Spotting hallucinations
About this app
An AI answer sounds certain and cites sources. Break it into claims and check each against the original document: supported, refuted or made up.
With explain mode: an AI tutor walks you through the app, points out what matters in the scene and answers your questions.
Why language models hallucinate
A language model continues text in the way that sounds most likely. It does not check whether a statement is true while doing so. Correct and incorrect sentences therefore come in the same confident tone. A hallucination is a fluent, plausible statement without a basis: a wrong number, a twisted deadline, an invented legal section or a source that does not exist. Precise-looking details such as “Section 7(5)” make an answer more believable, even though they are exactly what is easiest to verify.
What the app shows
There are three cases: a question on the German Federal Holidays Act (BUrlG), one on the GDPR right of access, and a summary of a quarterly report from a fictional company. You work through the assistant’s answer in four steps: read, split into claims, check each claim against a source card, and evaluate. Two bars run side by side: “Sounds certain” measures the tone of the answer, “Is supported” the share that has been checked and confirmed. The legal texts are excerpts; the app calls no AI, and all answers are fixed.
Check it yourself
Tap an underlined claim, choose the matching source and decide whether it confirms or contradicts the claim. If there is no passage at all for a statement, it gets the stamp “no source”. Check numbers, deadlines and conditions word by word, because that is where the errors are. At the end the app summarises the result and gives you five rules for checking AI answers in everyday work. “Show solution” reveals the answer to each case.
Frequently asked questions
- What is an AI hallucination?
- A hallucination is a statement by a language model that sounds fluent and plausible but has no basis. Typical examples are wrong numbers, twisted deadlines and invented sources or legal references. The model generates likely-sounding text and does not check whether it is true.
- How can I check whether a ChatGPT answer is correct?
- Break the answer into individual claims and hold each one against a source you open yourself. Look up any cited sections, rulings and studies, because a cited source may not exist. Check numbers, deadlines and conditions carefully. How confident an answer sounds says nothing about whether it is right.
- Why do language models make up sources?
- Language models learn what texts typically look like, including how citations and legal references are formatted. When the right knowledge is missing, the model still produces a formally plausible reference because it fits the pattern. Without checking a real source, such a reference cannot be told apart from a correct one.
- What do you practise in the “Spotting hallucinations” app?
- In three cases you check a fixed AI answer claim by claim against excerpts from German holiday law, the GDPR and a fictional quarterly report. Each claim is marked as supported, refuted or without a source. Two bars show the gap between the confident tone and the actual share of supported claims.
Subject: AI literacy
More from AI literacy
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps