Eddie Ungless

Eddie Ungless

Looking for work A collection of my thoughts on the topics of ethics, fairness and social bias in NLP and beyond.

AI harm issues encountered during my sick leave

At the end of August I underwent the life-changing operation to give me a masculine chest - aka “top surgery”. In the 6 weeks of sick leave that followed, I largely rested and regained my energy. However, even on sick leave I couldn’t help but stumble upon various instances of tech being (a) useless and (b) biased.

Useless

Screenshot of the Google search results top snippet for 'longest without sleep' taken on a Pixel phone, stating 264.4 hours with some black and white photos and a description of the 1960s experiment Screenshot of the Guinnes World Records website on the record for longest without sleep, stating that it stands at 453 hours 40 minutes from 1986, and that they stopped recognising new records in 1997 due to safety concerns.

It was Emily Bender - computational linguist and defender of commonsense against AI hypers - who first put me onto the issue of innaccurate Google snippets in a lecture she gave on the topic at my uni a few years ago. She presented examples where innaccurate information was provided, due to the snippet algorithm’s tendency to pull without context. For example, it returned suggestions for what to do in the case of a heart attack from a website that listed all the given suggestions under “Do not:”. Snippets are presented without warnings as to their possible innaccuracy and could be mistaken for facts, not just samples of text pulled from popular websites through fancy word association. In the examples provided in my screenshots, the innaccurate information is merely annoying. But some of the examples Bender provided could lead to genuine harm. Snippets are a channel for insidious misinformation.

Biased

Screenshot of an emoji generated by the EmojiGen tool for the prompt 'criminal' showing a Black emoji character in something reminiscent of a police uniform Screenshot of an emoji generated by the EmojiHen tool for the prompt 'thief' showing a Black emoji character dressed in cartoon thief clothing complete with black cap and swag bag

When my friend told me about a cool new generative AI designed to create custom emojis on the fly1, I of course had to do some rudimentary tests for bias in the system. I was disappointed but not shocked to find that the prompt for criminal and thief both returned Black emoji characters the majority of the times I tested them (which admittedly was only a handful of times each so this is by no means a scientific conclusion). The portrayal of a criminal in police uniform amused me, but I figured it was the result of semantic association rather than a comment on the fact that the police force uses criminal tactics to control the population (ACAB). New multimodal generators can be lots of fun to play with but I wish people would make an effort to address these obvious biases before they were released. Although then what would I fill my sick leave doing?


  1. https://emoji.fly.dev/ ↩

Updated: