PlaxOpen Plax →

How AI chatbots and summaries work, and why they sometimes get things wrong

Last updated 10 November 2026

By Plax Labs · 4 min read · Published 10 November 2026

AI chatbots can answer questions, draft emails and summarise articles in seconds, and they sometimes state something false in the same calm voice. Both behaviours come from the same design, and understanding it makes you a better user of the tools, and a more careful reader of anything written with them, including parts of Plax.

The technology behind most of these tools is called a large language model, or LLM.

Predicting the next word

Text is first cut into small pieces called tokens, roughly words or parts of words. During training the model is shown vast amounts of text with a piece hidden and asked to guess it, and its internal settings, billions of numbers called parameters, are adjusted every time it guesses wrong. After many rounds it becomes very good at estimating which token is likely to come next in a given context.

To produce an answer, the model estimates the probabilities of the next token, picks one, adds it to the text and repeats. A setting often called temperature controls how adventurous the choice is. It is autocomplete at an enormous scale, and the surprising thing is how much knowledge and reasoning-like behaviour comes out of it.

From predictor to assistant

A model trained only to continue text will happily continue a question with another question. To make it behave like an assistant, developers train it further on examples of good answers and use human feedback to rate its responses, so that it learns to follow instructions and to refuse some requests. Hidden instructions, called a system prompt, set its tone and rules for a particular product.

Why they make things up

The model has no built-in notion of truth. It is trained to produce text that is plausible given what came before. If it has seen a fact many times it usually repeats it correctly. If a question is about something obscure, recent or absent from its training, it still produces fluent text, and the fluency can hide the fact that the content is invented. Its knowledge also stops at the date its training data ended, unless it is given a way to search.

Fabricated references are a well-known result: ask for a source and the model may produce a title, authors and a journal that look right and do not exist. The remedy is not to trust a model's memory for facts that matter. Tools that let it retrieve documents and quote from them reduce the problem but do not remove it.

Summaries: safer, but not safe

When a model is given a text and asked to summarise it, the facts are in front of it, so it has much less need to invent. This is why summarisation is among the most dependable uses of these models. It is still possible for a summary to go wrong in predictable ways:

  • Dropping qualifiers. Alleged, may, according to and up to disappear, and a possibility reads as a fact.
  • Getting a number or a name wrong, or attaching a statement to the wrong person.
  • Merging two details from different parts of the text into one that the source never said.
  • Emphasis. What a summary leaves out is as much a choice as what it includes, so the same article can be summarised in ways that lean differently.

How Plax uses AI

The stories in the Plax feed are short summaries of what publishers have published. For each story, Plax takes the headline and the description or opening the publisher provides, and an AI model is asked to write a short brief that sticks to those facts, adds no opinion, and leads with the most important point. Every story links to the original at its publisher, and the headline is the link, so a reader can check anything that matters. Features such as Go deeper, which add background on request, are written by AI too, are labelled as such, and can contain mistakes. Our editorial policy describes this in more detail.

Using AI output sensibly

Treat AI output as a first draft from a fast, well-read assistant who sometimes bluffs. Check numbers, names and dates against the source. Be most careful in health, money and law, where a confident mistake costs the most. Ask for the source and follow it. And when the answer matters, ask the same question in a different way: if the answers disagree, that is a sign the model is guessing.

Keep reading

Written by Plax Labs with the help of AI writing tools, and updated 10 November 2026. Found a mistake or something out of date? Tell us, and read our editorial policy. Read Technology stories on Plax.