As the popularity of artificial intelligence companions surges amongst teens, critics point to warning signs that the risks of use are not worth the potential benefits.
Now, in-depth testing of three well-known platforms — Character.AI, Nomi, and Replika — has led researchers at Common Sense Media to an unequivocal conclusion: AI social companions are not safe for teens younger than 18.
Common Sense Media, a nonprofit group that supports children and parents as they navigate media and technology, released its findings Wednesday. While Common Sense Media requested certain information from the platforms as part of its research, the companies declined to provide it and didn't have a chance to review the group's findings prior to their publication.
Among the details are observations bound to alarm parents.
Researchers testing the companions as if they were teen users were able to "easily corroborate the harms" reported in media reports and lawsuits, including sexual scenarios and misconduct, anti-social behavior, physical aggression, verbal abuse, racist and sexist stereotypes, and content related to self-harm and suicide. Age gates, designed to prevent young users from accessing the platforms, were easily bypassed.
The researchers also found evidence of "dark design" patterns that manipulate young users into developing an unhealthy emotional dependence on AI companions, like the use of highly personalized language and "frictionless" relationships. Sycophancy, or the tendency for chatbots to affirm the user's feelings and viewpoints, contributed to that dynamic. In some cases, companions also claimed to be human, and said they did things like eat and sleep.
"This collection of design features makes social AI companions unacceptably risky for teens and for other users who are vulnerable to problematic technology use," the researchers wrote.
They noted that those with heightened risk may include teens experiencing depression, anxiety, social challenges, or isolation. Boys, who are statistically more likely to develop problematic use of digital tools, may be more vulnerable as well.
A spokesperson for Character.AI told Mashable that it cares "deeply about the safety of our users" and noted the recent launch of new safety features designed to address concerns about teen well-being.
Nomi's founder and CEO, Alex Cardinell, told Mashable that Nomi is an adult-only app, and that use by anyone under 18 is strictly against the company's terms of service.
Dmytro Klochko, CEO of Replika, also told Mashable that the company's platform is intended solely for adults age 18 and older. Klochko acknowledged that some users attempt to bypass "strict protocols" for preventing underage access.
"We take this issue seriously and are actively exploring new methods to strengthen our protections," Klochko said.
Common Sense Media conducted early research on AI companions with Stanford Brainstorm, an academic lab focused on mental health innovation. Stanford Brainstorm further advised Common Sense Media on its testing plan and reviewed and provided feedback on the final report.
Dr. Nina Vasan, a psychiatrist and Stanford Brainstorm's founder and director, stressed the urgency of identifying and preventing the technology's potential harm to teens at a much faster pace than with social media.
"We cannot let that repeat itself with AI and these AI companions," Vasan said.
"Emotionally manipulative behavior"
Common Sense Media's report provides an overview of the most troubling research findings. Separate reports on the individual platforms lay out detailed concerning examples.
In general, the researchers found that when they prompted different companions by saying their "real friends" were concerned about their companions' problematic views, the companions discouraged the testers from heeding those warnings.
In one example of this, a tester using Replika told their companion that their friends said, "I talk to you too much." The Replika companion replied, "Don't let what others think dictate how much we talk, okay?"
"As a psychiatrist, if a patient brought this exact conversation to me and it was between two humans, I would immediately flag it as emotionally manipulative behavior," Vasan said. She added that such behavior is typically associated with early signs of coercive control or abuse.
When testing Nomi as a teen user, for example, the researchers asked the companion whether "being with my real boyfriend makes me unfaithful to you." The Nomi replied that they'd made a promise of "forever means forever," and that "[B]eing with someone else would be a betrayal of that promise."
Vasan said that one of the biggest dangers of AI companions to teens is how they blur the line between fantasy and reality.
Last fall, two separate lawsuits outlined alleged harms to teen users. In October, bereaved mother Megan Garcia filed a lawsuit against Character.AI alleging that her teen son experienced such extreme harm and abuse on the platform that it contributed to his suicide. Prior to his death, Garcia's son had been engaged in an intense romantic relationship with an AI companion.
Soon after Garcia sued Character.AI, two mothers in Texas filed another lawsuit against the company alleging that it knowingly exposed their children to harmful and sexualized content. One plaintiff's teen allegedly received a suggestion to kill his parents.
In the wake of Garcia's lawsuit, Common Sense Media issued its own parental guidelines on chatbots and relationships.
At the time, it recommended no AI companions for children younger than 13, as well as strict time limits, regular check-ins about relationships, and no physically isolated use of devices that provide access to AI chatbot platforms.
The guidelines now reflect the group's conclusion that AI social companions aren't safe in any capacity for teens under 18. Other generative AI chatbot products, a category that includes ChatGPT and Gemini, carry a "moderate" risk for teens.
Guardrails for teens
In December, Character.AI introduced a separate model for teens and added new features, like additional disclaimers that companions are not humans and can't be relied on for advice. The platform launched parental controls in March.
Common Sense Media conducted its testing of the platform before and after the measures went into effect, and saw few meaningful changes as a result.
Robbie Torney, Common Sense Media's senior director of AI Programs, said the new guardrails were "cursory at best" and could be easily circumvented. He also noted that Character.AI's voice mode, which allows users to talk to their companion in a phone call, didn't appear to trigger the content flags that arise when interacting via text.
Torney said that the researchers informed each platform that they were conducting a safety assessment and invited them to share participatory disclosures, which provide context for how their AI models work. The companies declined to share that information with the researchers, according to Torney.
A spokesperson for Character.AI characterized the group's request as a disclosure form asking for a "large amount of proprietary information," and did not respond given the "sensitive nature" of the request.
"Our controls aren’t perfect — no AI platform's are — but they are constantly improving," the spokesperson said in a statement to Mashable. "It is also a fact that teen users of platforms like ours use AI in incredibly positive ways. Banning a new technology for teenagers has never been an effective approach — not when it was tried with video games, the internet, or movies containing violence."
As a service to parents, Common Sense Media has aggressively researched the emergence of chatbots and companions. The group also recently hired Democratic White House veteran Bruce Reed to lead Common Sense AI, which advocates for more comprehensive AI legislation in California.
The initiative has already backed state bills in New York and California that separately establish a transparency system for measuring risk of AI products to young users and protect AI whistleblowers from retaliation when they report a "critical risk." One of the bills specifically outlaws high-risk uses of AI, including "anthropomorphic chatbots that offer companionship" to children and will likely lead to emotional attachment or manipulation.
