Helpful, harmless and honest, are criteria of how we aspire AI models to behave. Whilst these criteria make some sense for task-related assistance such as coding or produce a report, it can start to break down when considering education - where we use AI to support the development of human intelligence. For example, **Helpfulness:** this is often defined through being useful and effective in assisting users with tasks, but pedagogical theories often counterintuitive. For example, you may wish to provide the child a task that guides them to be frustrated that produces motivation and improves memory consolidation when the answer-delivery is finally given in a well timed manner, like [[Productive failure|productive failure]]. **Harmlessness:** education is a [[Design science|design science]] and based in societal values, where the curriculum is often used as a way to reproduce the values of that society within their children. The decision of what types of knowledge is valued is inherently biased. Since in [[Intelligent tutoring system|tutoring systems]] the student is not a user but a subject, where AI's affordance of scalability presents [[(TODO) The dangers of standardised moral education with tutoring systems|dangers in standardised dissemination and indoctrination of values]]. **Honesty:** students lack strong prior knowledge and hence for effective learning teachers often use subtle lies that cover up complexities of subjects. Where is it right to lie?