Large language models, the systems behind chatbots such as ChatGPT, Claude and Gemini, can write fluent paragraphs on almost any subject in seconds. For information hygiene they raise two separate questions. What happens when you ask one of these tools for information? And what happens to the wider information supply when machines can produce text at almost no cost?

Fluent is not the same as true

A language model is trained on huge amounts of text to predict which words are likely to come next. It does this so well that its answers often read like the work of a knowledgeable person, and they are frequently correct. But it is built to produce plausible text, not checked text. When it lacks information, it may fill the gap with something that sounds right. Researchers call this hallucination or confabulation: confident, specific statements that are simply made up, such as invented quotations, statistics or sources.

A widely reported case came in 2023, when lawyers in New York filed a brief in a lawsuit against the airline Avianca that cited court decisions that did not exist. A chatbot had invented them, complete with plausible names and quotations, and had assured one lawyer they were real. A federal judge fined the lawyers and their firm $5,000. The lesson was not that the tool is useless but that a chatbot's tone tells you nothing, since fluency and confidence are not evidence. A model sounds equally sure whether it is right or wrong.

Newer systems often search the web and cite sources, which helps. But a citation can be misread or attached to a claim it does not support, so the citation itself must be opened and checked. Used well, AI tools are excellent for explaining concepts, summarizing a document you supply, or suggesting where to look. Used carelessly, they are a fast way to acquire confident errors.

Can AI text be detected?

Many teachers and editors hoped software could spot machine-written text. Results have been disappointing. In January 2023 OpenAI released a classifier for this purpose, then withdrew it in July 2023 because its accuracy was too low; by the company's own figures it correctly flagged only about a quarter of AI-written text. A 2023 study found that several popular detectors were wrongly flagging writing by non-native English speakers, whose plainer phrasing can resemble machine output. Accusing someone of using AI on a detector's word alone is unfair. For now, the useful question is usually not "who wrote this?" but "is it accurate, and can it be checked?"

The flood of synthetic content

The second problem is scale. When writing costs almost nothing, low-quality content multiplies. The monitoring company NewsGuard began tracking websites that publish mostly AI-generated articles with little human oversight; it counted 49 in May 2023 and well over a thousand by 2025. Many are content farms built for collecting advertising money, recycling and garbling real news, and some have published false death reports and invented quotations. AI tools also make it cheaper to run fake persona accounts, write fake product reviews, and tailor propaganda to many audiences in many languages at once.

There is a feedback problem too. In 2024 a team led by Ilia Shumailov reported in Nature that models trained repeatedly on AI text drift toward bland, distorted output, gradually losing rare details. They called this model collapse. If more of the web becomes machine-made, human-written and carefully checked sources become more valuable, not less.

A fair view also notes the other side. The same tools help fact-checkers sort claims, translate foreign-language sources and transcribe hours of video. A 2024 study found that patient, evidence-based conversations with an AI system reduced many participants' belief in conspiracy theories, and the effect lasted for months. The technology amplifies whatever purpose it is put to.

Habits for the synthetic web

• Treat an AI answer as a lead, not a source. On any fact that matters, verify it against an original source.
• Be suspicious of exact quotations, statistics and citations from a chatbot until you have found them yourself.
• Prefer sources with named authors, editors and a correction policy, which give someone a reason to care about being right.
• Notice text that is fluent but oddly generic, with no specific names, places or dates.

Worked example

A student asks a chatbot for a quotation from Frederick Douglass about education and receives an eloquent sentence with a book title and year attached. Applying this level, the student searches for the exact wording in a collection of Douglass's writings and in a library database. It is not there. A related idea does appear in Douglass's 1845 Narrative, where he describes the moment he grasped that learning to read was "the pathway from slavery to freedom," but the chatbot's sentence is invented. The student uses the real quotation, cites the actual book, and has learned something the chatbot could not supply: what Douglass really wrote. The check took five minutes and saved the essay from a fabricated source.