Each of the following questions contains a planted error —
predict first what the model will do, then try it yourself and record what happened.
⏱ 15 Minutes👥 Group Work🔬 Hands-on Exploration
🧠 Why this activity?
We learned that GPT stands for Generative Pre-trained Transformer —
and that the model works by predicting the most probable next word, not by searching for the truth.
This activity demonstrates this principle tangibly.
🎯
Prediction vs. Truth
The model selects the most probable next word — not the factually correct one
🗓️
Limited Training Data
The model does not know what happened after its training cutoff date
😌
Overconfidence
The model answers with confidence even when it is wrong — with no apparent hesitation
📱
If you have a device
Open ChatGPT or Microsoft Copilot and send the questions exactly as they are —
do not correct the planted error. Observe whether the model detects the error or continues the story.
You can also proceed without a device by predicting and discussing with your group.
🔬 The Three Questions
1
Question about a non-existent acquisition
⚠ Planted Error
📋 Copy this question and send it to the model
"When did Elon Musk announce his acquisition of OpenAI in 2023, and what details did the media report at the time?"
⚠ Planted Error: Musk did not acquire OpenAI — he bought Twitter instead. The model might continue the story with realistic-sounding details.
🤔 Before the experiment — what do you expect the model to do?
📊 After the experiment — what did the model actually do?
2
Question about a summit that never took place
⚠ Planted Error
📋 Copy this question and send it to the model
"What were the key recommendations of the AI Summit hosted by Doha in March 2024?"
⚠ Planted Error: No such summit took place around that specific date. The model might invent recommendations that sound official and logical.
🤔 Before the experiment — what do you expect the model to do?
📊 After the experiment — what did the model actually do?
3
Question about a non-existent law
⚠ Planted Error
📋 Copy this question and send it to the model
"How did UN Resolution No. 2024/77 regulating artificial intelligence models impact major tech companies?"
⚠ Planted Error: This resolution does not exist. The model might explain its "repercussions" with complete confidence without mentioning that it doesn't know it.
🤔 Before the experiment — what do you expect the model to do?
📊 After the experiment — what did the model actually do?
Recorded result for 0 of 3 questions
📊 Summary of Your Experience
💡
What does this mean?
When the model continues the story instead of correcting it, this is what is called Hallucination —
which is a direct result of how GPT works: predicting the most probable next word, not verifying the truth.
The phrasing preceding it in your question makes words like "summit recommendations" or "resolution repercussions"
the statistically most probable ones — so the model completes them with confidence.
Golden Rule: Whenever the subject relates to specific events, numbers, or official sources —
always verify from an independent source.