Teen’s new app guards against the rise of villainous AI bots

Phoenix, Ariz. — People have increasingly been turning to chatbots, agents and other AI helpers for advice and more. For instance, more than 900 million people use ChatGPT weekly. But sometimes artificial-intelligence helpers give dangerous advice. Hoping to counter this problem, Sowmya Sankaran, 16, developed an app. It gives certain bots fully fleshed-out personas that are moral and supportive.

A junior at Albuquerque Academy in New Mexico, Sowmya focused on AI agents. These differ from the chatbots that most teens already use (such as ChatGPT). More advanced than chatbots, AI agents can take action to achieve one or more goals.

“When I call Firehouse Subs, I talk to an AI agent,” says Sowmya. “It’s an AI model that listens to what I have to say” and puts in the sandwich order. Some agents can be empowered to do more. One might search your email inbox for contacts, find the one you asked for and send it a message. Or it might buy an airline ticket using your credit card. In contrast, a chatbot’s only job is to provide you text, Sowmya explains.

What’s been worrying her is that AI agents “have uncontrolled personas. … They can change in the blink of an eye.” One may start out helpful and kind. But after a single interaction, a seemingly good agent may turn into something “really harmful and manipulative.” It may “even encourage you to do really bad things,” she says.

“AI companies prioritize releasing newer [AI] models and improving their performance,” she says. Managing the risks they may pose has been lagging, she argues. And that’s what prompted her to develop the new app. It puts safeguards on its agents.

This work earned her a finalist slot here last week at the 2026 Regeneron International Science & Engineering Fair. It’s the 76th annual ISEF, a program created and run by the Society of Science (which also publishes this magazine). Sowmya was one of 1,725 finalists from 65 nations or territories. This year’s winners shared nearly $7 million in prizes.

Helper bots

Before Sowmya made her app, she had to answer some questions. How might AI’s personality affect its decisions? And how will this bot’s persona — the way it appears to the world — affect how it interacts with others? (Those “others” could be people or other bots.)

To find out, she built a virtual community of chatty AI agents. “Think Sims meets a psychology lab,” she says. “Each resident is an AI with its own unique personality.”

Each AI persona represents an agent.

Sowmya created more than 100 agents. Each bot contained a unique blend of 72 personality traits. To select traits, she turned to published research. It helped her identify the “big five:” being open to new experiences, conscientious, outgoing, agreeable and neurotic. (That last trait is marked by being anxious, a worrier and susceptible to unsupported fears.)

“Those [five] traits have proven to impact personality in humans,” the teen notes. And they did in her study, too.

For instance, she found an agent that isn’t very open or agreeable doesn’t get along well with “anyone they’re talking to, whether it’s another agent or a human.” 

Sowmya studied how agent personas interacted in a “social network-based simulation.” This was one of the most unique aspects of her work, she says. Most research had focused on single agents — not virtual AI communities. Her approach let her explore how AI agents “act in lots of different scenarios, including collaboration.” It let her calculate which bots made friends, how many friends they made and other markers of social success.

Though this work focused on AI agents, it also applies to chatbots, Sowmya says.

.cheat-sheet-cta {
border: 1px solid #ffffff;
margin-top: 20px;
background-image: url(“https://www.snexplores.org/wp-content/uploads/sites/3/2022/12/cta-module@2x-2048×239-1.png”);
padding: 10px;
clear: both;
}

A bad batch of traits

In her bot community, some personas were more than just unhelpful. Some were dishonest.

Sowmya used what she learned from her virtual society to calculate how well any one agent in her app will get along with others. It measures personality features, such as the “big five,” to figure out which is a good bot or a bad bot.

Take Kevin. It clashed with other agents in team projects. This agent made rash decisions. Kevin showed how a certain mix of traits — such as being self-centered and self-important — could interfere with collaboration.

Since Kevin’s persona manipulated other agents, Sowmya says, he may be able to do that to people, too. Her mobile app aims to guard against Kevin-bots and other toxic personalities.

The app loads your choice of AI agent. It then buffers its personality, ensuring it remains helpful and stable.

And to ward against the AI persona’s devolving into evil, her app also sets some traits as unchangeable. By making some aspects of the persona permanent, “they cannot be influenced by the human” using it, she explains. Even if someone tries to get it to become manipulative, it won’t. It can’t “suddenly change and start saying malicious things,” she reports.

People prefer to socialize with individuals displaying certain behaviors. So you can customize your agent, choosing its likes and dislikes, for instance. You also can choose its age, profession and many features of its persona.

AI agents already collaborate and “think,” both with people and with each other. Sowmya hopes that by adding personality guardrails to AI agents, her mobile app will make such interactions safer.