Chapter 1: What reasons does NLP ethics need?↗
Bias in data and algorithms, hate speech, stereotypes, system design and data handling form a complicated ethical palette in NLP. They invite a further question: when a system becomes “better,” by which standard is it better?
This note discusses Prabhumoye and colleagues’ Case Study: Deontological Ethics in NLP (NAACL 2021). The authors select generalization and respect for autonomy from the deontological tradition to identify problems and suggest design directions. Their normative case studies should not be read as four experimentally validated solutions.
The social effects reviewed include exclusion, overgeneralization, bias confirmation, unequal topic exposure and dual use. The task is not simply to find an “ethics parameter” inside a model, but to make the problem, affected people and grounds for judgment explicit.
Chapter 2: Two deontological threads↗
Deontology is a family of ethical theories concerned with duties, rights and constraints on action. It does not mean obeying any existing rule, or preserving tradition for its own sake. Consequentialism should likewise not be caricatured as improvisation without principles.
| Comparison | A deontological perspective | A consequentialist perspective |
|---|---|---|
| Starting point | What duties, rights and constraints does the action involve? | What consequences follow from the action or rule? |
| One criticism of discriminatory hiring | It violates a right to equal treatment | It produces exclusion and harms academic and social outcomes |
These distinguish possible arguments, rather than exhaust either tradition. There are also consequentialist theories that evaluate rules.
2.1 Generalization: consider actions together with reasons↗
The paper’s test is not simply whether things would go badly if everyone acted that way. It asks: if everyone with the same reasons took the same action, would that undermine the reasons on which the action depends?
Its contract example concerns someone who pretends to intend to honor an agreement but keeps it only when convenient. The strategy depends on others believing the promise; universal adoption undermines that condition. The paper also considers plagiarism undertaken for professional advantage. This is an ethical argument, rather than a statistical prediction about citation counts or a test completed merely by pointing to bad consequences.
In this note’s notation, a proposed maxim can first be organized as:
Here, is the relevant context, the agent’s stated reasons and the action. The tuple makes assumptions more explicit. It does not determine which contexts count as the same, whether the reasons are genuine, or what the ethical verdict should be. Extracting a reason is different from establishing that it justifies an action.
2.2 Autonomy: what is known, and who decides?↗
Respect for autonomy concerns people’s ability to make decisions pertaining to themselves. The paper treats informed consent as one important expression: a person understands the relevant facts and likely consequences, then decides whether to permit an action.
An interpreter offers an intuitive example. Hiring someone to translate ordinarily permits them to convey the speaker’s words, without authorizing arbitrary changes of meaning. Relevant information need not include every technical detail of every interaction. The question is which meaningful differences arise between reasonable expectations and what actually happens.
Chapter 3: Four NLP cases↗
3.1 Question answering: what is owed beyond an answer?↗
The paper examines QA systems around 2021. Improved accuracy can still leave users unable to judge an answer’s reliability, provenance or failure. A system that relies on the user’s ignorance to conceal uncertainty may exploit the information gap it is supposed to help overcome.
Different kinds of explanation also need separating. Attention weights do not become evidence of input importance or a complete decision mechanism merely because they are easy to visualize. Jain and Wallace (2019) showed issues including different attention distributions yielding similar predictions in the models and tasks they studied. This does not establish that every form of attention analysis is forever useless.
A natural-language explanation can also be an account generated after the answer. The original paper worries that using another black box to generate an explanation pushes the problem further down the line; it does not announce that explanations have become part of genuine model reasoning. A plausible explanation need not faithfully describe how an answer was produced. An inspectable knowledge-graph path is one direction the authors suggest, but visibility does not guarantee correct knowledge, adequate coverage or warranted trust. For more on these distinctions, continue to When we try to understand its understanding.
3.2 Content detection: beyond chasing words↗
Hate, offensive content and microaggressions need not contain fixed trigger words. The same words can also occur in quotations, rebuttals and different community contexts. The paper discusses limitations of surface-level detection and proposes presenting potential implications and reasons for flagging content to assist human judgment.
That does not mean a system can directly read intent. Who is affected, which contexts alter interpretation and why annotators disagree remain questions to investigate. Generating a deeper account is a direction to test, rather than a solution that settles bias.
3.3 Machine translation: whom does the system speak for?↗
Machine translation (MT) does more than replace words. Choices of tone, politeness, kinship terms and idiom strategy can change how the speaker is understood.
The paper uses differences in granularity between English “Ms.” or “Mr.” and Japanese forms of address. An appropriate choice can require information about profession, relationship and setting, rather than word-for-word correspondence. Translating English “aunt” into Hindi may likewise require a more specific kinship relation.
Idioms reveal another choice. Drawing on Baker, the paper lists five strategies:
| Strategy | Operation |
|---|---|
| Similar meaning and form | Retain a similar idiomatic form when appropriate |
| Similar meaning, different form | Use a target-language expression with different literal content |
| Paraphrase | Explain the idiom’s meaning directly |
| Omission | Leave the idiom out at a particular point |
| Compensation | Restore a related tone or effect elsewhere |
No seasoning suits every dish. The question is what authorizes the system’s choice, as well as whether the result is fluent. Where ambiguity materially changes the message, the authors suggest options and their implications in text interfaces, or a follow-up question or uncertainty notice in speech. This is a conditional clarification proposal, rather than a requirement to confirm every word.
3.4 Dialogue: understand the system’s social role↗
In children’s interactions with smart toys and virtual assistants, parental permission to use a device need not imply knowledge of what the interaction entails. Children may also attribute feelings or authority beyond the system’s actual capabilities.
The authors therefore identify two research questions: what social role the system occupies, and how it responds to social cues. Their proposal is not simply to make systems more human-like. It calls for understanding where human-like interaction is appropriate and where machine boundaries should be made clear.
The proposed “kids mode” could adjust content, sources, topics and language, while informing parents about salient features. It is a design direction in the paper, not a proven developmental intervention. Parental decisions and children’s own understanding and interests are also distinct questions.
Chapter 4: Using NLP to assist ethical analysis↗
The preceding direction applies ethics to NLP. Conversely, NLP could help organize actions, reasons, entities and relations in texts. The paper proposes potential applications involving traffic-police conversations and court transcripts.
Extracting “why someone did this” can still omit context, mistake a later defense for an actual motive, or inherit annotators’ judgments. Extraction may help examine arguments without becoming an authority that automatically determines their justification.
Chapter 5: Two principles are not the whole framework↗
The paper explicitly acknowledges that generalization and autonomy do not exhaust NLP ethics. Privacy, respect and dignity in interaction may require further principles. Where rights and duties remain unclear, different frameworks need to be compared for what they contribute.
How to define a conceivable world, the same reasons or logical consistency remains contested. These principles help turn intuitive judgments into arguments that can be examined. They do not install a moral compass that can never drift.
The original paper’s references provide the complete bibliographic entries for the studies discussed here.