Biology by Bradford
“Made with AI” has become a quality claim. It tells you nothing. Here is what it actually means on this site, and the checking that is the only part worth arguing about.
“Made with AI” has quietly become a quality claim. It shouldn’t be. Those three words cover a teacher who ran their own wording past a chatbot before printing it, and a thousand questions poured out of one and published unread. Same phrase, opposite objects.
Which means it tells you nothing about whether the thing in front of your class is any good. So here is what it means on this site, in enough detail that you can hold me to it.
The vibe coding is the boring half.
It’s a silly name for something fairly mundane. I describe what I want. Something gets built. I use it, it’s wrong, I say why, it gets rebuilt. Repeat until the thing works, usually at eleven at night with five classes waiting in the morning.
What comes out of that loop is machinery: the paper generator, the on-screen marking, the glossary layer, a hundred and fifty diagrams drawn as SVG so they stay sharp on a photocopy. Plumbing. Genuinely useful plumbing, and I couldn’t have written a line of it five years ago. But plumbing.
None of it decided a single piece of biology.
The questions are not the AI part.
There are 2,946 questions in the bank. 1,999 IB across all forty subtopics, 947 AP across the eight units. Drafted fast, yes. Then read. Every one.
Plenty don’t survive that. Wrong biology. An ambiguity I can hear a class arguing about before I’ve finished the sentence. A distractor no student would ever pick, which is a wasted option and a free mark. A statement that’s really HL wearing an SL label. That last one is the error a student actually feels, because it puts a question in front of someone who was never taught it.
Fifteen years of teaching these two courses is the thing doing the judging. The machine just made the judging worth doing at this scale. There is no version of this where I write three thousand questions by hand between lessons, and there is no version where I publish three thousand I haven’t read.
The checking is the bit worth arguing about.
This is the part I’d actually like teachers to interrogate, because nobody asks it. Not of AI resources, and not of anyone else’s either.
Nothing reaches the site without passing a validator. It isn’t a spellchecker. It tests for the specific ways a question bank rots:
- HL content tagged SL, as above.
- Markscheme points against the mark total. A four-mark question with three creditable points is a question nobody can get full marks on.
- Extended-response structure: one part worth 7 or 8, three parts summing to 15, as the real rubric requires.
- Every referenced figure exists, and every figure is referenced by something.
- Answer-key spread. Both banks once sat near 70% on a single letter. B for IB, A for AP. That doesn’t teach biology, it teaches guessing. They are held to roughly a quarter each now, and every paper generated deals its answers out evenly so no single letter takes more than about 30% of it.
- Length cues. If the right answer is reliably the wordiest one, students stop reading the biology and start counting words.
I’d be curious how many hand-made departmental question banks would pass that last one. Mine didn’t, until it was measured.
Where the lines are.
Nothing is published unread. Not one question, however good the draft looked.
A student is never handed the answer mid-paper. While a paper is being sat the glossary gives the word alone, no definition. A hint gives the vocabulary the markscheme rewards, not the point itself. The assumption behind the whole design is that a student who games a practice paper has only wasted their own afternoon. So honesty is made the easier path rather than a rule to police.
I don’t imitate anyone’s exam product. Writing original questions in the published format is fair, and it is what students need to rehearse. Copying an exam board’s interface, look or branding is a different thing, and I turn that work down when it is offered.
Nothing about a student comes back to me. No accounts, no logins, answers never leave the browser. Not in aggregate, not for improving anything.
Nobody ever asks a department what its questions were checked against. They ask whether a machine was involved.
What’s still rough.
The Turkish and Chinese support is a seed translation. Machine-drafted, not yet checked by a native speaker. The language packs carry that flag in their own metadata and it stays there until someone who speaks the language has been through them. If that is you, I’d take the help gladly.
And there are 2,946 questions, so some of them are wrong. That isn’t false modesty, it’s arithmetic. Tell me and I’ll fix it and say that I have.
Why it’s free, and what that isn’t.
Because I’d rather it were used than owned. I teach these courses, the tools I wanted didn’t exist, and it costs me nothing to leave the door open.
That is a decision about my project. It is emphatically not a verdict on teachers who write and sell their own resources. That is real work. Properly done it takes weeks, and it deserves paying for. I’ve bought plenty of it and been glad to.
Only one of those questions tells you anything about what your students are about to sit. Ask it of me by all means. Then ask it of the last worksheet that came round the department, and see whether anyone can answer.