Humor and Translation: Biased Artificial Intelligence
By Mark Herman
A long-standing joke is that the difference between a language and a dialect is that a language is a dialect supported by an army and a navy. In truth, every language is a cluster of dialects. And even when two people speak the same dialect (that is, employ the same grammar and vocabulary), they can differ in pronunciation, speech rhythm, and speech speed, and even vary their speech to a considerable extent depending on the social situation in which they find themselves.
So, how does artificial intelligence (AI) cope with these variations? The short answer is that, as of now, it doesn’t.
In the May 2025 edition of Berkeley News Online1, a publication of the University of California at Berkeley (UCB), there is an interview with Professor of Linguistics Nicole Holliday. She is a sociolinguist, studying everything relating to language as a social object, focusing in particular on sociophonetics—the influence of social factors on the way people pronounce words, and the way variations in pronunciation can both reflect and shape social identities.
According to Holliday, harm arises when AI decides that the speech of certain people is substandard and therefore they cannot be hired for or promoted to certain positions or be welcome in certain social situations.
Holliday cites as an example the AI system Zoom Revenue Accelerator, which purportedly knows exactly how fast people should talk to be “professional.” This is absurd. According to Holliday, it “goes against what we know about successful human interaction, since speech rate is something that speakers negotiate amongst themselves within each conversation in real time.”
A young person may speak rapidly to another young person, but if he or she wants to be understood, the speech will be slower when talking to an older person or a person who has trouble understanding the first person because of hearing loss, or because the second person speaks a somewhat different dialect or has learned the first person’s language as a second language. And it is well known that some people are even adept at code switching: they are able to completely change their speech patterns depending on whom they are talking to. Nonetheless, an AI system can cause an employer to consider a job applicant as “not an effective communicator” regardless of how the applicant will actually “communicate” when performing the job in question.
Since AI systems are trained mostly on members of dominant social groups, they are most likely to be biased against marginalized groups. This means that a racist, agist, or xenophobic (or even a sexist) employer could knowingly use a particular AI system to discriminate while falsely claiming that the “objective” AI system, rather than the employer, is eliminating the undesired job applicants.
Obviously, an AI system trained on a limited set of data will be limited. Holliday gives the hypothetical example of a voice assistant trained on the speech of 1,000 white women from San Francisco. It will not perform well when other people use it because it “doesn’t know how to deal with regional, gender and age differences.”
On the other hand, large data sets present their own pitfalls. Holliday states that if an AI system were to “vacuum . . . up . . . [huge amounts of] speech [data] from the internet,” the data set would be so large that it would be a challenge, if not actually impossible, for the AI system developer to moderate or manage the content. Therefore, any biases inherent in the data could remain unknown even to the AI system developer.
In general, Holliday denies “hat[ing] AI and I’m not against technology. What I’m against is the technology pretending that it can do sociological things, because sociological things are fundamentally human. It can’t do that. These systems are making sociological judgments about language, which they are not equipped to do and which can cause significant social harms”
Notes
1. Park, Jessica. “Linguistics Professor Explores the Social Consequences of Biases in AI Linguistics,” Berkeley News Online (May 29, 2025).
Submit items for future columns via e-mail to mnh18@caa.columbia.edu. Discussions of the translation of humor and examples thereof are preferred, but humorous anecdotes about translators, translations, and mistranslations are also welcome. Include copyright information and permission if relevant.
