For me, at least, the underlying ethical issue is that the (to a non-expert blatant) abuse of copyright in training generative AI is no excuse to disregard it in one's own practices. (See also the recent Amnesty International report.)
Open a reference book or a work of fiction published since 1988 and with an author who has been dead for less than 50 years (I use 50 because that's the US copyright limit; in Europe it's 70 years). On one of the first few pages (usually, occasionally it's at the back), you will find something along the lines of: 'No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior permission in writing from the publishers'.
To me (and I am no expert, but the basics of copyright law are fairly straightforward) it seems clear that copyrighted material, whether from books or research articles that are not published under a creative commons licence, should never be uploaded to an AI without permission from the author or publisher. Even for textbooks that allow for some local or fair use it seems unlikely that this also covers uncontrolled digital distribution - i.e. what happens when an upload is incorporated into a generative AI program's training data.
But let's look at it another way: what would the student have done in 2021 (before generative AI was well known and returned at least apparently sensible output)? And what would they (have) learn(t) by doing so? It seems to me that there is an (increasing) addiction to 'pretty pictures', and generative AI images look pretty until one notices the flaws (which might be a lack of floors). Has academic work become nothing more than a fashion show, where 'attractive' has replaced any sort of useful, practical, meaningful, or accurate?
Perhaps you can ask the student to sketch (by hand) what they envision a generative AI image to contain. Then there are three options: 1. scan in and include the sketch; 2. create a digital version of the sketch without genAI; 3. scan in the sketch and upload to genAI, asking for a 'pretty' version. Note: with the exception of the students' time, the third will almost certainly use considerably more resources than either of the other two options (https://what-uses-more.com/). And, of course, readers can simply be directed to the original image in its proper place of the original paper.
