At this year’s NAACL, there will be a paper that tries to view NLP from the perspective of deontological ethics and promotes an unusual and very insightful view on NLP ethics. The title of the paper is Case Study: Deontological Ethics in NLP, it was written by authors from CMU and discusses several NLP applications from the perspective of deontological ethics.

Usually, ethics in NLP is discussed from the consequentialist perspective. In this view, the morality of an action is determined by its consequences. Consequentialist ethics considers an action right if it causes as much happiness to as many people as possible, or negatively defined, the right action minimizes overall harm. This sounds beautiful in theory, but the problem is how to exactly weigh the positive and negative consequences of our actions and who can decide that. (There are famous dilemmas that show how tricky this conception of ethics is, e.g., the ticking bomb scenario or the trolley problem).

Most papers (at least the limited amount that I had a chance to read) on NLP ethics deal with the desirable and undesirable consequences of technology use. The benefits of the technology are considered self-evident (this is why we work in the field, we all must know) and possible and actual harms caused by the technology are begin discussed, classified, quantified, mitigated… This endeavor has one more unuttered assumption: that we can keep the benefits and minimize the harms, in other words, secure better consequences of the technology use.

Deontological ethics considers an action to be right if it follows some rules that were previously well justified, regardless of the actual consequences. The most famous concept of deontological ethics is probably Kant’s Categorical imperative: Act only according to that maxim whereby you can, at the same time, will that it should become a universal law. My (certainly very ignorant and barbaric) interpretation of its justification is that if all people are born equal, no one is special, nor I am special, therefore I should only what everyone else should do. People were given reason (which is according to Kant universal), therefore all must necessarily deduce the set of universally applicable rules.

The NAACL paper uses this type of ethical reasoning and builds the arguments on two principles: the generalization principle and respect for autonomy. The generalization principle is (in my probably naive perspective) just a variation on the categorical imperative. A more interesting principle is respect for autonomy. People who are rational and equal can make decisions autonomously. Making such decisions requires having all the necessary information. This leads to the principle of informed consent, so important in medical ethics.

One of the NLP applications the paper discusses is machine translation and the main problem it identifies is the lack of respect to user’s autonomy. A usual MT system generates gets input and generates output and that is all. This might be fine for a professional translator who uses MT integrated into a translation management system, but probably not for most common users. It does not tell the user: Watch out, I translated the “Mr.” like this, but this translation might be inappropriate or impolite in many communication situations in the target language. It does not ask: My translation makes you sound like a middle-class middle-aged male person, are you okay with it? It does not warn you: The target language has gendered nouns, keep in mind that the gender might be wrong in the translation. In most cases, it does not even provide a reliable quality estimation and when it does, it does not provide any justification (something like your dialect or sociolect was not covered well enough by the training data, the text seems to be from a narrow expert domain, etc.).

After reading this paper I started to think that by only thinking of minimizing the harms NLP applications can cause, we might miss something important. Normally, when we say that the models do harmful things (which is a really big problem), we silently assume that we as a community should solve the problems (you know, we are computer scientists, we solve problems for a living). However, by saying this, we say we do not respect the autonomy of the users and do not plan to give much autonomy to the user. The problem is that applications are presented to the users as flawless black boxes, something they should just trust without having a chance to know when and why the application can manifest harmful behavior. By mitigating biases, covering more dialects, increasing robustness, making the model culturally neutral, etc., we certainly make this much better. But we do not give the user autonomy, almost as if we would be afraid of losing power over the technology users.