Learning goals
- You can explain in two sentences how a language model produces an answer.
- You recognise typical errors such as hallucinations and outdated knowledge.
- You can tell assistance from automation.
- You know why AI helps with some tasks and hurts with others of similar difficulty.
How a language model answers
A language model such as GPT, Claude or Gemini has learned from very large amounts of text which word is likely to follow another. When you ask a question, it produces the most likely continuation word by word, based on everything in the conversation. It does not look anything up unless it is explicitly connected to search or to your documents.
Three practical rules follow:
- Context is everything. The model only knows what was in its training and what you give it in the conversation. Better context, better answers.
- Confident language is not proof. A model phrases a wrong answer just as fluently as a right one.
- Knowledge has an expiry date. Without web search, the model only knows the state of its training. Prices, laws and responsibilities may have changed since.
Hallucinations and how to catch them
When a model invents facts, sources or quotes, this is called hallucination. It happens because the model produces a plausible continuation even when it lacks the information.
The most effective everyday countermeasures:
- Provide sources: Upload the document or paste the text and require answers only from it.
- Ask for evidence: Have the model cite the location for every statement, and open it.
- Allow a way out: Tell the model it may say "not found" or "UNCLEAR". Without that permission it is more likely to guess.
- Check numbers: Always verify sums, percentages and dates yourself once.
The jagged frontier
In 2023, researchers from Harvard Business School and Boston Consulting Group studied how consultants work with GPT-4. On tasks inside the model's capabilities, they were 25.1 % faster on average, completed 12.2 % more tasks and delivered more than 40 % higher quality. On a similarly difficult task outside that frontier, they were 19 percentage points less likely to be correct with AI than without.
The researchers call this the "jagged frontier". It does not run along "easy" and "hard"; it cuts across tasks that look equally demanding to people. So: try AI on real examples where you know the right result before you hand it a task for good.
A second distinction matters. Assistance means a person works with AI and decides. Automation means AI completes a step without human review. According to the Anthropic Economic Index of September 2025, 77 % of business use via the API shows automation patterns; in chat it is about half. Companies automate through APIs; people collaborate in chat.
The frontier test
Find out where the frontier runs in your own job.
Do the task in your day-to-day work
You do this task in your day-to-day work, not here in the course. Tick off the steps once you have done them.
What comes out of it: A short list of three tasks with your classification. You will need it in Module 3.
Enter your result here
Enter what came out of the task. Everything is saved automatically.
See an example
This is what a filled-in result could look like. Your entries stay unchanged.
Your three test tasks
- Summarise the weekly sales report from CRM figuresAI answer: partly correct · Classification: inside the frontier · Note: Good structure, two figures copied wrongly
- Draft a reply to a customer complaintAI answer: correct · Classification: inside the frontier · Note: Tone right after one correction
- Work out the notice period in a supplier contractAI answer: wrong or invented · Classification: outside the frontier · Note: Invented a clause
Used again in module 3
Your result appears there automatically, and you can take it over with one click. You will also find all results in the workbook.
Quiz
1. What does a language model do when you ask it a question and it is not connected to search?
2. Which measure reduces the risk of invented facts the most?
3. What does the "jagged frontier" from the HBS and BCG study describe?
4. What separates assistance from automation?