Learning and research
How to use AI for understanding rather than copying, and why every search needs a source that is not the model itself.
What works well
- Having a difficult text explained paragraph by paragraph, with your own follow-up questions.
- Having the same matter explained at three levels, from simple to precise.
- Generating practice exercises and then checking the solution yourself.
- Having your own draft commented on, asking for gaps rather than praise.
- Collecting counter-arguments to your own position.
What does not work
| Intent | Why it fails |
|---|---|
| Looking up facts | No register, no date, no source |
| Finding literature | Citations get invented |
| Learning what is current | Knowledge ends at the training cut-off |
| Having it calculate | Numbers are treated as text |
| Quoting | Verbatim quotes are frequently altered |
Research that holds
- 01
Use the model to sharpen the question
Which terms, which boundary, which sub-questions? That is the strength: turning a vague question into a searchable one.
- 02
Search real sources
A database, an official journal, the literature. A tool connected to a source is a different thing from a model with a memory.
- 03
Have what you found summarised
With the instruction to answer only from the supplied texts and to cite.
- 04
Check every citation
Does it exist? Does it say what is claimed? Two clicks per item.
Learning rather than copying
The difference is the order. Summarise, compute or apply it yourself after having it explained and you retain it. Take a finished answer and you have a text and no understanding. That is not a moral claim but one about memory.
Why citations get invented
A reference has a very regular form: name, year, title, journal, pages, identifier. A model producing the most probable continuation reproduces that form perfectly, even where the content never existed. The more regular the form, the more convincing the invention.
Practical consequence: case numbers, DOIs, ISBNs and section references are the least reliable statements there are and simultaneously the most credible-looking. See Why models hallucinate.
For academic work
- The institution's rules govern. Most now require disclosure of where and how AI was used.
- Detection tools are unreliable and disproportionately accuse non-native speakers. They do not serve as evidence.
- Every adopted statement needs a checked primary source, however it was found.
- Model, version, date and parameters belong in the methods section where AI was used for analysis.
The learning sheet
FREE ACCOUNT
Learning sheet: a topic in four rounds
A routine that uses a model for revision rather than for copying, with the prompt for each round.
Worksheet2 items