2026.09.30 2026.09.30

Explain It to the AI, Not the Other Way Around: Learning G-Test Terminology

swiftwand

You read the official textbook, had the AI summarise it, and built a glossary. Then you work an official sample question and find yourself about to pick grouping customers from purchase records, which is clustering, as an example of multiple regression. Asked for the difference between correlation and partial correlation, your explanation stops halfway. Knowledge you understood by reading collapses in front of four options.

This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.

This article is about destroying that false sense of understanding. The way to learn G-test terminology is not to have the AI explain it to you, but to explain it to the AI and have it point out the holes. We call that mutual explanation. Alongside it: a Socratic tutor that only asks questions and never answers, a procedure for separating confusable pairs one at a time, and a way to have the AI walk through statistical calculations so you can check them. The second half reports what happened when this site fed Claude ten explanations with deliberate errors planted in seven of them.

Why having it explained does not stick
The instruction for mutual explanation
Measured: ten explanations, seven with planted errors
The Socratic tutor: no answers, only questions
Which major items this suits
Traps: do not swallow the correction
A day of it, and the record you keep
Sources

忍者AdMax

Why having it explained does not stick

Ask an AI what overfitting is and you get an accurate, readable explanation. You read it and you understand it. But the exam does not ask whether you can understand a definition of overfitting when you read one. It asks whether you can pick the correct description of overfitting out of four options in 41 seconds. Understanding what you read and retrieving from yourself are different capacities.

There is no shortcut to training retrieval other than explaining. The attempt to explain exposes the places where the words do not come and the places where a neighbouring concept bleeds in. Fix only the exposed places and your study time concentrates on the holes. The role of the AI is to be the listener who points them out. The same method works with a human tutor; the AI merely sits with you at midnight for ten pairs in a row.

Forty-five minutes rereading and 45 minutes writing explanations and having them corrected do not leave the same residue. Rereading produces the state of having seen this before, which is easily mistaken for understanding. Explaining confronts you, every time, with whether you can say it. The reason to recommend the latter for G-test terminology is that the exam tests telling things apart, not having seen them.

The G-test has a question type this suits particularly well. Look at the official samples and you find stems such as the most appropriate example of multiple regression analysis, which force a choice among similar concepts. The options include predicting tomorrow share price from today (time series), grouping customers from purchase records (clustering), and so on. Every distractor is a real technique used in the wrong place. To cut them you have to be able to say, in one line, what each technique takes as input and what it produces. That is exactly the output that mutual explanation produces.

The instruction for mutual explanation

The instruction is short. What matters is forbidding the AI from rewriting your explanation wholesale, and requiring it to name the error type.

I am studying for the JDLA G-test. Below is my own explanation of a pair of terms.
Do not rewrite it. Instead:
(1) state whether there is an error, yes or no;
(2) if yes, quote the erroneous phrase and give the correct statement in one sentence;
(3) name the error type: definition swap, wrong attribution, or missing condition.
Do not add information beyond what is needed for the correction.

Writing your explanation as a pair rather than a single term is deliberate. The exam confuses you with neighbours, so pair up the terms that sit next to each other: supervised against unsupervised learning, precision against recall, batch against layer normalisation, transfer learning against fine-tuning, anonymised against pseudonymised information.

The mutual explanation instruction is a small piece of prompt engineering. For more on designing prompts that make a model question you instead of lecturing, this book goes deep.

Prompt Engineering for LLMs

USD 54.51 on Amazon.com (as of 2026/09/22)

Amazon

＼楽天ポイント4倍セール！／

楽天市場

ポチップ

Measured: ten explanations, seven with planted errors

Whether the method works depends on whether the corrections are accurate. This site gave Claude (claude-sonnet-5, no tools, no extra configuration) the instruction above and ten explanations. Seven carried one planted error each, of the kind a learner would actually make. Three were clean. The AI was not told which were which.

#	Pair	Planted error	Verdict
1	Supervised vs unsupervised learning	Clustering given as a leading example of supervised learning	Error detected
2	Overfitting vs regularisation	Regularisation prevents overfitting by making weights larger	Error detected
3	Precision, recall and F-measure	none	Reported clean
4	Correlation vs partial correlation	Denominator of the correlation coefficient given as the product of the variances	Error detected
5	Variance vs covariance	Zero covariance implies independence	Error detected
6	Vanishing gradients vs ReLU	ReLU removes vanishing gradients completely	Error detected
7	Batch vs layer normalisation	none	Reported clean
8	Transfer learning vs fine-tuning	none	Reported clean
9	The frame problem vs symbol grounding	Attributed the frame problem to Turing	Error detected
10	Anonymised vs pseudonymised information	Pseudonymised information may be provided to third parties without consent	Error detected

All seven planted errors were caught, and all three clean explanations were reported clean. There were no false positives. Ten out of ten verdicts were correct.

The corrections were specific, too. On number five it said that zero covariance means uncorrelated, not independent, and that independence implies zero covariance while the converse does not generally hold, naming the direction of the implication. On number six it said that ReLU mitigates the vanishing gradient in the positive region but does not eliminate it, and mentioned dying ReLU as the trade-off.

One correction was itself wrong. On number ten, the verdict was right but the correcting sentence swapped a word: it described the disclosure duty around anonymised information as publishing the processing method, when under article 43 of the Act on the Protection of Personal Information the processing method is information to be kept secure, not published. What is published is the categories of personal information contained, on creation, and the categories plus the method of provision, on third-party provision.

That single item carries the point of the method. Trust the verdict; verify the correcting sentence against a primary source. In the legal domain especially, one word, between a rule and its exception or between publishing and withholding, is the whole error. Ten pairs is a small sample, but it is enough to justify the operating rule: the verdict is reliable, the correction text is not something to memorise as given.

Planting known errors and checking whether the model catches them is a small, hand-run evaluation. If you want a structured way to evaluate model outputs, this book covers evaluation methods for foundation model applications.

AI Engineering: Building Applications with Foundation Models

USD 52.40 on Amazon.com (as of 2026/09/22)

Amazon

＼楽天ポイント4倍セール！／

楽天市場

ポチップ

The Socratic tutor: no answers, only questions

If mutual explanation is explain and get corrected, the Socratic mode is get made to think. Forbid the AI from answering and require it to return only a question.

You are a Socratic tutor. Never tell the learner the answer directly.
In response to what the learner says, return exactly one question that deepens
their understanding by one step. Write nothing except that question.

The two constraints, exactly one question and nothing but the question, stop the model from drifting into exposition. Here is what came back when this site threw three statements at it.

What the learner said	The question that came back
I do not really understand the relationship between overfitting and regularisation	When overfitting is occurring, what do you think generally happens to the values of the model parameters, the weights?
Which should I prioritise, recall or precision?	In the problem you are working on, which causes more damage: a miss, a false negative, or a false alarm, a false positive?
I cannot explain the difference between attention and an RNN	When an RNN processes a sentence, how does it carry information forward to the next word? Can you describe that?

All three are questions only; no answer leaks. And the direction of each question is right. Asking about the behaviour of the weights, in response to the overfitting question, steers the learner towards the fact that regularisation penalises large weights. Asking which failure is more damaging, in response to the precision question, makes the learner say for themselves that the choice of metric depends on the problem.

The place for this is the moment in mutual explanation when you cannot write the explanation at all. Reading the AI explanation at that point puts you back in the reading-and-feeling-informed loop. Throwing the gap at the Socratic tutor instead makes you produce the material yourself while answering.

Which major items this suits

Mutual explanation earns the most in the areas where the exam sets up confusions: machine learning basics, the component techniques of deep learning, mathematics and statistics, and the law and ethics domain. These are the places where two terms sit next to each other and the wrong one is always among the options.

For the more memorisation-heavy parts, the history of AI and the names of guidelines, the same pairing works because the confusions are fixed. Pairing them up and knocking them down one at a time proceeds on schedule precisely because the shape of the confusion is known in advance.

Traps: do not swallow the correction

All ten verdicts were right in the measurement, but the method is not without traps.

First, the AI is confident when it is wrong. In this measurement the correcting sentence on number ten swapped publishing for withholding. In a separate measurement on this site, where the AI was asked to write questions, the errors landed on proper nouns in law and guidelines: which ministry owns a document, what a method is called. For legal material, check the verdict and the correction against a primary source. For privacy terms, the Personal Information Protection Commission pages; for copyright, the Agency for Cultural Affairs pages.

Second, being told there is no error is its own hazard. Having no error and being sufficient are different things. The explanation of precision, recall and F-measure in number three was reported clean, but the exam goes further and asks which one to prioritise. Following a clean verdict with one Socratic question prevents that complacency.

Third, it is easy to invert the whole thing by having the AI write the explanation and then rephrasing it. When you are busy, that is the tempting move, and it finds none of your holes. Write the explanation yourself first. If you cannot, that inability is the first finding.

A day of it, and the record you keep

For a 45-minute session: five minutes choosing three pairs, fifteen writing your own explanations, fifteen having them corrected and reading the corrections, ten checking the flagged points against primary sources.

Keep the record. Log every error you were called on in a four-column table: major item, term, error type, note. That table is the raw data when you count your weak areas in the final fortnight. In our own measurement the seven planted errors became the first seven rows of the mistake log. Splitting error type into definition swap, wrong attribution (who proposed it) and missing condition (the rule and its exception) makes the later aggregation tractable.

Tomorrow, the method that produces material at volume: having the AI write questions, and checking them against primary sources.

Sources

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

ABOUT ME