Hey {{ first_name | human }},
Here comes the heatwave…
TL;DR: The 60 Second briefing
🧪 AI feedback: A recent study demonstrates that AI will change the feedback it provides you if it knows your race or gender even when all else is equal.
🚨Norway AI ban: Norway has put a ‘near’ total ban on AI use in the primary phase of education.
📚 AI+education news
🚨 Norway limits AI > What it is: Norway is introducing a near-ban on generative AI for pupils aged 6–13 from the start of the next school year. The default position will be that children in Years 1–7 should not use AI tools, because the government believes they risk bypassing the essential work of learning to read, write and calculate.
The restrictions then ease gradually. Pupils aged 14–16 will be allowed to use AI cautiously under teacher supervision, while students aged 17–19 will be expected to learn how to use it appropriately in preparation for further study and employment
Why this matters: The policy sits within a wider Norwegian attempt to reverse some aspects of classroom digitisation. Following declining assessment results, the country has already restricted smartphones in schools and now plans to fund more printed books, partly reversing its previous shift towards tablets and reduced handwriting. The interesting distinction is that Norway is not rejecting AI altogether. It is sequencing access according to age and educational purpose: foundational knowledge and skills first, supervised use later, and greater independence once pupils are old enough to evaluate and use the technology more responsibly.
🧪 AI Feedback depends on inheritable characteristics > What it is: A Stanford study examining how pupil background information changes the feedback produced by AI writing tools. The researchers tested 600 persuasive essays using GPT-4o, GPT-3.5 Turbo and two Llama models. Each essay remained identical, but the accompanying pupil profile was altered to include characteristics such as race, gender, attainment, disability, English fluency and motivation.
Pupils described as male were more likely to receive direct criticism and specific actions for improvement. Those described as female received more emotional praise, including language such as “love” and “wonderful”.
The models also appeared to reproduce racial and cultural stereotypes. Feedback for Latino pupils sometimes assumed weaker English, while feedback for Asian pupils was more likely to emphasise responsibility, effort and respect.
Do this next:
Before adopting an AI feedback tool, ask the provider how it prevents pupil characteristics from influencing the feedback unnecessarily. Useful questions include:
What information about pupils can the system access?
Does it know a pupil’s name, gender, ethnicity, SEND status, attainment level, English proficiency, behaviour history or previous performance?
Which of these data points are actually used when generating feedback?
Can schools switch off or restrict the use of pupil-profile data?
Has the provider tested whether identical work receives different feedback when pupil characteristics are changed?
What bias testing has been carried out across gender, ethnicity, disability, language background and prior attainment?
How was the model trained or fine-tuned to reduce stereotypical or lowered-expectation feedback?
Does the system use fixed curriculum criteria and rubrics, or does it make broader judgements about the pupil?
Can teachers see why a particular piece of feedback was generated?
How often is the tool audited for inconsistent or discriminatory outputs?
What happens when a school identifies a biased or inappropriate response?
Is pupil data used to train or improve the provider’s models?
Schools should also test the product themselves. Submit the same piece of work under different fictional pupil profiles and compare the feedback. Look particularly at whether the level of challenge, tone, specificity and next steps remain consistent.
The key question for providers is not simply, “Is your AI personalised?” It is: What pupil data does the system know, how does that information affect the feedback, and what evidence shows that it does not lower expectations for particular groups?
🎯AI concepts every teacher should know: 3. Embeddings
Last week, I mentioned that LLMs do not see words as such, they see parts of words and a useful metaphor for understanding that is the idea of morphology. However, an important step LLMs do next is what is called ‘embeddings’. LLMs do not see words or even letters. They see numbers. Embeddings is the process of assigning a number to a token. When the number has been assigned the system works similar to map coordinates. When we know two coordinates, we can explain different relationships between them, like how far apart they are from each other or which is further north or south.
LLMs also measure the ‘relationship’ between tokens. The closer the assigned numbers, the more related those token are going to be, happy and sad could be close together, but guitar and pasta would be further away.
‘Till next week.
Mr A 🦾
Help a colleague save time by sharing this newsletter; distributing these ideas helps a friend get home on time and keeps our energy focused on what matters most: great teaching.
Safety & Privacy Notice
The tools and workflows mentioned are intended for professional productivity and educational enhancement. Users must ensure that any AI implementation remains compliant with their local data protection regulations and institutional safeguarding policies.
Data Privacy: Do not enter personally identifiable information (PII), sensitive student records, or confidential institutional data into public AI models.
Verification Required: AI-generated content can be inaccurate, biased, or out of date. Always maintain a "human-in-the-loop" approach by reviewing and fact-checking all outputs before use.
Professional Judgement: These suggestions do not substitute for formal legal, clinical, or safeguarding advice. Final responsibility for accuracy and appropriateness remains with the professional user.
