Back to site

SchoolGround AI · Working paper

Language models in schools

Safety, equity and independent learning

Summary

Large language models (LLMs) power many of the AI chatbots students already use. Their educational value depends on how they are used: access alone does not establish safety, fairness or learning.

This paper examines three concerns: unreliable and unsafe responses, unequal support, and the replacement of independent practice. It proposes a shared school environment in which teachers guide AI use and research tests whether students learn. The aim is to integrate LLMs as educational tools while identifying and reducing their harms.

1. Student use is an existing reality

  • Students already use chatbots for schoolwork. In a Pew survey conducted in September–October 2025, 54% of U.S. teens aged 13–17 reported doing at least a little schoolwork with chatbot help. [1]
  • Use continues outside school controls. In another U.S. survey, 59% of students who encountered an AI block on a school device or network said they switched to a personal device. [2]
  • Implication: schools need an educational response to tools students already use. This does not mean that every task should involve AI.

2. Safety cannot be assumed

  • Fluent answers can be incorrect. They can invent facts, reasoning and references. Fluent language does not guarantee a correct answer. [3]
  • Safety measures can fail. Assessments found failures in responses to teen mental-health needs and in some teen safeguards. These findings concern specific products, versions and sensitive uses. [4] [5]
  • Privacy needs protection. Student information needs clear rules for collection, access and use. UNESCO calls for privacy protection and age-appropriate educational validation. [6]

Implication: check answers, protect information and provide a clear route to human help.

3. Support is not necessarily equitable

  • Use is unequal. In a UK undergraduate survey, more advantaged students were more likely to use certain AI tools. The study did not establish why. [7]
  • Reliance differs too. In a U.S. teen survey, 20% from households below $30,000 used chatbots for all or most schoolwork, versus 7% at $75,000 or more. [8]
  • The same meaning can receive different treatment. Experiments found racial bias when models judged speakers using different English dialects. [9]

Implication: provide shared access, teach everyone how to use it and test whether support is fair across learners.

4. Task performance is not the same as learning

  • Assistance can replace necessary practice. In a high-school maths trial, a general GPT-4 assistant improved assisted work but students later performed worse without AI than those who practised without it. A tutor with learning safeguards largely avoided that loss. [10]
  • Students report concerns about their learning. In a U.S. survey, 39% of AI-using teens reported missing out on learning when using it to finish assignments. [2]
  • Instructional design can improve outcomes. A structured AI tutor produced better immediate test results than an active-learning class in a trial with 194 university physics students. Long-term benefits were not tested. [11]

Implication: students must practise reasoning, checking sources and solving problems themselves. Assess what they can explain and solve without assistance. These studies do not establish that all LLM use weakens critical thinking.

5. A school-led approach to AI use

SchoolGround AI proposes treating LLMs as educational tools with defined purposes and limits. As with calculators, teachers should decide when assistance supports a learning goal and when independent work is necessary.

  • Teacher-defined context. Teachers choose the learning goal, the materials and the help students may use.
  • Shared access and guidance. Provide a common tool and guidance, with support for different learning needs.
  • Independent reasoning. Ask students to make an attempt, check the answer and explain what they learned.
  • Human responsibility. Teachers guide learning. Safeguarding staff handle concerns. The model supports their work.

The calculator comparison describes a role, not equal reliability. LLMs can invent answers and respond unpredictably, so they need additional checks. [3]

These are proposed design principles. Their educational value and practical limits must be tested. Include tasks completed entirely without AI.

6. Evaluation must guide implementation

SchoolGround AI proposes a shared environment built around school, grade and class knowledge. Research and classroom practice should shape it together.

  1. Define success before a pilot.Teachers and researchers define the skill, permitted help and success measures before a pilot.
  2. Measure independent learning.Compare ordinary teaching, structured AI support and independent practice. Check reasoning, retention and new problems without AI.
  3. Evaluate risks and distribution of benefits.Record wrong answers, unsafe responses, privacy failures and differences across student groups. Check teacher workload too.
  4. Revise, reassess or suspend.Publish limits, review model changes and suspend uses that do not meet agreed standards.

Conclusion: student use makes an educational response necessary; it does not make any particular product effective. SchoolGround AI aims to support learning while eliminating preventable harms. Its safety, fairness and learning benefits remain claims to test, not established outcomes.

References and scope

Sources checked on 27 October 2026. Surveys describe use; experiments test specific effects; guidance sets expectations. This is a selected evidence review, not an exhaustive systematic review. The school requirements above are our proposals.

  1. [1]

    Pew Research Center. 24 February 2026.

    How Teens Use and View AI (opens in a new tab)

    Survey. 1,458 U.S. teens aged 13–17; surveyed 25 September–9 October 2025. Schoolwork: 10% all/most + 21% some + 23% a little = 54%. Self-reports, not global prevalence.

  2. [2]

    Common Sense Media · Survey conducted by NORC. 18 August 2026.

    Teens in the AI Era: Schoolwork and Skills That Matter (opens in a new tab)

    Survey. 1,017 U.S. teens aged 13–17; 30 April–14 May 2026. Figures 5 and 9: learning concerns and bypassing blocked tools. The 39% refers to AI users; 59% to users who encountered a block. Self-reports, not causal effects.

  3. [3]

    National Institute of Standards and Technology (NIST). July 2024.

    Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (opens in a new tab)

    Risk framework. NIST AI 600-1, especially sections 2.2 (confabulation) and 2.4 (data privacy). Identifies risks; does not measure their frequency in schools.

  4. [4]

    Common Sense Media, with Stanford Medicine’s Brainstorm Lab. Updated 14 November 2025.

    AI Chatbots for Mental Health Support: Risk Assessment (opens in a new tab)

    Safety assessment. ChatGPT, Claude, Gemini and Meta AI in teen mental-health scenarios. A specific use case and testing period; not a rating of all classroom uses.

  5. [5]

    Common Sense Media · Youth AI Safety Institute. Updated 7 October 2026.

    ChatGPT for Teens: Risk Assessment (opens in a new tab)

    Product safety assessment. More than 4,000 test prompts around the August 2026 launch. Reported crisis-referral and parental-safeguard failures, alongside improved refusal of direct harmful assistance. Product-specific findings, not a test of every LLM.

  6. [6]

    Miao, F. & Holmes, W. · UNESCO. 7 September 2023.

    Guidance for generative AI in education and research (opens in a new tab)

    Education guidance. Guidance on privacy, human agency, age-appropriate use, inclusion and educational validation. Recommendations, not proof that a product is safe.

  7. [7]

    Stephenson, R. & Armstrong, C. · HEPI / Kortext. 12 March 2026.

    Student Generative AI Survey 2026 (opens in a new tab)

    Survey. 1,054 full-time UK undergraduates; December 2025. Report 199, pp. 9–12: methods and differences in use by social grade. The cause is unclear; cost alone is not established. Sponsored by education-technology company Kortext.

  8. [8]

    Pew Research Center. 24 February 2026.

    Demographic differences in how teens use and view AI (opens in a new tab)

    Survey analysis. Same survey as [1]. All/most schoolwork with chatbot help: 20% in households below $30,000 versus 7% at $75,000+. An association with income, not evidence that income or AI caused learning loss.

  9. [9]

    Hofmann, V., Kalluri, P. R., Jurafsky, D. & King, S.. 28 August 2024.

    AI generates covertly racist decisions about people based on their dialect (opens in a new tab)

    Peer-reviewed experiment. Nature 633, 147–154. Dialect experiments across 12 model versions, including matched-meaning text. Demonstrates bias in tested models; does not measure classroom learning or every current model.

  10. [10]

    Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. & Mariman, R.. 25 June 2025.

    Generative AI without guardrails can harm learning: Evidence from high school mathematics (opens in a new tab)

    Peer-reviewed randomized trial. PNAS 122 (26), e2422633122; Table 1. Nearly 1,000 students, one Turkish high school, GPT-4, mathematics. The guarded tutor largely avoided learning loss but did not significantly improve unassisted exam results. Short-term outcomes.

  11. [11]

    Kestin, G., Miller, K., Klales, A., Milbourne, T. & Ponti, G.. 3 June 2025.

    AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting (opens in a new tab)

    Peer-reviewed randomized trial. Scientific Reports 15, 17458. 194 eligible Harvard physics students, two lessons; immediate post-tests. Evidence for a specific structured tutor, not proof of long-term gains or effectiveness at every school age.

SchoolGround AI · Working paper 01Back to top ↑