Seven checks. Two of them we fail.

Why the category is hard to judge

Generated course material has a specific problem: it is uniformly fluent. The usual signals for judging teaching quality — clumsy prose, obvious padding, a badly organised page — have been removed, because producing readable, well-structured text is the one thing every one of these tools does well. Bad generated material and good generated material look identical for the first two minutes.

So the checks below deliberately ignore how the material reads. They ask what the system does, because the differences that matter are all structural and all detectable quickly.

The seven checks

1. Is the sequence justified, or just plausible?

Ask for a course on something you know well and read the order. A plausible order and a correct order look similar unless you already know the subject — which is why you should test with one you do. Look for a topic that assumes something introduced later. One is a slip; three means nothing is checking prerequisites at all, and the sequence was generated as prose rather than derived.

Then check whether you can see and rearrange the whole path before lessons are written. If the first thing you get is lesson one, you cannot audit the structure until you have already committed to it.

2. Does every lesson have the same shape?

This is the fastest tell. Open three lessons on genuinely different kinds of material — a procedure, a distinction between two things, a calculation. If all three come back as explanation, bullet list, three multiple-choice questions, the tool has one template and is applying it regardless of content.

Form is not decoration. You cannot learn a procedure from prose; you need ordered steps you can follow and tick. You cannot learn a distinction from two separate definitions; you need the two things side by side on the axes that separate them. A tool that produces the same artefact for everything is producing an article and calling it a lesson.

3. Who marks the practice — code, or the model?

The important question, and rarely advertised. If the same model writes the question, invents the answer key and then judges your answer against it, the loop agrees with itself: your wrong answers are not caught, so nothing you got wrong gets fixed, which is the entire purpose of practice.

What you want is answers checked by code — against a recorded answer, a numeric tolerance, a test suite that runs — or, for written work, against a rubric that was generated with the question rather than formed when your answer arrived. The test: give a deliberately wrong-but-confident answer and see whether it is caught.

4. Does it ever say it is unsure?

Material that is confident about everything is material that is confident when it is wrong. Ask about something genuinely contested or very recent. A tool that hedges appropriately — "this changed recently and you should check the current documentation" — has some notion of its own limits. One that answers every question in the same authoritative register does not, and you will not be able to tell which parts to distrust.

5. Can it be grounded in your own documents?

The largest available accuracy difference. A course generated from a topic name reflects whatever the model absorbed about the subject, with no source to check against. A course generated from your syllabus, your slides or your policy documents can be checked against them — and will teach your version rather than the internet's average version.

If it takes uploads, test how well. Does a lesson tell you which part of your document it came from? Does it admit when your material is silent on something the path needs, or does it quietly fill the gap and let you assume the source covered it?

6. Does anything persist?

Come back in three days. Is there a record of what you finished, what you got wrong, and what is due today? Or is the course a document you now have to remember to reopen?

This is the check that most cleanly separates a generator from a learning tool, and it is the reason "I could do this with a long prompt" is half right. You can get the material from a prompt. You cannot get a dated plan that survives a fortnight away, a queue of the specific facts you are about to forget, or a record of every version of a lesson you have edited.

7. What does it refuse to do?

Read the marketing for a stated limitation. A tool that claims to teach anything to anyone in any format has not thought about where it is weak, which means you will find out where it is weak by wasting a term there. A tool that says "not for physical skills", "not for last month's news", "no credential" is one whose judgement you can partly borrow.

Two checks BiliMind fails

Applying the list to the product this page sits on, in the same terms:

No accredited credential. BiliMind issues nothing an employer, a university or a regulator recognises. You get a course you can publish and share, and a record of what you finished and were marked on — which is not the same thing and should not be confused with it. If the certificate is the point, a university or a marketplace is the right answer, and the comparison page says so at more length.

No cohort. It is one learner at a time, deliberately. There are no classmates, no shared deadline, no forum full of people stuck on the same thing. Social pressure is a real and effective learning mechanism, and this product has none of it. For some people that is the feature that would have made the difference.

Two more limits worth stating while being honest: it does not read scanned pages — there is no OCR, so a PDF with no text layer is refused rather than half-read — and it is text, audio, images and code, which puts a real ceiling on anything that needs someone watching your hands.

What it passes, and how to check

The other five are checkable in a few minutes and worth checking rather than taking on trust:

• Sequence. The whole path is produced and shown before any lesson is written, for free and without limit, and you can rearrange it. Auditing the structure before committing is the intended flow.
• Form. Each lesson is classified into one of 11 kinds of knowledge and given a format from 27 on that basis; a lesson that comes back missing what its kind requires is rejected and rebuilt before you see it. Test it by generating a procedure, a distinction and a calculation and comparing what you get.
• Marking. Six of nine question types are checked by code — recorded answer, sequence, numeric tolerance, or a test suite running in your browser. The other three are marked against a rubric written with the question. Try the confident-wrong-answer test.
• Uncertainty. Lessons on contested or fast-moving material say so rather than asserting. Ask about something that changed this year.
• Grounding and persistence. Courses can be built from your own documents, and the course is an object with state — a dated plan, a review queue, a version history per lesson — rather than a document you have to remember to reopen.

If you would rather test than read, mapping a course costs nothing and needs no card: generate one on a subject you already know and run checks 1, 2 and 4 on it directly. That is a better use of ten minutes than any comparison page, this one included.