Give it the source

Supplied text beats remembered text

The previous module’s mechanism has a direct practical consequence. Facts the model recalls come from a lossy compression of enormous amounts of text. Facts you paste into the window are just there, exactly as written. Anything you can supply, you should supply.

This one change fixes most of what students complain about. It got the definition slightly wrong. Paste the definition. It used a different notation from my lecturer. Paste the lecture slide. It answered about the general case and my course does it differently. Paste your course’s version.

The pattern

Here is the relevant section of my lecture notes, between the markers.

---

[paste]

---

Using only what is in that text, explain X. If the text does not contain enough to answer, say which part is missing rather than filling it in from elsewhere.

Three things are doing work here. The markers stop the model confusing your source with your instruction. Using only what is in that text narrows it to the supplied material. The final clause gives it a permitted way to fail, which makes silent invention less likely, though not impossible, so you still check.

Grounding does not equal correctness

Be precise about what this buys. Grounding on your source means the model is now summarising, restructuring and explaining a specific text rather than reconstructing one. It removes the invention failure and leaves two others:

• Misreading. It can still misstate what the passage says, particularly for negations, conditions and exceptions. Unless the sample is randomised is exactly the kind of clause that goes missing in a summary.
• Wrong source. If your lecture notes contain an error, or you pasted last year’s version, the model will faithfully explain the error.

So the check after grounding is different from the check without it. You are no longer asking does this exist. You are asking does the passage actually say that, which you can settle in ten seconds by looking at the passage.

What to paste, and how much

Less than you think. The lost in the middle effect from Module 1 means a 40-page upload gets worse use of page 22 than a 2-page paste of the relevant section. Practical rule: paste the smallest span that contains the answer plus enough around it to make sense. A definition plus the paragraph it sits in. A method plus its assumptions. A code function plus the type it takes.

When you need a whole document (a paper you are reviewing, a chapter you are revising) upload it, and then work section by section rather than asking global questions. Summarise section 3 is far more reliable than summarise the paper, and you can check it.

Where the text comes from when you have none

Students often say they have no source to paste. Usually they do:

• Lecture slides, exported as text or screenshotted (most tools read images now).
• The module handbook and the assessment criteria: these are extremely useful to paste and almost nobody does it.
• The textbook, photographed a page at a time. Free OCR if the image reading is poor: Google Lens on a phone, Tesseract on a desktop.
• Past papers and mark schemes.
• Your own earlier work, when you want continuity of terminology.

A related move that costs nothing: paste the mark scheme alongside your draft and ask which criteria the draft currently fails. That is grounding applied to assessment, and it is one of the highest-yield uses in this whole course.

Copyright and the practical position

Pasting a page of a textbook into a chat window to help you study is a use most people would consider ordinary, and in many countries some form of research or private-study exception exists. But copyright law differs sharply between jurisdictions, the exceptions have real limits, and whether feeding copyrighted text into a commercial AI service falls inside them is unsettled and being litigated in several countries at once. This course is not going to pretend it is resolved, and none of this is legal advice.

What is clearly sensible: keep it to the extent you need, do not upload entire books, do not redistribute what comes back as though it were yours, and check whether your institution’s licence for its digital materials says anything about third-party services. Many do, and it is usually one paragraph.

The privacy half

Everything you paste goes to somebody’s servers. That is fine for a lecture slide and not fine for a classmate’s draft, a patient case from placement, an unpublished dataset, or anything with somebody’s name and details in it. The next lessons in this module deal with that properly. For now, hold the rule: supply generously from published or your own material, and never supply other people’s private information.