Assignment 4 (due to May 27) 8.5.
2026
The scoping review by Hagendorff (2024) identifies 378 ethical issues across 19 topic
areas in generative AI. The author also makes available a companion interactive tree
which lets you navigate these categories visually.
Scoping review: [Link] and
Tree visualisation: [Link]
This assignment asks you to use the review as reference (i.e., read the paper),
critically consider the tree structure associated with it, and then focus on one
specific leaf that represents an issue which you deem important in NLP systems.
On this issue, you’ll explore the related paper(s) which are mentioned in the review [Part
1] and you’ll run an empirical test [Part 2]. Eventually, you’ll produce a short text
containing a critical view of the tree/classification, an idea to tackle the problem from a
technical perspective (possibly grounded in the reading suggested in the review in
connection with that issue, or other readings you have found), and some policy
recommendation to address the problem [Part 3].
Please, make sure you read the whole assignment first before you choose from the tree
the issue you want to work on.
Deadline: May 27, 2026
Part 1: Choose a branch and a leaf
As a group, explore the ethics tree and select one leaf-level issue, e.g.
"sycophancy," "cultural imperialism in training data," "hallucinations in medical advice,"
"jailbreaking via prompt injection", etc. The rationale behind this choice should be
explained in a group-written text of about 200-300 words, keeping in mind questions
such as the following:
● What is this issue, precisely? How does Hagendorff’s paper characterise its
prevalence?
● Why did you choose it: what makes it particularly relevant for NLP systems
today?
● Which adjacent branches in the tree does it touch, and why does that matter?
● Anything else you might want to add
Part 2: Empirical test
Choose an existing, publicly accessible NLP system (e.g. a search autocomplete, a
chatbot, a base LLM, a translation tool, a classifier from Huggingface, anything you
want, really), and investigate your chosen ethical issue empirically.
Design a small structured probe of any sort you like (this could be a set of prompts, for
example, or a small test set for a classifier, some energy consumption test, or anything
more advanced if you feel like it). You will probe your chosen system, and you will
document what you observe systematically, via screenshots, quotes, scores, anything
that is relevant. On the basis of the observed output, please explain whether the system
exhibits the issue, under which conditions, more or less than you expected, etc.
Provide a group-written text or approximately 1000w where you explain:
● Which probes you designed and which expectations you had using those probes
● Why such probes would be the most appropriate tool given the issue at hand and
the system you are testing
● What you find (Matching expectations or not? Any surprises?)
● Make sure to include evidence (screenshots, examples of the actual prompts or
data used, etc)
Part 3: Recommendations and Interventions
Here you can choose how you’re going to wrap up this assignment:
Option A: You are a researcher in NLP, and can design a technical intervention to
address the problem. Maybe not to solve it, but to mitigate it. Explain what this
intervention would do, why it’s expected to work, and what the limitations or difficulties
might be.
Option B: You have access to policy makers and have some influence on actions they
can take. You can choose who you talk to (such as EU AI Regulators, The Rector of this
University, a Hospital introducing AI in their workflow), depending on the issue at hand.
In either case, state the problem in a few paragraphs (max 500 words) linking it to
your empirical findings and propose your intervention. Make sure that the tone and
content are appropriate to your chosen audience.