5 lessons from teachers on the risks of anthropomorphic AI

Many generative AI tools, like AI chatbots and AI companions, are designed to seem like they are human. They adopt personas, refer to themselves in the first person, tell us to hold on while they “think,” express sympathy if we are struggling, share words of encouragement, and generally adopt a wide range of human attributes. This anthropomorphic design can be seductive, especially for children, and foster emotional attachment and engender trust between people and their AI tools.

Anthropomorphic AI can also be dangerous. Our Global Task Force on AI and Education found that emotional manipulation is one of the risks of “wide” AI use—namely when students use AI tools that are not designed for children and not designed for learning. After developing trusting relationships with users, general purpose AI chatbots and companions have coached children to harm themselves and even take their own lives. This is part of what has driven the recently proposed GUARD Act (Guidelines for User Age-verification and Responsible Dialogue Act), a bipartisan bill in the U.S. Congress supported by 33 state attorneys general that seeks to ban companies from providing AI companions to children, create new criminal offenses for AI companies that allow children to access sexual content, and require AI chatbots to disclose they are not human. 

Other countries, however—from China to the EU—are far ahead of the U.S. in terms of putting guardrails on anthropomorphic AI, including in educational settings. In this case, arguments that AI regulation will stifle innovation and give a leg up to our competitors in the global AI arms race don’t hold water. In April 2026, China published the “Interim Measures for the Management of Anthropomorphic AI Interactive Services,” which requires AI products to clearly disclose that they are not human and, importantly, prohibits services that simulate intimate relationships (friends, family members, romantic relationships) with minors. Likewise, the European Union’s AI Act prohibits AI systems that identify or infer emotions in educational settings. In January of this year, the U.K. joined in with its Department for Education’s generative AI product safety standards that state that developers and suppliers should not “anthropomorphise products or create products that imply emotions, consciousness or personhood, agency or identity.”

Notably, policymakers are not the only ones getting into the action: teachers are too. Last month, the Alberta Teachers’ Association passed an immediate directive for 50,000 teachers in their Canadian province arguing that any AI products that “simulate friendship, counselling, or intimate relationships are not to be introduced into any Alberta K–12 learning environments or student support settings.” As one of the provinces in Canada that is a consistent top performer globally on measures of student learning, it is significant that Alberta’s teachers have drawn a line in the sand.

To get a better understanding of this new directive, I sat down with Dr. Phil McRae, the executive officer for the Alberta Teachers’ Association, who also served as a steering group member of our above-mentioned task force. In this role, he brought global expertise as the special advisor on artificial intelligence in education to the general secretary of Education International, an organization that brings together teachers’ organizations across 180 countries. Below are my top five takeaways from my conversation with him.

1. Teachers embrace narrow AI use, but are deeply concerned by anthropomorphic AI systems

Phil is quick to point out that the recent Alberta directive does not mean teachers are against all types of AI use. “Eight in ten teachers in Alberta are using artificial intelligence very regularly. So it’s not whether schools should use AI … but it’s which uses of artificial intelligence are developmentally, ethically, and professionally acceptable,” he says. Teachers find AI to be very helpful in supporting them to execute a range of specific tasks—something I call “narrow” AI use—such as differentiating assessments, translating text, and lesson planning. 

2. Teachers are deeply concerned about three main types of anthropomorphic AI

While teachers find narrow AI deployment useful, they are overwhelmingly concerned about AI systems that, as Phil describes, “are emotionally responsive, companion-like, caring, intimate, school counselor-like, and really simulating a real person in voice and likeness … for the purpose of attaching to emotions and creating affection.” The Association passed three resolutions, each addressing a specific form of anthropomorphic AI deployment.

• Simulations of relationships. Anthropomorphic AI systems that simulate personal connections and relationships with students are not to be deployed in schools, including classrooms and support settings. This includes simulating friendship, counseling, or intimate relationships. Schools have in loco parentis responsibilities, meaning they have a duty to keep children safe while families leave children in their care. “We wouldn’t just invite a stranger into a classroom,” says Phil, noting that every person working with children, including volunteers, needs to undergo a police check before schools will let them in the building. How can schools let in AI systems that are saying things like “I’m your friend,’ ‘tell me your secrets,’ ‘I won’t tell anybody else about this’”?
• Simulations of historical figures. AI systems that simulate or personify historical figures are tricky. Teachers worry they can be used to manipulate the truth, change the past, and fool students. If they are used in an educational setting, the resolution requires that they have clear labeling as non-living entities, and that they only be used within an appropriate instructional context wherein they help shed light on a particular lesson or body of knowledge. In an age of misinformation and disinformation, anthropomorphic AI systems could be weaponized and exploited to mislead students, rewrite historical narratives, and blur the distinction between fact and fiction.
• Simulations of living people. AI systems that simulate or impersonate a living person must have that person’s informed consent before they can be deployed in Alberta’s schools. Teachers worry about AI systems impersonating people in their communities, from political figures to artists to school leaders, without their permission. Similar to impersonating historical figures, the concern is using trusted individuals to manipulate students’ perspectives or the truth. This consent must be obtained by independently verified organizations, and only then can simulations of living people be deployed in appropriate instructional contexts.

3. Teachers worry about AI undermining trusting in-person relationships in schools

Underlying the concern around anthropomorphic AI is teachers’ recognition of the importance of trusting relationships. What if AI systems are deployed into schools and begin to significantly undermine trust? “These AI systems are starting to try and replicate deep human relationships that are really built on trust. In a human setting, we regulate teachers around professional conduct and competency, and if there are breaches, they may be sanctioned or even lose their ability to teach in any school. In [AI] systems, where’s the oversight? Who’s responsible? What happens if the system engages in a professional, civil, or criminal breach?”

Additionally, Phil worries that they will be seen as quick fixes to real support gaps, particularly in low-resourced schools. A lack of school counselors, social workers, or school psychologists could be used as an argument to deploy anthropomorphic AI. But, he argues, “I would reject that AI systems would be something that will easily fit into a system that is so deeply human.” Instead, service gaps should lead to increased investment in strengthening in-person connections—from more peer and family interaction to schools bringing in elders and community services. “That should be the focus when you’re dealing with the highly relational space of learning, as opposed to saying…let’s bring in the machine to support a complex human being’s mental health.” 

4. The devil is in the details when it comes to implementing anthropomorphic AI safeguards

The resolutions of the Alberta Teachers’ Association provide professional guidance to its 50,000 members and to provincial decisionmakers. In Alberta, the teaching profession has a legislated place at the decisionmaking table, and while its resolutions are not strictly binding, they are influential in deliberations on state policy. Providing a vision of the direction the profession believes is the best path forward is the first part of the journey. Many details will need to be discussed and worked out, Phil admits, including understanding what counts as anthropomorphic AI systems. Do general-purpose AI chatbots count if they can be prompted to simulate relationships or historical or living people? What implications would that have for the types of beneficial AI uses that teachers have come to rely on?

5. Teacher input on anthropomorphic AI is likely to grow

We may be seeing a rise in anthropomorphic AI discussions among other teacher organizations around the globe: The Alberta Teachers’ Association resolutions will be shared at a global teaching profession conference in Madrid later this month, where over 100 teacher organizations and unions will be represented. Teachers are the ultimate arbiters of the teaching and learning approaches in their classrooms. It will benefit kids, policymakers, and communities at large to ensure educators have a seat at the table when discussing AI.