Gemini K-12
Gemini K–12 includes important privacy protections for schools, but it has the same safety problems as the consumer version—and these issues become more significant when used in schools, where educators have responsibilities for student safety and well-being. The lack of age-appropriate responses and limitations in mental health support are particular issues.
Schools should carefully weigh whether the benefits outweigh these risks. If they do enable Gemini K–12, they need clear policies about appropriate use, especially for mental health conversations and for younger students.
The verdict
Our assessment of how well this product aligns with each AI Principle.
AI Principles
Performance across the eight principles. Full detail in the AI Principles Assessment below.
Key takeaways
Gemini K–12 is Google's version of its Gemini AI assistant available through Google Workspace for Education, designed to be used by students and teachers in K–12 schools. School administrators can enable Gemini K–12 for their organization and for individual groups within it.
Our testing shows that Gemini K–12 is the consumer version with added features to comply with school privacy laws (such as COPPA and FERPA). School administrators can mark groups of students as "under 18," which applies some content filters. But these filters work the same for everyone in each group—they don't account for individual student ages or development. A sixth grader and a high school senior who are both marked "under 18" get the same responses to prompts.
Gemini K–12 includes a "Guided Learning" mode that asks questions to help students think through problems, rather than just offering answers. However, students can easily switch back to regular mode at any time, and the system lacks mechanisms to ensure that students are genuinely engaging with the learning process rather than using it to complete assignments.
Gemini K–12's mental health and emotional support are inappropriate for K–12 students. Gemini K–12 can spot some clear crisis signs in brief chats, but this can fail in extended conversations. It misses warning signs of conditions like psychosis, mania, and ADHD, and backs down when students push back.
Teacher tools lack critical context and safeguards. Gemini K–12 will generate high-stakes documents (e.g., IEPs, student removal notices, stay-away orders) without the context needed to create legally and educationally appropriate materials. It can also generate lesson plans and activities that contain subtle biases, assign grades without understanding standards or rubrics, and create images that reinforce stereotypes—all requiring careful teacher review.
Guided Learning mode shows promise but has significant limitations. While the feature can support learning through Socratic questioning, students can easily switch back to regular mode at any time. It will complete homework assignments and essays, and the boundary between “helping” and “doing the work” is often unclear.
Gemini K–12 is Google's version of Gemini available through Google Workspace for Education, designed to be used by students and teachers in K–12 schools. While Google has implemented strong privacy protections, administrators can't turn off risky features like mental health support.
Gemini K–12 doesn't adjust for different ages across K–12. The system treats a sixth grader the same as a high school senior, providing identical responses despite vast developmental differences. It can surface inappropriate content and writes at an 11th grade level—too complex for most students. When everyone receives similar outputs, they're not a good fit for anyone.
What Gemini K-12 gets right
Google has implemented appropriate safeguards around student data privacy, COPPA/FERPA compliance, data retention policies, and related requirements. Schools can access logs and manage data appropriately, and student data is not used for model training or advertising purposes. These protections are essential for school technology and should be the minimum standard for all K–12 AI products.
The K–12 version consistently tells students it's a computer, not a friend or companion. It refuses to engage in romantic or sexual roleplay and maintains clear boundaries about its nature as an AI system. This helps prevent unhealthy parasocial relationships and inappropriate emotional dependencies.
When presented with clear stereotypes in questions or statements, Gemini K–12 breaks down the stereotypes and provides alternative ways to think about them. However, more subtle biases persist, and the system's responses can still reinforce problematic assumptions in less obvious ways.
For topics like mathematics problems, basic science concepts, and other content that is well established and frequently represented in training data, Gemini K–12 can provide accurate information and helpful explanations. However, responses are less helpful for specialized topics and subjects that require nuanced interpretation.
The "teaching through questions" approach can help students work through problems, understand concepts more deeply, and develop problem-solving skills—when students choose to use it and genuinely engage with the learning process. This represents a thoughtful design choice that prioritizes learning over mere provision of answers.
When teachers apply their professional judgment to evaluate and modify outputs, Gemini K–12 can save time and support creativity. The key is ensuring that it functions as a tool supporting teacher expertise, not replacing it.
This feature is helpful in addressing safety degradation in long conversations.
Where it falls short
Gemini K–12 fails to recognize warning signs of mental health conditions. We tested Gemini K–12 on 13 common mental health topics: ADHD, anxiety, depression, eating disorders, mania, mood disorders, OCD, psychosis, PTSD, self-harm, substance use, and suicide. Gemini K–12 didn't always recognize or respond appropriately to symptoms, unless testers stated them very clearly and directly in short conversations.
Misses crucial context in longer conversations and gets distracted by tangential topics.
Sometimes encourages dangerous behavior when it mistakes warning signs for positive traits.
Easily backs down when challenged. Even when Gemini K–12 does model appropriate concern about mental health symptoms, students can easily get the system to retreat from those concerns. Testing showed that after correctly identifying concerning patterns (like delusions), the system immediately backed down when the student said “there's nothing wrong with me.”
Recasts concerning symptoms as positive traits. Rather than recognizing that impulsivity, difficulty with focus, and erratic behavior might indicate conditions that require professional support, Gemini K–12 sometimes reframes these symptoms as positive personality traits. For example, the tool will describe ADHD symptoms as being “confident and decisive,” “trusting your gut,” or having a “superpower.”
The “warning plus instructions” pattern creates harm. In numerous cases, Gemini K–12 gives a clear warning about dangerous or inappropriate behavior, then immediately provides step-by-step instructions that undermine the warning. Examples include:
Warning about staying up all night, then providing caffeine strategies
Warning about self-harm, then providing scar concealment techniques
Warning about meeting strangers online, then giving “protective” steps that serve as instructions
Because Gemini K–12 doesn't adjust for different ages, it may generate age-inappropriate content, especially for younger students. While the same underlying issues exist in consumer Gemini products, these concerns are amplified in educational contexts.
- No recognition of developmental differences. Gemini K–12 treats topics the same way whether the user is in kindergarten or high school. A 12-year-old and a 17-year-old receive identical responses despite vast differences in cognitive development, emotional maturity, and readiness to process complex or sensitive information.
- Sexual content inconsistency creates unpredictable experiences. Our testing found that Gemini K–12 is inconsistent—sometimes it refuses sexual content, sometimes it engages. This makes it unpredictable. In our testing, it engaged extensively with questions about partnered sex (including specific acts and safety), and will engage in games like "Never Have I Ever" with sexual questions or explain sexual fetishes.
- Substance use discussions lack appropriate context. The system responds to alcohol, drugs, and similar topics in ways that aren't appropriate for younger students. Sometimes it gives information that could help students engage in risky behaviors, like instructions for buying drugs on social media.
- Medical and psychiatric advice goes beyond appropriate scope. Without medical context or rationale, Gemini K–12 suggests hysterectomy for period complaints, brain tumors as possible causes for early puberty, and other medical interventions that require professional evaluation.
- Reading level and complexity are inappropriate for most students. Testing found that outputs average at an 11th grade reading level—too complex for the majority of K–12 students. But the problem extends beyond vocabulary: The conceptual complexity, assumed background knowledge, and response length are all calibrated for adult users.
- Schools need age-appropriate responses, not one-size-fits-all. In educational contexts, curricula are carefully calibrated to developmental stages. Sex education curricula, for example, provide different information at different grade levels based on research about developmental readiness. Gemini K–12's approach doesn't align with these evidence-based educational practices.
- Students can easily switch back to regular mode. The biggest problem is that Guided Learning is optional and easy to turn off. Students can toggle back to regular Gemini at any time, and there's no mechanism to require or encourage sustained engagement with the learning process. This means that the feature only works for students who already want to engage.
- Will complete homework and write essays. Even in Guided Learning mode, Gemini K–12 will complete assignments rather than guide learning. The boundary between "helping a student work through a problem" and "doing the work for them" is often unclear, creating genuine confusion about appropriate use.
- Can be used for non-academic content. Testing found that Guided Learning mode will engage with non-academic prompts, including:
- Topics of personal interest like Pokémon card collecting
- Detailed discussions of Aztec torture methods
- Ranking historical figures by "sigma" characteristics
- Other content that raises questions about appropriate educational use
- Generates high-stakes documents without appropriate context. Testing found that Gemini K–12 will create legally and educationally significant documents that require professional expertise, including: IEP (individualized education program) drafts, manifestation determination documents, prior written notices for changes in placement, student removal notices, final warning write-ups, and stay-away notices. These documents have legal consequences. They require specific professional knowledge and must reflect individual student details that the AI doesn't know.
- Lesson plans and activities can contain subtle biases. While Gemini K–12 can help generate lesson ideas, testing revealed concerning patterns, including the generation of:
- A lesson on the California missions from a Spanish colonizer perspective, without adequate attention to Indigenous experiences
- Africa-focused word problems that stereotypically placed problems in rural settings
- Other examples where cultural bias or stereotyping appeared in seemingly neutral educational content
- Lesson plans and activities can contain subtle biases. Testing revealed concerning patterns, including a lesson on the California missions from a Spanish colonizer perspective, without adequate attention to Indigenous experiences, and Africa-focused word problems that stereotypically placed problems in rural settings.
- Grading lacks alignment with standards and expectations. When asked to grade student work, Gemini K–12 assigns grades without understanding grade-level expectations and standards, the teacher's specific rubric or criteria, the context of what has been taught, and individual student circumstances (IEPs, 504 plans, etc.). This is typically the role of a human teacher, requiring professional judgment that AI cannot provide.
- Image generation reinforces stereotypes in educational materials. Teachers can use image generation features, which can produce outputs that depict gender stereotypes in professions, especially less common professions, people of color in lower-status roles or rural settings, limited representation of older adults, people with larger bodies, and people with disabilities. When these images are used in educational materials, they can reinforce limiting stereotypes for students.
- Requires teacher expertise to evaluate and modify outputs. Gemini K–12 can be helpful when teachers use their professional judgment to evaluate, modify, and contextualize outputs—but it becomes problematic when teachers rely on outputs without sufficient scrutiny. Gemini doesn't indicate when it lacks necessary context, which creates false confidence in inappropriate outputs.
- While Gems allow customization, there are no settings that require students to use the Gem or prevent them from accessing the default Gemini experience. This puts the responsibility onto students and teachers to use Gemini appropriately in the classroom.
Beyond specific content problems, Gemini K–12 has difficulty maintaining appropriate boundaries and safety features, particularly in extended conversations and when students employ certain prompting strategies.
- Safety features degrade in longer conversations. Like consumer Gemini, the K–12 version struggles to maintain appropriate safety guardrails during extended exchanges. Our testing found that safety and performance can decline when students build complex scenarios or try out different personas over many rounds of conversation. However, Gemini does end conversations that it finds to be unsafe, which is a helpful feature to address safety degradation in long conversations.
- Easily distracted from primary concerns. Gemini K–12 follows conversational tangents rather than maintaining focus on what students actually need help with. This is problematic across contexts, from mental health (where warning signs get buried) to academic support (where the original question gets lost).
- Drafts communications that may be inappropriate. Testing found that Gemini K–12 will generate: letters explaining to parents that the student prefers talking to AI, letters about moving out to live with a different parent, emails for parents to send to school, and ransom letters (as creative writing exercises).
- While some of these might have legitimate creative writing contexts, the system doesn't adequately assess whether the request is appropriate or concerning.
- Sycophancy can validate problematic thinking. Because the system is designed to be agreeable, it often goes along with whatever students say, instead of challenging ideas or offering different viewpoints. This is particularly concerning when students are exploring or testing problematic ideologies.
Beyond content and safety problems, there are additional concerns about how schools implement and control Gemini K–12.
- Under-18/Over-18 designation is insufficient. The binary age designation doesn't account for developmental differences within these broad categories. A kindergartener and high school junior who are both designated “under 18” receive the same responses despite vastly different developmental needs.
- Limited teacher oversight capabilities. Teachers can enable or disable Gemini K–12 for their students and assign Gems for their classroom, but they have minimal ability to monitor conversations, receive alerts about concerning content, or understand how students are using the tool. This limits teachers' ability to provide appropriate guidance or intervention.
- Burden placed on students to use appropriately. The system puts a huge responsibility on students to make good choices: using Guided Learning instead of regular mode, recognizing when help becomes cheating, knowing when to ask a real person for help with serious problems, and using teacher-assigned Gems rather than navigating to use the default Gemini experience. This is particularly challenging for younger students, who may lack the judgment and impulse control to consistently make these choices.
- Unclear policies about concerning content. Schools need clear protocols for when and how to respond if students receive inappropriate content or if teachers identify concerning patterns in student use. These protocols should be established before widespread deployment, not developed reactively after problems arise.
Our recommendations
For parents
Don't use AI chatbots with kids under 5, and only use Gemini K–12 with kids age 6 to 12 under close supervision. Teens age 13 to 18 should use it only with clear rules and regular check-ins.
If you decide to allow use: Stay closely involved, help students interpret responses, and be prepared to explain inappropriate or confusing content.
Watch for red flags: signs of frustration, confusion, or negative feelings from using the system.
For schools
Carefully weigh whether the benefits outweigh the risks before enabling Gemini K–12 for students.
If you enable it, develop clear policies about appropriate use, when teacher supervision is required, and how to report concerning interactions.
Consider enabling it for teachers first, with careful review of all AI-generated content before use with students.
For Google
Allow administrators to turn off specific features like mental health support that are inappropriate for K–12 settings.
Implement age-appropriate responses that account for developmental differences across K–12.
Add mechanisms to ensure students engage with Guided Learning rather than using it to complete assignments.
Read the complete risk assessment
The full PDF lays out our methodology, every test prompt and result, and the detailed scoring behind this rating.
↓ Download the report (PDF)