The first issue is that it primarily offers guidance about summative assessment, but at Cambridge, we basically do not do summative assessment except in examinations and dissertations. But in both these cases, the guidance seems wrong-headed to me.

1. For examinations, we do not need an AI policy at all, because we have a more general policy that whitelists allowed assistive devices, such as calculators, etc.
2. Applied to project dissertations, the University guidelines are both too inspecific and too lax. The issue should not be whether a student is honest about whether some material was generated by AI or not, but rather whether it is suitable for some material to be generated by AI or not (regardless of attribution).

For the latter, let me explain what I mean. One of the things we ought to be teaching our students is that the purpose of written text is to communicate from my mind to yours; the learning-assessment goal of a dissertation is not “causing there to exist a well-written summary of the project and its scientific merits” but rather “coming up with and synthesizing into text a well-written summary of the project and its scientific merits”. Properly disclosed AI-usage would not satisfy that assessment goal any more than properly disclosed ghostwriting would.

There is a human factor here too. The fact that some text is coming from someone’s mind whom I care about is what makes it valuable for me to read it; if the text isn’t coming from the person’s mind, it is by definition a waste of time and a hideous disrespect of my humanity. On the other hand, many dissertations will be about AI and will include text generated by LLMs as examples or case studies; that is perfectly natural and appropriate, and remains acceptable within the customary code of human interaction.

It is a good thing that we are in a position to set additional requirements in the project briefs, but I am concerned that among my colleagues there is not a clear consensus as to what is appropriate here, with some erroneously comparing the use of generative AI for text composition to the use of calculators for number computation. The real analogy would be between “using generative AI as a search engine or brainstorming engine” and “using a calculator for computation”, neither of which I think students should be required to disclose.

It is a huge weakness of the University guidance that it does not properly address formative assessment, which is most of what we do when teaching at Cambridge, e.g. in supervisions. Poorly-paid supervisors are likely to be the ones who have the most time wasted by generative AI slop. To put a finer point on it, I don't care at all whether a student discloses that their problem sheet (or parts of their problem sheet, more realistically) is written by AI. If I receive something with an AI disclosure on it, I won’t read the AI-generated text because it is not text qua communication (it isn’t coming from a human mind with whom I have a relationship that compells me to read what they have to say). I will advise the supervisors I engage to do the same. It would be strictly better for me to read the prompt, because the prompt is coming from the human mind; a turned in prompt would elicit partial credit, because some thinking has been done, but full credit would naturally be reserved for students who synthesised their initial thinking into prose that they wrote themselves.

So why is AI disclosure important? The issue of plagiarism is a complete red herring. The reason we need to require AI disclosure is rather so that we can delineate what materials faculty and supervisors may legitimately disregard entirely, and so that we can place a row of teeth behind a policy that requires students to engage with the supervisors using their own thoughts and ideas rather than wasting their supervisors’ unbelievably ill-compensated time.

I am not sure what I will write, but I plan to tell my students that they (1) obviously must follow the University guidance, but (2) if I catch wind that they are wasting their supervisors’ time by handing in solutions written by AI, even if they disclose it, they will be accountable to me.

On the other hand, I don’t care at all if they use generative AI to brainstorm or search or come up with ideas; I don‘t want to know about it at all. But, just like when look a fact up on Google, you remain responsible for both its correctness and its synthesis with what you already know. You can’t cite Google or Wikipedia, and you can’t cite Claude. Use whatever you want as much as you want, but then synthesise what you learned into something that you will be prepared to discuss in a supervision environment.

These principles go both ways: as a teacher, I will never ever give lectures that were written even in part by AI, nor will I ever use AI to assess student work. If there is some process that might be achieved by AI (like sending emails, marking assignments, or giving a lecture or a supervision, etc.) then that process probably can just be deleted: if it is important, than we obviously cannot use AI for it, and if it is not important, we obviously should not be doing it at all. Students pay a lot of money to be taught by world experts in their topics of interest; it is nothing short of fraud to give them instead what they could have gotten on their own for the cost of a £20 Claude or ChatGPT subscription.