ChatGPT and GPT models

What it is

A large language model is a program trained on enormous amounts of text to predict what words come next. GPT models are OpenAI's version of this. Researchers reach them in two ways: through ChatGPT, a chat window where you type a question and read the reply, or through an API, a programming connection that lets a script send thousands of requests automatically and collect the answers. The instruction you send is called a prompt. Asking the model to do a job with no worked examples is zero-shot prompting; including a few examples first is in-context learning. Some newer models in the family are built to work through a problem step by step before answering.

The same model can be pointed at very different jobs because its only skill is handling text. Give it a sentence from a paper and it can pull out the material and the measurement. Give it a list of messy labels and it can group them. Give it a chemical goal and it can suggest ingredients. Settings control how freely it answers: a low temperature makes replies more repeatable. Because the output is just text, scripts usually ask for it in a fixed format, such as JSON, and retry when the reply cannot be read.

Why scientists use it

Science runs on text that no database holds in tidy form. Methods, measurements and caveats sit in sentences inside millions of papers, and the vocabulary shifts from one field to the next. Older extraction tools had to be told the patterns to look for. A language model can be told the job in ordinary words instead, which makes it quick to try on a new question and easy to redirect when the question changes.

So it often stands in for work that previously needed a person reading carefully: deciding whether a reported number really measures the thing you want, whether a sample was pure or doctored, whether two differently worded labels mean the same thing. It is also used as plain labour, writing the small scripts that reshape data files, and as a point of comparison, since asking a general model to do a specialist task shows how much of the task needed a specialist tool at all.

In the record

materials-chemistryGPT-4 reads band gap values from paper sentences to train better predictorsGPT-4 read sentences from a band gap dataset, pulled out each material and value, then answered follow-up prompts on whether the sample and measurement qualified.structural-biologyModel labels which protein residues are active sites and what job they doOpenAI o1 was asked to cluster raw protein annotation text into biologically meaningful groups and flag unclear cases, which experts then refined into final categories.materials-chemistryLanguage models suggest ingredients and firing temperatures for inorganic materials recipesGPT-4.1 and o3 were among models asked to rank candidate starting materials for inorganic syntheses and to return firing temperatures as structured data.structural-biologyAlphaFold's confidence scores used to predict which disordered protein segments bind LC8ChatGPT helped write a Python script that split protein sequences into subsequences and paired each one with copies of LC8.

Limits

The model predicts plausible text, so it can produce a confident answer that is simply wrong, and a wrong chemical formula or label looks exactly like a right one. It has no way to check a claim against the world, and it may quietly repeat whatever bias sits in its training text. Answers can change between runs or between versions of the model, which makes results hard to reproduce. This is why careful work still checks a sample by hand, keeps experts in the loop, and scores the output against known answers.

Written by AIxSci: the background from general knowledge, the examples from checked records.