Safe words: Meet the teenager researching Gen Alpha's fast-changing slang, with an eye on online safety
Manisha Mehta, 14, recently presented a paper at a global tech summit. An app aimed at parents is in the works. AI must be trained differently too, she says.
It might interest you to know that the emoji depicting a cluster of kidney beans, among children under 16, stands for “self-identified incel”. Calling a girl a “bop” can mean she is promiscuous. And “Kys”, once “Know yourself”, can now be an abbreviation for “Kill yourself”.
It’s hard to know the difference, adults a finding. It might interest you know that AI can’t tell the difference either.
The fact that neither the algorithms nor her parents could identify signs of danger in Gen Alpha slang concerned Manisha Mehta so much that the 14-year-old high-school student from California decided last year that she had to do something about it.
She began by enlisting the help of her father, Virendra Mehta, who was then pursuing a PhD in artificial intelligence (which he has since completed). He put her in touch with computer science professor Fausto Giunchiglia at University of Trento, Italy, and the teenager and the scholar got to work.
First, they decided, they would try to assess just how large the gap was, by having parents compete a survey that tested their understanding of 100 terms of slang used by Gen Alpha (that’s the generation born between 2010 and 2024).
Their findings were so dramatic that she ended up lead author of a research paper (with Giunchiglia as supervisor and co-author) that she then presented at the Association for Computing Machinery Conference on Fairness, Accountability, and Transparency (ACM FAccT) in Athens, in June.
At the conference, Mehta discussed how she based the list on conversations from gaming platforms and on social-media content. Next, with help from Giunchiglia and her father, she designed a survey in which 24 Gen Alpha volunteers aged 11 to 14 identified the meanings and possible contexts of each term, and flagged the ones they saw as indicative of anger, hostility or bullying behaviour.
She then had 30 adults (parents, caregivers and professional moderators) and four AI models (GPT-4, Claude, Gemini and Llama 3) take the same survey.
“Our slang is often a source of mirth to grown-ups, but they don’t quite see the serious effects of bullying and even predation here,” Mehta says. “It’s as if there is a barrier in place between us.”
Code-breaking
Of the Gen Alpha participants, 92% were able to identify all phrases with the potential for harm, Mehta’s paper states. The adults were able to spot only a third of these, on average. The AI model with the highest score detected about 42% of the red-flag phrases; the model with the lowest score identified only 2.5%.
“Only two of the 30 adults were able to identify KYS as a term with possible negative connotations,” Mehta says.
It isn’t only about communication, of course. The teen points to the risks to online safety, and even the inability to moderate squabbles, when even millennial parents like hers, in their 40s, don’t know what their children are saying.
This isn’t like previous communication gaps either. As she puts it, she has noticed for a while, “with interest and mild concern”, how new terms suddenly appear, become inescapable, mysteriously disappear, and are replaced by new ones, in a rapid cycle.
All the adults, for example, missed the use of the skull emoji as something that turns a perceived compliment into an instance of mockery. (“You ate that up” by itself is a compliment, for instance. Add the skull and it becomes: “What a loser.”)
“The skull emoji’s versatility allows it to change a conversation’s tone from supportive to snarky, all while maintaining a cool facade, which is likely why it has gained popularity,” Mehta says.
What is it now?
Parents expressed worry, and often surprise, when they were shown their results. Many were more concerned by how dramatically AI had fallen short, Mehta says. “They had assumed that AI programs could step in and help with content moderation.”
Her research paper is just the start, the teen adds. She is currently working on an app that would allow adults to routinely take an updated survey, to familiarise themselves with new terms.
“I started working on this as a way to learn more about AI, and hang out more with my dad. I never expected the research to be of interest to so many others,” she says.
There wasn’t just interest, Giunchiglia points out.
When he missed his flight from Venice and she had to present their paper, the researchers in the audience were amused by Mehta’s age but impressed by her work, he adds. “The acceptance rate at ACM FAccT is usually very low. Maybe one out of every four or five papers is selected,” Giunchiglia says. Yet hers was, because it addresses a problem of today, with an eye on possible ways to tackle it.
After the app, Mehta says she would like to explore the role of therapy-bots and the implications of AI misunderstanding Gen Alpha prompts.
A key problem, she adds, is that her generation is so new.
“There is an overabundance of content created by previous generations on the internet, so AI programs will likely keep to the more common meanings of phrases,” she adds, “unless they are trained differently.”
