Research ethics asks who a piece of work could harm, not only what it achieves.
Major conferences now require a broader impact or ethics statement.
Consent, privacy and licensing apply to scraped datasets as much as to people.
Dual use means the same method can serve a good purpose and a bad one.
Stating limitations honestly is part of the work, not a weakness in it.
Research ethics asks who could be harmed by a piece of work, not only what it achieves.
In machine learning that question is now a formal part of publication.
NeurIPS and other major venues require a broader impact or limitations statement.
Reviewers can and do reject work on ethical grounds alone.
Data is where most problems start, because the people in a dataset did not agree to be in it.
Scraped data carries consent, privacy and licensing questions that scale does not excuse.
Personal information in a training set can be memorised and later extracted from the model.
Dual use means one method serves a good purpose and a harmful one without any change to it.
Face recognition unlocks a phone and enables surveillance using the very same weights.
Release decisions therefore involve a judgement about who gets the capability.
Staged release and gated access are attempts to keep the benefit while slowing the harm.
Bias in a dataset becomes bias in a model and then becomes a decision about a real person.
Measuring that across groups is part of evaluating the work rather than an optional extra.
Honest reporting matters too, including failures, negative results and cherry-picked examples.
Stating what your method cannot do is a strength of a paper and never a weakness in it.
One question on this card. Every option explains itself once you pick, right or wrong.
11 concepts in this module, then a graded test with every answer explained.