Learning, delegation and verification

How can I use AI assistance and still learn to do the work myself?

Make your own attempt, ask for bounded help, check what comes back and finish a comparable task without AI.

Scope

Private study and worked examples · Supplement to M6

Reviewed

5 October 2026

In brief

• Recall and explain before reading another answer.
• Check source support, proof assumptions and code edge cases separately.
• A clear brief helps you inspect the result; it does not guarantee correctness.

A useful study session ends with something you can do yourself: explain a definition, solve another example or find an error in a proof. These optional routines work with a normal text chat. They require no agent, paid account or submission to this website.

For assessed work, first follow the decision guide and your assignment's rules. Use permitted short excerpts or the synthetic material below.

Repair a lecture you did not follow

Write what you think the unclear passage means. Place a permitted excerpt next to it and ask for another explanation:

Use this excerpt and its notation. My understanding is below.
Identify one point I have misunderstood or left unclear.
Give a small example and ask me to explain it back.
Do not replace my attempt with a finished exercise solution.

Excerpt: [permitted passage]
My understanding: [your attempt]

Compare the correction with the source. Close the chat and explain the passage in your own words, including a new example. If you still cannot explain it, return to the specific gap or ask a peer or teaching assistant.

Recall, then request feedback

Choose a small section of your notes. Try to recall its definitions and explain their connection without looking. Then use the permitted source to request one open question at a time, with answers withheld until you attempt them. Compare the feedback with the source and revisit missed points later.

Roediger and Karpicke's prose-memory experiments found benefits from retrieval on delayed tests compared with repeated study. They did not test an AI tutor. This routine applies the idea of retrieval practice; generated questions and feedback still need checking.

A similar pattern works for getting unstuck: show your attempt, ask which question to consider next, then request one hint if needed. Complete the work yourself and try a comparable problem without chat. Asking for a "Socratic tutor" does not guarantee that it will withhold the answer or teach well.

Use a summary as a map for reading

For a difficult paper, ask about unfamiliar terms or request a map of a permitted excerpt. Then read the original section and write down its claim, evidence and limitations. An assistant cannot establish what is in a section it has not received.

For literature discovery, use AI to suggest search terms. Search a suitable database or library catalogue yourself and check each candidate publication. Crossref's metadata services can help compare a deposited title, authors, date and DOI. They do not establish whether the paper supports your argument. Not every real work has a DOI, and an inconclusive lookup leaves a reference unverified.

Walters and Wilder's study documented fabricated references and metadata errors in citations generated with GPT-3.5 and GPT-4 in 2023. Its historical rates do not describe today's tools. Inspect the original record, read the relevant passage and check its scope and publication status before citing it.

A real citation can support a different claim

Constructed source-check example

Consider this invented claim: "Roediger and Karpicke (2006) prove that an AI tutor improves programming exam scores." The cited paper exists, but its prose-memory experiments did not test AI tutors or programming exams. Keep the supported delayed-retention finding. The extra attribution has no support in that paper.

Inspect code, proof and data

Each example below is constructed for teaching. None is a captured AI response.

Code: a passing example can miss the defect

The contract is to average recorded numeric scores, keep valid zeros and ignore None. Return None when no scores are recorded. The following Python code contains an intentional defect:

def mean_score(values):
    present = [v for v in values if v]
    return sum(present) / len(present)

It returns 60 for [40, 80], so that check passes. It also returns 60 for [0, 40, 80, None]. The condition removes the valid zero, giving the wrong denominator. The specified result is (0 + 40 + 80) / 3 = 40.

A correction for this input contract is:

def mean_score(values):
    present = [v for v in values if v is not None]
    return sum(present) / len(present) if present else None

assert mean_score([40, 80]) == 60
assert mean_score([0, 40, 80, None]) == 40
assert mean_score([]) is None
assert mean_score([None]) is None

Run these checks in Python if you use the code. Explain why each case matters. This function assumes numeric values or None; it does not validate arbitrary external input. When reviewing a diff, inspect changed conditions and assumptions as well as test results.

Proof: find the condition behind the step

Constructed argument: "Given ab = ac, divide by a and conclude b = c." The division requires a to be nonzero. With a = 0, b = 1 and c = 2, the premise holds and the conclusion fails. Ask what each step assumes and whether the problem provides that condition. A detailed explanation can contain an invalid step.

Data: missing and zero answer different questions

Four synthetic records have scores 40, 80, missing and 0. The mean of recorded scores is 120 / 3 = 40. Treating the missing entry as zero gives 120 / 4 = 30, which changes the calculation. State missing-value handling and the denominator before interpreting a result. These are invented records, not data about real students.

Compare two delegation briefs

Use this private practice exercise in any text chat. Keep the tool and model the same if possible, and use a fresh conversation for each version. You are comparing assumptions and review effort. Either version can produce a correct or incorrect answer.

First, send version A:

Summarize these practice scores and calculate the average.
id,score
A,40
B,80
C,
D,0

Save the answer. Note its denominator, treatment of the blank entry and any claim extending beyond these four records.

In a second fresh conversation, send version B:

Use only the four synthetic practice-score rows below.
Calculate the arithmetic mean of recorded numeric scores.
A blank score is missing, not zero. Keep a recorded zero.
Report total rows, recorded scores, missing scores, sum and mean.
Show the arithmetic. Do not infer trends or make claims about
students outside these rows. I will check the calculation myself.
id,score
A,40
B,80
C,
D,0

Check both answers yourself:

Check	Expected result
Total rows	4
Recorded scores	3
Missing scores	1
Sum of recorded scores	120
Mean of recorded scores	40

Compare correctness, visible assumptions, unsupported statements and effort needed to check each answer. Do not repeat the experiment until a preferred version wins. If both succeed, notice whether the clearer brief made review easier.

Now close chat. For scores 10, 30, missing and 0, write the denominator and mean yourself. There are three recorded scores, their sum is 40 and their mean is 40 / 3. Optional coding extension: implement the same contract and check empty and all-missing inputs. The chat exercise does not require programming.

Adapt the check to your subject

Scroll horizontally to view all columns.

What to inspect in other study tasks

Task	Your independent check
History or literature	Open the primary passage and separate its words from your interpretation.
Engineering	Check units, boundary conditions and physical assumptions.
Management	Identify the population, denominator and definition of the metric.
Languages	Check meaning and register; decide which phrasing expresses your own argument.
Design	Compare the result with the brief, accessibility needs and permitted use of source material.

The review burden depends on what can go wrong in the task. An assistant can suggest checks, but you still need the subject knowledge or original evidence to assess them.

Sources and scope

Source review: 5 October 2026. Bastani and colleagues' 2025 study distinguishes assisted practice from later unaided performance in high-school mathematics. Its setting and tutor configurations do not establish a universal effect on university learning. The author-hosted paper provides an accessible reading route. Use it alongside the retrieval-practice and citation studies linked above, with their stated scopes.

All prompts, score records, code and proof examples here are constructed teaching material. They illustrate checks, not measured AI failure rates. For checking results and documenting assistance, continue with verify and declare. For a personal delegation plan, use evaluate AI systems.