/w=3840,quality=90,fit=scale-down)
Raphael Thys
When it comes to using A.I. to enforce inclusive language, be careful what you wish for
Photo: Sei/Unsplash
While working at a fairly evolved digital organization, my colleagues developed a customized bot within the business messaging app, Slack. Most people who use Slack for communication know you can utilize bots inside the application to set reminders, schedule meetings, or even make jokes.
Our team, however, decided to explore how we could use an automated bot to encourage more inclusive language. While the team was diverse, we didn’t interact in a way that enabled non-verbal communication, which is vital in synchronizing common meaning. I found this out firsthand during one such experiment when we created a “Guys Bot.”
The premise of the Guys Bot was simple. When someone typed the masculine word guys in chat (as in “Good morning guys!”), the bot would send an automated message to the original poster. The message would state, “Excuse me, it looks like you said guys. Maybe consider more inclusive language like friends, pals, or teammates.”
This experiment prompted a debate at our virtual water cooler on preferred alternatives to the word guys when addressing a mixed-gender group of people. Friends, teammate, or mates were all frontrunners.
“I like folks. Folks Is hip, fresh, easy to type. And it’s used regularly in some of the other affinity groups,” chimed in one colleague.
The custom Slack emoji below this comment spun like a slot machine with thumbs up, hearts, and a 100 different other icons.
Other people chimed in.
“Yeah, folks is great.”
“I like to spell it folx. It’s edgy.”
It went for a few minutes. Eventually, the buzz died down, and we returned to our normal work routines. It was a harmless exchange, except for one thing. The word “folks” carries a completely different meaning to me. It makes my skin crawl.
While growing up just outside of Chicago, my best friend’s older brother was a member of the Latin Kings, a prominent nationwide gang. He was deep into the gang, which made him a target on the street for rival gangs. The Latin Kings’ main rival was the Folks — or Folk Nation — a loose affiliation of gangs including Gangster Disciples and others.
Around the time my friend’s brother turned twenty, rival Folks snatched him up. They stuck a blade in his gut and cut him from his belly button to his esophagus, like how you’d butcher a farm animal. He survived, but surgeons had to crack his rib cage open to keep him alive. He has a three-finger wide scar that goes up his torso to just above the t-shirt neckline.
Whenever I hear the word Folks, I think back to being a teenager learning how to recognize gang signs of friend and foe. Pitchforks thrown up — one of the Folks’ hand signs — meant danger to me. Pitchforks thrown pointing down in front of the wrong people could start a brawl or get you shot.
The Latin Kings have their own symbols. If you point your hand up to the sky and tuck your middle and ring-finger in — like you’re Spider-Man shooting web — you’re making the Latin Kings gang sign. If you do this with both hands and bring them together just right, you can make their five-pointed crown.
Latin King symbols published at stophoustongangs.org
This is a lot of context to squeeze into a 20-word Slack post on what the word “folks” meant to me, but after a few days of mulling it over, I responded to the thread. I’m all for inclusive language, but this single word was significant to me. I felt I should share the backstory, given how an innocent term can vary in meaning. So I crafted a response and hit send.
After two days of crickets, someone responded. “Wow, thank you for sharing.”
I didn’t expect my colleagues to strike the word folks from their vocabulary. I didn’t expect anyone to change their behavior at all. I merely wished to express a unique perspective among the group and remind them of their assumptions and biases. Even with an educated and diverse set of peers from many backgrounds, language and communication contain edge cases you can’t account for.
I point this out because I’ve spent much of my life parsing through written and spoken words. As a child, I learned basic sign language to communicate with deaf kids in my class. In the military, I was an Arabic linguist who translated thousands of hours of voices during the Iraq War. Now I’m a technologist leveraging natural language processing for human good. Given that I work in artificial intelligence (AI), I’m realizing something far more nefarious happening than a “Guys Bot.” Machines are destroying the way we communicate.
Consider that words are symbols developed to convey meaning. Human minds do this well, and we are seeing near-human performance in AI-powered language tools. Despite humanity’s innate ability to communicate and to build machines that mimic our human communication, we still don’t get it 100% right and often misunderstand people’s intended meaning.
It’s difficult to normalize language to be inclusive, inoffensive, and neutral because we all interpret meaning differently based on lived experiences. Yet this is exactly what we’re asking artificial intelligence to do for us at a billion-person scale. Once these rules are machine-enforced, they’re inescapable and create much bigger issues. AI-powered recommendations let us tab ahead to autocomplete our thoughts in real-time. These functions are now key features in our email, word processors, and browsers. The machines are nudging us into a sterile, common tongue, and we accept it as a matter of convenience. But what’s the cost? We’re trading uniqueness and precision for convenience. We’re also handing the power of acceptable speech to machine-enforced tools, not individuals.
Many people in the tech world are familiar with the story of Timnit Gebru, the ex-Googler ousted over her research on the ethics of large language models. Most people don’t know that the co-author of that paper, Emily Bender, addressed something more poignant about machines eating our language — AI will not capture underrepresented groups and shifts in the social norms of language.
MIT’s Technology Review went so far as to address this in a review of Gebru and Bender’s research.
”… An AI model trained on vast swaths of the internet won’t be attuned to the nuances of this vocabulary and won’t produce or interpret language in line with these new cultural norms.
It will also fail to capture the language and the norms of countries and peoples that have less access to the internet and thus a smaller linguistic footprint online. The result is that AI-generated language will be homogenized, reflecting the practices of the richest countries and communities.”
Because AI needs representation in data and massive scale to fit billions of common combinations, it can reflect only the input data — the language of privilege.
After unpacking this, I find three key points.
First, there is an active desire to only include neutral language in AI-powered tool. While many of us can agree that some language is hateful or offensive, society will always disagree on the margins. The discourse around edge cases is essential for society and shouldn’t be enforced by machines.
Decisions on acceptable language deployed at Google-scale will affect billions of people and could be governed by a small number of people and platforms. This is a significant power dynamic we’re actively living in. Who gets to decide what is acceptable or unacceptable language? I don’t believe this should be the few and the powerful. The dark part of my soul also registers this — excluding abusive language from the digital world doesn’t wipe it from society. In some cases, maybe offensive language should be included in language models.
Second, society is constantly changing. The issue here is that AI relies on scale to register change. Small movements could get squashed by louder voices and normalized, neutered language. If you can’t amass enough gravity to trend in the digital world, movements for change may not rise above the virtual noise floor.
Finally, as pointed out by Bender’s research, small groups with limited access won’t be captured by the algorithm. The poor, the few, and those that spend most of their time in the physical world will not be represented in online language. The Latin Kings in Humboldt Park, the far-removed populations in the Appalachian Highlands, and those with special accessibility considerations will not make it into the AI language blender. We either don’t have the data, won’t get the data, or the data won’t be big enough to make an impact. It will filter these groups out. AI-powered language tools then become exclusive, not inclusive.
In the physical world, people don’t communicate like a Wikipedia article. We use shorthand. Tone and tempo. Humor. Facial expressions. We can make a hilarious joke that — typed-out — would get flagged by every language censor online today. We can tell where people were born based on their accents and phrasing. Humans are truly unique, and our communication is dynamic.
This was evident in my recent conversation with Dr. Chris Tucker, author of the book A Planet of 3 Billion and Chairman of the American Geographical Society. Dr. Tucker is an InQTel alumni and has been in technology and entrepreneurship for a couple of decades. He asked me a fascinating question: “Who’s doing computer vision and AI for sign language?”
I responded, “Sign language isn’t just about recognizing hand gestures, although I have seen some related prototypes.”
Tucker agreed. American Sign Language (ASL) is not just a collection of hand gestures for verbs, nouns, and adjectives. It’s a language embedded in the culture of the Deaf and hard-of-hearing. When this group communicates, they are using the full spectrum of communication. Tone, tempo, body language, facial expressions. Word and symbol combinations that make individual communication uniquely human.
Tucker and I brainstormed on the data points we would have to collect to capture the level of the individuality of people communicating via sign language. Suggesting that we create an AI system that scrubs the uniqueness of Deaf people seemed preposterous. A criminal act.
“Why do we let the machines do it to us hearing people?” I asked Tucker and then recounted the Guys Bot story. Ironically, the one-handed sign for the Latin Kings is also the sign for “I Love You” in ASL.
The truth is we would never accept an AI system that reduces ASL to simple hand symbols. And we shouldn’t. Similarly, we should resist AI tools nibbling around the edges of all our communications. The goal should be to complete the fullest representation of individual communication. Period. This should apply to all communication systems that leverage AI. We should demand AI systems that promote individuality and uniqueness. We should demand transparency in how AI is developed and the data used to build it. Because we all deserve this.
But we’re currently letting an algorithm chomp away our uniqueness like Pac Man gobbling pellets on the screen.