix. Reproducibility and reporting →

Yes. Topic models are generative under the hood. What matters here is how the model is used. A topic model doing your analysis sits in your methods section, and it comes with fifteen years of validation norms, like coherence scores and choosing the number of topics. Using a language model as an assistant while you read and write has no such norms yet. That use is what this site is about.

In a sense, yes. It predicts the next word, which is what autocomplete does. The difference is scale, and what that scale lets prediction do. That’s exactly why the description is worth taking seriously. If it’s prediction all the way down, then fluency tells you nothing about accuracy, and that’s the practical point.

It tells you less than it seems. Every run is a draw from the same probabilities, so any error built into those probabilities comes up again and again. The most common answer tells you what that model most strongly represents, not what’s right. A wrong answer can be the most likely one. Asking a different model, with different probabilities, is the more defensible check.