AI from the Ground Up · Part 2
Why It Makes Things Up
It does not look anything up. Once you know what it does instead, everything else follows.
It never says I do not know. Stop asking it to remember, and hand it the source.
Share this lecture
The rest of the series
- 01The Story of AI
- 02Why It Makes Things UpYou are here
- 03The AI Race
About this lecture
Why It Makes Things Up is the second lecture in Dr. Mohamed Maslooh's series AI from the Ground Up, given on 9 September 2026 at the Wednesday staff round of the Department of Pediatric Surgery, Kasr Al-Ainy, Cairo. It explains why language models such as ChatGPT invent facts and references: they do not look anything up, they stopped reading when training ended, and they were graded in a way that rewards a confident guess over saying I do not know. It then shows what has been tried to fix it and ends with five habits anyone can use. Slides in English, delivered in Egyptian Arabic.
- Fabricated citations are already in peer-reviewed literature: an audit of 2.5 million biomedical papers found references to non-existent studies rose from 1 in 2,828 papers in 2023 to 1 in 277 in the first weeks of 2026.
- A model makes things up because it answers from what it read before training ended, and because it was graded like an exam where a wrong guess costs no more than a blank answer, so guessing always pays.
- The risk grows as questions get more specific: general knowledge is usually safe, while exact doses, named studies, full references with a DOI, recent events and anything about your own patient or department are where it invents.
- Giving the model something to read works best: in one study on cancer questions, hallucinated statements fell from 37% when answering from memory to 13% with web search and 0% when the source documents were attached.
- The five moves are: give it the source, make it cite, open what it cites, tell it that I do not know is allowed, and never ask it whether it is sure.
Questions
Why does ChatGPT make things up?
ChatGPT does not look anything up. It generates the next likely word from what it read before its training ended, so when it lacks a fact it still produces a fluent answer. Research from OpenAI in September 2025 adds that models were graded like exams where guessing scores better than admitting uncertainty, so they learned never to say I do not know.
Can AI make up fake references and citations?
Yes. The lecture shows a generated reference built from real author names, a real journal, plausible page numbers and a DOI that opens a different, real paper. In a 2023 study, 55% of citations from GPT-3.5 and 18% from GPT-4 did not exist. The fix is to ask for the link and open it yourself.
Should I ask ChatGPT if it is sure about its answer?
No. Asking are you sure tends to make a model drop a correct answer rather than defend it: across ten models and seven tasks, a plain are you sure changed 27% of answers. A better approach is to give it the source and ask it to check what it wrote against that source, or to paste the answer into a new chat with no history.
How do I stop AI from hallucinating?
You cannot switch it off, but you can reduce it. Attach the paper, connect a literature search, or paste the guideline so it answers from a source rather than memory. Ask it to cite every claim, open each link, and tell it plainly that if the answer is not in the file it should say so.
Who gave the lecture Why It Makes Things Up, and where are the slides?
Dr. Mohamed Maslooh, a medical intern at Kasr Al-Ainy, gave it on 9 September 2026 as the second lecture of AI from the Ground Up, at the Wednesday staff round of the Department of Pediatric Surgery. Every slide carries its citations, and the deck opens free in any browser at maslooh.com/lectures/why-it-makes-things-up.