SASHA Learning Brief
AI Homework Help: Start With the Teacher’s Instructions
A practical way to check AI homework help against the current assignment and rubric, with careful limits on what retrieval and educational guidance establish.
Start with the task in front of you
Imagine a teacher asking a learner to compare two explanations using evidence from a particular passage. An AI response supplies an impressive summary of the whole topic, but never compares the explanations. The writing may sound convincing while missing the assignment. This fictional example illustrates a useful starting question for homework support: what, exactly, is the learner being asked to practise?
Before opening a learning tool, find the latest assignment instructions, the relevant classroom example and any marking criteria or rubric. Check whether the teacher has issued a correction or clarified the task. Note the expected method, sources and form of response. The practical routine proposed here is intended for discussion with an educator; it is not a validated assessment or a guarantee of learning gains.
Check which help is allowed
England’s Department for Education recommends that schools consider how their homework policies address generative AI and explain when its use is appropriate. Its guidance also requires professional judgement when checking generated content. This is guidance for England, rather than a universal school policy. The learner’s own school rules and current teacher instructions remain the place to establish permitted help.
Our suggested first check is simple: distinguish help understanding a question, hints during practice and producing work for submission. Permission for one does not automatically settle the others. If the instruction is unclear, ask the teacher before using AI on the assessed task. Any required acknowledgement of assistance should describe the help actually used, in the format the school requests.
Understand what retrieval can establish
Retrieval-augmented generation, or RAG, combines a language generator with retrieved material. In their original 2020 paper, Patrick Lewis and colleagues tested models using a Wikipedia index and reported factuality improvements compared with a particular generation-only baseline. The experiments concerned language tasks, including question answering and question generation. They did not test whether a homework tool follows an individual teacher’s rubric.
The paper also reports imperfect results. It therefore provides no basis for claiming that retrieval eliminates hallucinations. For a proposed homework workflow, our practical inference is to inspect both the material supplied and the response produced: is this the current assignment, and does the explanation match it? A relevant document or a visible citation alone cannot establish that every generated statement is correct.
Keep people and privacy in view
UNESCO’s 2023 guidance takes a human-centred approach to generative AI in education. It calls for attention to privacy, age-appropriate use and the educational suitability of tools. These principles support a cautious approach to choosing homework assistance. They do not certify a particular product, prove that AI use improves attainment or replace a school’s decisions about safeguarding and teaching.
For an initial demonstration, our recommendation is to use a fictional assignment and invented learner responses. Avoid uploading identifiable pupil work, personal information or a teacher’s restricted materials merely to see what happens. For real classroom documents, first establish that the school permits their use in that service and that sharing permissions are in place. Respect the service’s age requirements and the school’s supervision arrangements.
Try one small, reviewable exchange
Here is an original exercise for an adult-led, permitted demonstration. Use a fictional task that asks for two differences between two short descriptions. Ask the tool to restate the task and identify the supplied criteria before offering a hint. Compare that restatement with the instructions yourself. Then let the learner attempt one difference and explain which words in the descriptions support it.
Review the hint against the same criteria. Did it help with comparison, introduce an unsupported claim or effectively write the answer? If the response conflicts with the assignment, stop and return to the classroom material. Record the unresolved question for the teacher. An instruction to admit uncertainty is useful to try, but its presence in a prompt does not ensure the tool will follow it.
Make the next teaching question visible
A brief observation from that fictional exercise could read: ‘Identified two differences; needed a prompt to connect each one to evidence.’ This records both the attempt and the assistance. Keep it descriptive rather than treating the tool’s judgement as a grade. The next conversation can then focus on whether the learner can explain the comparison with the supports their teacher allows.
The same scrutiny should apply when discussing SASHA: which instructions were used, what assistance was given and what can an educator check? The sources below establish no SASHA feature availability or effectiveness. Begin with a modest, inspectable goal: help the learner understand the task, preserve their own reasoning and leave the teacher a clear question when something remains uncertain.
Sources and further reading
These sources inform the discussion; they do not certify SASHA or establish its effectiveness. Follow each link to inspect the original context.