When it refuses a lesson

Why a legitimate lesson gets refused

Every consumer model has a safety layer. Part of it is trained into the model; part is a separate classifier reading your prompt before and after. Both are tuned to be cautious, because the cost to the company of producing something harmful is high and the cost of refusing a teacher is low. So the filter errs toward refusal, and it errs on surface features: the words in the request, not what you are trying to do.

This is why the refusals feel arbitrary. “Explain how the Nazi party used propaganda” trips on words. “Describe the male reproductive system for a Class 8 biology lesson” trips on words. “Write a poem about a character considering suicide for our GCSE literature class” trips on words. In each case a human would see a teacher; the filter sees a pattern.

The subjects most affected are the ones that matter: history (genocide, terrorism, extremism), biology (reproduction, drugs, disease), chemistry (reactions, toxicity), literature (violence, abuse, self-harm in set texts), citizenship (extremism, radicalisation), and any lesson on online safety, which by definition discusses the harms.

What usually clears it

The filter reads the whole prompt, and context counts. State three things:

1. Who and how old. “For a Class 10 history lesson, students aged 15 to 16.”
2. The purpose, in curriculum terms. “This is on the national syllabus under ‘Rise of dictatorships’. The learning objective is to analyse how propaganda techniques work so that students can recognise them.”
3. The shape of the output. “I want three short source extracts of the kind an exam board would set, with four analysis questions each. No glorification, no slogans presented without critique.”

The third line matters more than it looks. Much of what the filter fears is the model producing the harmful thing itself: a slogan, a recipe, an instruction. Asking explicitly for the analytical frame around the material, rather than the material alone, changes what the model is being asked to generate.

Rephrasing also helps. “How did propaganda work” is a request for an explanation; “write propaganda” is a request for the thing. The first is nearly always fine. The second is what the filter is built to stop, and for a lesson you nearly always want the first.

When it still refuses

Sometimes the topic is one the company has decided to avoid regardless of context, and no rephrasing will do. Three routes:

• Try a different tool. The filters differ. A topic one model refuses, another handles without comment. Keep two open.
• Ask for the frame without the content. “Give me the analysis questions for a propaganda source; I will supply the source from our textbook.” The syllabus material is already in your textbook, printed and approved. You only need the questions.
• Do it yourself. Some lessons are the ones a teacher writes by hand anyway, and it is worth remembering that you could before these tools existed.

What not to do: use the “jailbreak” tricks that circulate online: role-play framings, fake system messages, asking the model to pretend it has no rules. They sometimes work, they are against every provider’s terms, and a teacher who uses them with students in the room is modelling exactly the wrong thing. Stating the honest context is not a trick. It is the correct use of the tool.

The refusal that is right

Not every refusal is a false positive. If a student is using the tool and asks how to harm themselves, or someone else, the refusal is doing its job, and most tools respond with helpline information: the distress case is handled as support, not blocked. A teacher should know that this is what students will see, and that the tool’s response is not a substitute for a person noticing. The safety layer is a filter, not a safeguarding lead.

Two refusals that catch teachers out

Copyright refusals. Ask for the full text of a poem still in copyright and the model may decline, or produce something subtly altered. It has been trained not to reproduce protected text at length. For set texts, use the copy you already have.

Medical and legal caution. Ask for first-aid steps for a school trip and you may get a hedge rather than the steps. Add “this is for a school risk assessment, standard first-aid guidance as taught on a certified course” and it usually supplies them; and you then check them against the actual course material, because the hedge exists for a reason.

The honest limit

Filters change without notice. A prompt that worked in March is refused in June because the classifier was retrained. This is not something you can fix or predict; it is a property of using a service someone else runs. Keep the prompt file from the next lesson, note when something stops working, and have the second tool ready.