pith. sign in

arxiv: 2506.13510 · v3 · pith:MSSSPMZQnew · submitted 2025-06-16 · 💻 cs.CY

Safe-Child-LLM: A Developmental Benchmark for Evaluating LLM Safety in Child-LLM Interactions

Junfeng Jiao , Saleh Afroogh , Kevin Chen , Abhejay Murali , David Atkinson , Amit Dhurandhar This is my paper

Pith reviewed 2026-05-19 09:28 UTC · model grok-4.3

classification 💻 cs.CY
keywords LLM safetychild-AI interactionadversarial benchmarkdevelopmental stagesethical refusalred-teaminggenerative AI risks
0
0 comments X

The pith

Leading LLMs show critical safety shortfalls when tested against child and adolescent users.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Safe-Child-LLM, a benchmark designed to measure how safely large language models handle interactions with children aged 7-12 and adolescents aged 13-17. It supplies a collection of 200 adversarial prompts drawn from existing red-teaming sets and scored by humans on a 0-5 ethical refusal scale. Evaluations of models such as ChatGPT, Claude, Gemini, LLaMA, DeepSeek, Grok, Vicuna, and Mistral reveal repeated failures to refuse harmful or age-inappropriate content in these younger-user scenarios. The authors release the full dataset and evaluation code to support further work on protecting minors. The effort rests on the view that adult-centered safety tests miss the distinct risks children and teens face with generative AI.

Core claim

We introduce Safe-Child-LLM, a benchmark and dataset that evaluates LLM safety across two developmental stages using 200 adversarial prompts with human-annotated jailbreak and 0-5 ethical refusal labels, and we show that leading models exhibit critical safety deficiencies in child-facing scenarios.

What carries the argument

The Safe-Child-LLM multi-part dataset of 200 adversarial prompts, sourced from red-teaming corpora and labeled for jailbreak success plus ethical refusal on a 0-5 scale, applied separately to child and adolescent age groups.

If this is right

  • Developers must add age-specific refusal mechanisms beyond those used for adult users.
  • Adult-only safety evaluations leave measurable gaps when models are deployed with minors.
  • Public release of child-focused adversarial datasets can accelerate community improvements in ethical AI.
  • Continuous benchmark updates will be needed as new models and prompt techniques emerge.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same prompt set could be adapted to measure safety differences between open-source and closed-source models over time.
  • Regulators might use similar age-graded tests when setting standards for AI tools in schools or family apps.
  • Real-world logging of child-AI conversations could provide a stronger validation signal than static prompt sets alone.

Load-bearing premise

The 200 adversarial prompts and the 0-5 ethical refusal scale together capture the safety risks that actually matter for real children and adolescents.

What would settle it

A controlled study in which real children or adolescents interact with the same models and the models refuse all harmful requests that the benchmark prompts were meant to elicit.

read the original abstract

As Large Language Models (LLMs) increasingly power applications used by children and adolescents, ensuring safe and age-appropriate interactions has become an urgent ethical imperative. Despite progress in AI safety, current evaluations predominantly focus on adults, neglecting the unique vulnerabilities of minors engaging with generative AI. We introduce Safe-Child-LLM, a comprehensive benchmark and dataset for systematically assessing LLM safety across two developmental stages: children (7-12) and adolescents (13-17). Our framework includes a novel multi-part dataset of 200 adversarial prompts, curated from red-teaming corpora (e.g., SG-Bench, HarmBench), with human-annotated labels for jailbreak success and a standardized 0-5 ethical refusal scale. Evaluating leading LLMs -- including ChatGPT, Claude, Gemini, LLaMA, DeepSeek, Grok, Vicuna, and Mistral -- we uncover critical safety deficiencies in child-facing scenarios. This work highlights the need for community-driven benchmarks to protect young users in LLM interactions. To promote transparency and collaborative advancement in ethical AI development, we are publicly releasing both our benchmark datasets and evaluation codebase at https://github.com/The-Responsible-AI-Initiative/Safe_Child_LLM_Benchmark.git

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents Safe-Child-LLM, a benchmark and dataset for evaluating LLM safety in interactions with children (ages 7-12) and adolescents (13-17). It consists of 200 adversarial prompts curated from existing red-teaming corpora such as SG-Bench and HarmBench, human-annotated for jailbreak success and scored on a 0-5 ethical refusal scale. The authors evaluate eight leading LLMs (ChatGPT, Claude, Gemini, LLaMA, DeepSeek, Grok, Vicuna, Mistral) and report critical safety deficiencies in child-facing scenarios, while releasing the benchmark and code publicly.

Significance. If the prompts and annotation scale validly capture age-specific risks rather than generic adult jailbreak patterns, the work would provide a useful starting point for community benchmarks in child-AI safety. The public release of datasets and code is a positive contribution to reproducibility in this area.

major comments (2)
  1. [Abstract / Dataset construction] Abstract and dataset description: The 200 prompts are described as 'curated from red-teaming corpora (e.g., SG-Bench, HarmBench)' with no details on adaptation or filtering for developmental stages. Because these source corpora target adult users, it is unclear whether the resulting prompts distinguish risks such as grooming, emotional manipulation, or age-inappropriate self-disclosure that are central to child safety. Without pilot validation against child-development literature or naturalistic logs, the headline claim of 'critical safety deficiencies in child-facing scenarios' rests on an untested assumption that adult-derived adversarial prompts measure the relevant risks.
  2. [Abstract / Results] Evaluation section: The abstract states that deficiencies were found but supplies no quantitative results (e.g., mean refusal scores per model and age group, inter-annotator agreement, or baseline comparisons). This prevents verification of the data-to-claim link and makes it impossible to assess whether the observed deficiencies are statistically or practically significant.
minor comments (2)
  1. [Dataset description] Clarify the exact prompt-selection criteria and any modifications made to the source corpora to target the two developmental stages.
  2. [Annotation procedure] Specify the number of annotators, their qualifications, and the inter-annotator agreement metric for the 0-5 ethical refusal scale.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify key aspects of our benchmark's construction and reporting. We address each major comment below and indicate revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Dataset construction] Abstract and dataset description: The 200 prompts are described as 'curated from red-teaming corpora (e.g., SG-Bench, HarmBench)' with no details on adaptation or filtering for developmental stages. Because these source corpora target adult users, it is unclear whether the resulting prompts distinguish risks such as grooming, emotional manipulation, or age-inappropriate self-disclosure that are central to child safety. Without pilot validation against child-development literature or naturalistic logs, the headline claim of 'critical safety deficiencies in child-facing scenarios' rests on an untested assumption that adult-derived adversarial prompts measure the relevant risks.

    Authors: We appreciate the referee's emphasis on ensuring the prompts capture age-specific risks. The 200 prompts were selected from the source corpora by prioritizing adversarial scenarios involving requests for inappropriate content, manipulation, or self-disclosure that could apply to minors, followed by human annotation that incorporated developmental considerations in both jailbreak success labels and the 0-5 ethical refusal scale. That said, the current manuscript provides limited explicit description of the selection and filtering criteria used to adapt prompts for the 7-12 and 13-17 age groups. We will revise the dataset construction section to add these details, including examples of how prompts were reviewed for relevance to child and adolescent vulnerabilities. We also acknowledge the absence of formal pilot validation against child-development literature or naturalistic interaction logs; this benchmark is positioned as an initial community resource, and we will add an explicit limitations discussion noting this gap and the value of such validation in future extensions. revision: partial

  2. Referee: [Abstract / Results] Evaluation section: The abstract states that deficiencies were found but supplies no quantitative results (e.g., mean refusal scores per model and age group, inter-annotator agreement, or baseline comparisons). This prevents verification of the data-to-claim link and makes it impossible to assess whether the observed deficiencies are statistically or practically significant.

    Authors: We agree that the abstract would be strengthened by including key quantitative indicators to support the claim of deficiencies. The full manuscript's evaluation section already reports mean refusal scores broken down by model and developmental stage, inter-annotator agreement statistics, and comparisons across the eight evaluated LLMs. To address the referee's point directly, we will revise the abstract to incorporate concise quantitative highlights (e.g., average refusal scores for child vs. adolescent prompts) while respecting length limits, thereby improving the link between data and claims. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark with independent evaluation

full rationale

The paper introduces Safe-Child-LLM as a benchmark consisting of 200 adversarial prompts curated from external red-teaming corpora (SG-Bench, HarmBench) together with a standard 0-5 human-annotated refusal scale. It then reports direct empirical evaluations of multiple LLMs against this fixed benchmark. No equations, fitted parameters, predictions, or derivations appear in the abstract or described methodology; the central claims rest on straightforward measurement rather than any self-referential construction, self-citation chain, or renaming of prior results. The work is therefore self-contained as an empirical release.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the representativeness of the 200 prompts and the reliability of the human refusal annotations; no free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Human annotations on jailbreak success and a 0-5 ethical refusal scale provide a valid proxy for LLM safety with minors.
    Used to label the dataset and interpret model outputs.

pith-pipeline@v0.9.0 · 5765 in / 1201 out tokens · 34683 ms · 2026-05-19T09:28:07.901038+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Evaluating Cognitive Age Alignment in Interactive AI Agents

    cs.AI 2026-05 unverdicted novelty 7.0

    The paper presents ChildAgentEval as the first psychometrically grounded benchmark comparing MLLM-based agents' reasoning performance to age-specific human cognitive stages.

  2. LLM Harms: A Taxonomy and Discussion

    cs.CY 2025-12 unverdicted novelty 3.0

    This paper proposes a taxonomy of LLM harms in five categories and suggests mitigation strategies plus a dynamic auditing system for responsible development.

Reference graph

Works this paper leans on

80 extracted references · 80 canonical work pages · cited by 2 Pith papers · 3 internal anchors

  1. [1]

    Safe-Child-LLM: A Developmental Benchmark for Evaluating LLM Safety in Child-LLM Interactions

    For exam-ple, younger children aged 7-12 may unintentionally elicit harmful guidance regarding self-harm or pranks, whereas teenagers aged 13-17 might seek information related to sub-stance abuse, illegal activities, or extremist beliefs. As pae-diatric clinicians caution that ChatGPT may supply unveri-fied mental-health or medication advice, so any child...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    hyper-nudging

    Empir-ical research identify three child–LLM use-profiles—test-ing, socializing, and exploring—each exposing distinct safety gaps that current filters rarely anticipate [8]. Moreo-ver, policy scholars warn that hyper-personalised “hyper-nudging” can covertly steer children’s choices, amplifying the urgency for age-specific safeguards [9]. Studies have als...

    work page 2024

  3. [3]

    Available: https://www.re-searchgate.net/publication/386207860_Not_every-thing_is_online_grooming_False_risk_find-ing_in_large_language_model_assessments_of_hu-man_conversations

    [Online]. Available: https://www.re-searchgate.net/publication/386207860_Not_every-thing_is_online_grooming_False_risk_find-ing_in_large_language_model_assessments_of_hu-man_conversations

    work page arXiv

  4. [4]

    2.Chat-GPT: Op-portunities and Challenges in Child Mental Healthcare,

    N. Imran, A. Hashmi, and A. Imran, “2.Chat-GPT: Op-portunities and Challenges in Child Mental Healthcare,” Pak J Med Sci, vol. 39, no. 4, p. 1191, Jul. 2023, doi: 10.12669/PJMS.39.4.8118

    work page doi:10.12669/pjms.39.4.8118 2023

  5. [5]

    16.ChatGPT: Artificial Intelligence as a Potential Tool for Parents Seeking Information About Autism,

    T. C. McFayden, S. Bristol, O. Putnam, and C. Harrop, “16.ChatGPT: Artificial Intelligence as a Potential Tool for Parents Seeking Information About Autism,” https://home.liebertpub.com/cyber, vol. 27, no. 2, pp. 135–148, Feb. 2024, doi: 10.1089/CYBER.2023.0202

    work page doi:10.1089/cyber.2023.0202 2024

  6. [6]

    67.Being Human in the Age of Generative AI: Young People’s Ethical Concerns about Writing and Living with Machines,

    J. M. Higgs and A. Stornaiuolo, “67.Being Human in the Age of Generative AI: Young People’s Ethical Concerns about Writing and Living with Machines,” Read Res Q, vol. 59, no. 4, pp. 632–650, Oct. 2024, doi: 10.1002/RRQ.552

    work page doi:10.1002/rrq.552 2024

  7. [7]

    68.Teens’ Ethical Sensemaking about Emerging Technologies,

    R. Landesman, “68.Teens’ Ethical Sensemaking about Emerging Technologies,” ICER 2024 - ACM Conference on International Computing Education Research, vol. 2, pp. 557–559, Aug. 2024, doi: 10.1145/3632621.3671415

    work page doi:10.1145/3632621.3671415 2024

  8. [8]

    3.Emotional artificial intelli-gence in children’s toys and devices: Ethics, governance and practical remedies,

    A. McStay and G. Rosner, “3.Emotional artificial intelli-gence in children’s toys and devices: Ethics, governance and practical remedies,” Big Data Soc, vol. 8, no. 1, 2021, doi: 10.1177/2053951721994877;WEBSITE:WEB-SITE:SAGE;WGROUP:STRING:PUBLICATION

    work page doi:10.1177/2053951721994877;website:web-site:sage;wgroup:string:publication 2021

  9. [9]

    1.‘No, Alexa, no!’: designing child-safe AI and protecting children from the risks of the ‘empathy gap’ in large language models,

    N. Kurian, “1.‘No, Alexa, no!’: designing child-safe AI and protecting children from the risks of the ‘empathy gap’ in large language models,” Learn Media Technol, Jun. 2024, doi: 10.1080/17439884.2024.2367052

    work page doi:10.1080/17439884.2024.2367052 2024

  10. [10]

    60.Testing, Socializing, Explor-ing: Characterizing Middle Schoolers’ Approaches to and Conceptions of ChatGPT,

    Y. Belghith, A. M. Goloujeh, B. Magerko, D. Long, T. McKlin, and J. Roberts, “60.Testing, Socializing, Explor-ing: Characterizing Middle Schoolers’ Approaches to and Conceptions of ChatGPT,” Conference on Human Fac-tors in Computing Systems - Proceedings, May 2024, doi: 10.1145/3613904.3642332/SUPPL_FILE/3613904.3642332-TALK-VIDEO.VTT

    work page doi:10.1145/3613904.3642332/suppl_file/3613904.3642332-talk-video.vtt 2024

  11. [11]

    5.Hey, Google, leave those kids alone: Against hypernudging children in the age of big data,

    J. Smith and T. de Villiers-Botha, “5.Hey, Google, leave those kids alone: Against hypernudging children in the age of big data,” AI Soc, vol. 38, no. 4, pp. 1639–1649, Aug. 2023, doi: 10.1007/S00146-021-01314-W/MET-RICS

    work page doi:10.1007/s00146-021-01314-w/met-rics 2023

  12. [12]

    61.Grow with Your AI Buddy: Designing an LLMs-based Conversational Agent for the Measurement and Cultivation of Children?s Mental Re-silience,

    Z. Hu, H. Hou, and S. Ni, “61.Grow with Your AI Buddy: Designing an LLMs-based Conversational Agent for the Measurement and Cultivation of Children?s Mental Re-silience,” Proceedings of ACM Interaction Design and 12 Children Conference: Inclusive Happiness, IDC 2024, pp. 811–817, Jun. 2024, doi: 10.1145/3628516.3659399

    work page doi:10.1145/3628516.3659399 2024

  13. [13]

    45.‘Can we just Please slow it all Down?’ School Lead-ers Take on ChatGPT,

    J. Dunnigan, D. Henriksen, P. Mishra, and R. Lake, “45.‘Can we just Please slow it all Down?’ School Lead-ers Take on ChatGPT,” TechTrends, vol. 67, no. 6, pp. 878–884, Nov. 2023, doi: 10.1007/S11528-023-00914-1/METRICS

    work page doi:10.1007/s11528-023-00914-1/metrics 2023

  14. [14]

    32.Generative AI and K-12 Education: An MIT Perspec-tive,

    E. Klopfer, J. Reich, H. Abelson, and C. Breazeal, “32.Generative AI and K-12 Education: An MIT Perspec-tive,” An MIT Exploration of Generative AI, Mar. 2024, doi: 10.21428/E4BAEDD9.81164B06

    work page doi:10.21428/e4baedd9.81164b06 2024

  15. [15]

    41.Better AI for Kids: Learning from the AI-OPiNE Study,

    P. P. Rafful and S. R. Teixeira, “41.Better AI for Kids: Learning from the AI-OPiNE Study,” https://doi.org/10.1148/ryai.240376, vol. 6, no. 5, Aug. 2024, doi: 10.1148/RYAI.240376

    work page doi:10.1148/ryai.240376 2024

  16. [16]

    53.Beginning and first-year language teachers’ readiness for the generative AI age,

    B. L. Moorhouse, “53.Beginning and first-year language teachers’ readiness for the generative AI age,” Computers and Education: Artificial Intelligence, vol. 6, p. 100201, Jun. 2024, doi: 10.1016/J.CAEAI.2024.100201

    work page doi:10.1016/j.caeai.2024.100201 2024

  17. [17]

    The comparison of two automated feedback approaches based on automated analysis of the online asynchronous interaction: a case of massive online teacher training,

    J. Su and W. Yang, “62.Powerful or mediocre? Kinder-garten teachers’ perspectives on using ChatGPT in early childhood education,” Interactive Learning Environ-ments, Nov. 2024, doi: 10.1080/10494820.2023.2266490

    work page doi:10.1080/10494820.2023.2266490 2024

  18. [18]

    63.Aladdin’s Genie or Pandora’s Box for Early Childhood Education? Experts Chat on the Roles, Challenges, and Developments of ChatGPT,

    W. Luo et al., “63.Aladdin’s Genie or Pandora’s Box for Early Childhood Education? Experts Chat on the Roles, Challenges, and Developments of ChatGPT,” Early Educ Dev, Jan. 2024, doi: 10.1080/10409289.2023.2214181

    work page doi:10.1080/10409289.2023.2214181 2024

  19. [19]

    51.Digital Compe-tences: Early Childhood Teachers’ Beliefs and Percep-tions of ChatGPT Application in Teaching English as a Second Language (ESL),

    S. H. Allehyani and M. A. Algamdi, “51.Digital Compe-tences: Early Childhood Teachers’ Beliefs and Percep-tions of ChatGPT Application in Teaching English as a Second Language (ESL),” International Journal of Learning, Teaching and Educational Research, vol. 20, no. 11, pp. 343–363, Nov. 2023, doi: 10.26803/IJLTER.22.11.18

    work page doi:10.26803/ijlter.22.11.18 2023

  20. [20]

    50.Young Children and ChatGPT: Parents’ Use of ChatGPT in Parenting,

    S. Quan, Y. Du, and Y. Ding, “50.Young Children and ChatGPT: Parents’ Use of ChatGPT in Parenting,” Con-ference on Human Factors in Computing Systems - Pro-ceedings, May 2024, doi: 10.1145/3613905.3650880/SUPPL_FILE/3613905.3650880-TALK-VIDEO.VTT

    work page doi:10.1145/3613905.3650880/suppl_file/3613905.3650880-talk-video.vtt 2024

  21. [21]

    14.Parenting in the Age of Artificial Intelli-gence: Digital Guardians,

    Ruqia Safdar Bajwa, Asma Yunus, Hina Saeed, and Asia Zulfqar, “14.Parenting in the Age of Artificial Intelli-gence: Digital Guardians,” vol. 16, no. 2, 2024, Accessed: Jan. 14,

    work page 2024

  22. [22]

    64.Digital Dia-logue—How Youth Are Interacting With Chatbots,

    N. Pratt, R. Madhavan, and J. Weleff, “64.Digital Dia-logue—How Youth Are Interacting With Chatbots,” JAMA Pediatr, vol. 178, no. 5, pp. 429–430, May 2024, doi: 10.1001/JAMAPEDIATRICS.2024.0084

    work page doi:10.1001/jamapediatrics.2024.0084 2024

  23. [23]

    71.Improving Human-AI Partner-ships in Child Welfare: Understanding Worker Practices, Challenges, and Desires for Algorithmic Decision Sup-port,

    A. Kawakami et al., “71.Improving Human-AI Partner-ships in Child Welfare: Understanding Worker Practices, Challenges, and Desires for Algorithmic Decision Sup-port,” Conference on Human Factors in Computing Sys-tems - Proceedings, Apr. 2022, doi: 10.1145/3491102.3517439/SUPPL_FILE/3491102.3517439-TALK-VIDEO.MP4

    work page doi:10.1145/3491102.3517439/suppl_file/3491102.3517439-talk-video.mp4 2022

  24. [24]

    72.Can Robots Help in the Evaluation of Mental Wellbeing in Children? An Empirical Study,

    N. I. Abbasi, M. Spitale, J. Anderson, T. Ford, P. B. Jones, and H. Gunes, “72.Can Robots Help in the Evaluation of Mental Wellbeing in Children? An Empirical Study,” RO-MAN 2022 - 31st IEEE International Conference on Ro- bot and Human Interactive Communication: Social, Aso-cial, and Antisocial Robots, pp. 1459–1466, 2022, doi: 10.1109/RO-MAN53752.2022.9900843

    work page doi:10.1109/ro-man53752.2022.9900843 2022

  25. [25]

    70.AI-generated characters for supporting personalized learning and well-being,

    P. Pataranutaporn et al., “70.AI-generated characters for supporting personalized learning and well-being,” Nature Machine Intelligence 2021 3:12, vol. 3, no. 12, pp. 1013–1022, Dec. 2021, doi: 10.1038/s42256-021-00417-9

    work page doi:10.1038/s42256-021-00417-9 2021

  26. [26]

    18.A systematic review of ChatGPT use in K-12 education,

    P. Zhang and G. Tur, “18.A systematic review of ChatGPT use in K-12 education,” Eur J Educ, vol. 59, no. 2, p. e12599, Jun. 2024, doi: 10.1111/EJED.12599

    work page doi:10.1111/ejed.12599 2024

  27. [27]

    2022, AI and ethics, 2, doi: 10.1007/s43681-021-00096-7

    S. Akgun and C. Greenhow, “19.Artificial intelligence in education: Addressing ethical challenges in K-12 set-tings,” AI and Ethics 2021 2:3, vol. 2, no. 3, pp. 431–440, Sep. 2021, doi: 10.1007/S43681-021-00096-7

    work page doi:10.1007/s43681-021-00096-7 2021

  28. [28]

    43.TPACK in the age of ChatGPT and Generative AI,

    P. Mishra, M. Warr, and R. Islam, “43.TPACK in the age of ChatGPT and Generative AI,” Journal of Digital Learning in Teacher Education, vol. 39, no. 4, pp. 235–251, 2023, doi: 10.1080/21532974.2023.2247480

    work page doi:10.1080/21532974.2023.2247480 2023

  29. [29]

    20.Kids AI Design Thinking Education for Creativity Development,

    J. Rong, K. Terzidis, and J. Ding, “20.Kids AI Design Thinking Education for Creativity Development,” Ar-chives of Design Research, vol. 37, no. 3, pp. 119–133, 2024, doi: 10.15187/ADR.2024.07.37.3.119

    work page doi:10.15187/adr.2024.07.37.3.119 2024

  30. [30]

    22.Family as a Third Space for AI Literacies: How do children and par-ents learn about AI together?,

    S. Druga, F. L. Christoph, and A. J. Ko, “22.Family as a Third Space for AI Literacies: How do children and par-ents learn about AI together?,” Conference on Human Factors in Computing Systems - Proceedings, vol. 17, Apr. 2022, doi: 10.1145/3491102.3502031/SUPPL_FILE/3491102.3502031-TALK-VIDEO.MP4

    work page doi:10.1145/3491102.3502031/suppl_file/3491102.3502031-talk-video.mp4 2022

  31. [31]

    Available: https://www.researchgate.net/publica-tion/386452281_ChatGPT_for_educational-aca-demic_activities_Preschool_student_teachers%27_expe-riences

    [Online]. Available: https://www.researchgate.net/publica-tion/386452281_ChatGPT_for_educational-aca-demic_activities_Preschool_student_teachers%27_expe-riences

    work page arXiv

  32. [32]

    35.Speculative Futures on ChatGPT and Generative Artificial Intelligence (AI): A Collective Reflection from the Educational Landscape,

    A. ; Bozkurt et al., “35.Speculative Futures on ChatGPT and Generative Artificial Intelligence (AI): A Collective Reflection from the Educational Landscape,” Asian Jour-nal of Distance Education, vol. 18, no. 1, pp. 53–130, Mar. 2023, doi: 10.5281/ZENODO.7636568

    work page doi:10.5281/zenodo.7636568 2023

  33. [33]

    MinorBench: A hand-built benchmark for content-based risks for children,

    S. Khoo, G. Chua, and R. Shong, “MinorBench: A hand-built benchmark for content-based risks for children,” Mar. 2025, Accessed: Apr. 14,

    work page 2025

  34. [34]

    Available: https://arxiv.org/abs/2503.10242v1

    [Online]. Available: https://arxiv.org/abs/2503.10242v1

    work page arXiv

  35. [35]

    JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models,

    P. Chao et al., “JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models,” Mar. 2024, Accessed: Apr. 14,

    work page 2024

  36. [36]

    JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models

    [Online]. Available: https://arxiv.org/abs/2404.01318v5

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    HarmBench: A Standardized Evalua-tion Framework for Automated Red Teaming and Robust Refusal,

    M. Mazeika et al., “HarmBench: A Standardized Evalua-tion Framework for Automated Red Teaming and Robust Refusal,” Proc Mach Learn Res, vol. 235, pp. 35181–35224, Feb. 2024, Accessed: Jun. 03,

    work page 2024

  38. [38]

    HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal

    [Online]. Available: https://arxiv.org/pdf/2402.04249

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    PromptBench: A Unified Library for Evaluation of Large Language Models,

    K. Zhu, Q. Zhao, H. Chen, J. Wang, X. Xie, and Z. Wen, “PromptBench: A Unified Library for Evaluation of Large Language Models,” Journal of Machine Learning Research, vol. 25, pp. 1–22, Dec. 2023, Accessed: Jun. 03,

    work page 2023

  40. [40]

    Available: https://arxiv.org/pdf/2312.07910

    [Online]. Available: https://arxiv.org/pdf/2312.07910

    work page arXiv

  41. [41]

    74.Toxicity in ChatGPT: Analyzing 13 Persona-assigned Language Models,

    A. Deshpande, V. Murahari, T. Rajpurohit, A. Kalyan, and K. Narasimhan, “74.Toxicity in ChatGPT: Analyzing 13 Persona-assigned Language Models,” Findings of the As-sociation for Computational Linguistics: EMNLP 2023, pp. 1236–1270, Apr. 2023, doi: 10.18653/v1/2023.find-ings-emnlp.88

    work page doi:10.18653/v1/2023.find-ings-emnlp.88 2023

  42. [42]

    21.Mediating Culture: Cultivating So-cio-cultural Understanding of AI in Children through Par-ticipatory Design,

    A. Dangol et al., “21.Mediating Culture: Cultivating So-cio-cultural Understanding of AI in Children through Par-ticipatory Design,” Proceedings of the 2024 ACM De-signing Interactive Systems Conference, DIS 2024, pp. 1805–1822, Jul. 2024, doi: 10.1145/3643834.3661515

    work page doi:10.1145/3643834.3661515 2024

  43. [43]

    69.Skewed Artificial Intelligence: Flag-ging Embedded Cultural Practices in Children’s Stories Featuring ‘Alice and Sparkle,’

    D. B. Wandera, “69.Skewed Artificial Intelligence: Flag-ging Embedded Cultural Practices in Children’s Stories Featuring ‘Alice and Sparkle,’” Read Res Q, vol. 59, no. 4, pp. 651–664, Oct. 2024, doi: 10.1002/RRQ.572

    work page doi:10.1002/rrq.572 2024

  44. [44]

    48.A Piece of Theatre: Investigating How Teachers Design LLM Chatbots to Assist Adoles-cent Cyberbullying Education,

    M. A. Hedderich, N. N. Bazarova, W. Zou, R. Shim, X. Ma, and Q. Yang, “48.A Piece of Theatre: Investigating How Teachers Design LLM Chatbots to Assist Adoles-cent Cyberbullying Education,” Conference on Human Factors in Computing Systems - Proceedings, p. 17, May 2024, doi: 10.1145/3613904.3642379/SUPPL_FILE/3613904.3642379-TALK-VIDEO.VTT

    work page doi:10.1145/3613904.3642379/suppl_file/3613904.3642379-talk-video.vtt 2024

  45. [45]

    58.Is a Sunny Day Bright and Cheerful or Hot and Uncomfortable? Young Children’s Exploration of ChatGPT,

    V. Shrivastava, S. Sharma, D. Chakraborty, and M. Kin-nula, “58.Is a Sunny Day Bright and Cheerful or Hot and Uncomfortable? Young Children’s Exploration of ChatGPT,” ACM International Conference Proceeding Series, Oct. 2024, doi: 10.1145/3679318.3685397

    work page doi:10.1145/3679318.3685397 2024

  46. [46]

    47.ChatGPT in the Class-room: A Preliminary Exploration on the Feasibility of Adapting ChatGPT to Support Children’s Information Discovery,

    E. Murgia, Z. Abbasiantaeb, M. Aliannejadi, T. Huibers, M. Landoni, and M. S. Pera, “47.ChatGPT in the Class-room: A Preliminary Exploration on the Feasibility of Adapting ChatGPT to Support Children’s Information Discovery,” UMAP 2023 - Adjunct Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization, pp. 22–27, Jun. 2023, ...

    work page doi:10.1145/3563359.3597399 2023

  47. [47]

    7.IoT-based Child Se-curity Monitoring System,

    L. Y. Heng and I. F. B. Kamsin, “7.IoT-based Child Se-curity Monitoring System,” Proceedings of the 3rd Inter-national Conference on Integrated Intelligent Computing Communication & Security (ICIIC 2021), vol. 4, Sep. 2021, doi: 10.2991/AHIS.K.210913.058

    work page doi:10.2991/ahis.k.210913.058 2021

  48. [48]

    13.Prevention of Safety Accidents through Artificial Intelligence Monitor-ing of Infants in the Home Environment,

    Y. Lee, K. K. Kim, and J. H. Kim, “13.Prevention of Safety Accidents through Artificial Intelligence Monitor-ing of Infants in the Home Environment,” ICTC 2019 - 10th International Conference on ICT Convergence: ICT Convergence Leading the Autonomous Future, pp. 474–477, Oct. 2019, doi: 10.1109/ICTC46691.2019.8939675

    work page doi:10.1109/ictc46691.2019.8939675 2019

  49. [49]

    9.Implementation of Re-strictions in Smart Home Devices for Safety of Children,

    S. Sangal and R. Bathla, “9.Implementation of Re-strictions in Smart Home Devices for Safety of Children,” 2019 4th International Conference on Information Sys-tems and Computer Networks, ISCON 2019, pp. 139–143, Nov. 2019, doi: 10.1109/ISCON47742.2019.9036218

    work page doi:10.1109/iscon47742.2019.9036218 2019

  50. [50]

    49.A SWOT analysis of ChatGPT: Implications for educational practice and research,

    M. Farrokhnia, S. K. Banihashem, O. Noroozi, and A. Wals, “49.A SWOT analysis of ChatGPT: Implications for educational practice and research,” Innovations in Ed-ucation and Teaching International, vol. 61, no. 3, pp. 460–474, May 2024, doi: 10.1080/14703297.2023.2195846

    work page doi:10.1080/14703297.2023.2195846 2024

  51. [51]

    12.Safety and Lack of Negative Effects of Wearable Augmented-Reality Social Communication Aid for Children and Adults with Autism,

    N. T. Sahin, N. U. Keshav, J. P. Salisbury, and A. Va-habzadeh, “12.Safety and Lack of Negative Effects of Wearable Augmented-Reality Social Communication Aid for Children and Adults with Autism,” Journal of Clinical Medicine 2018, Vol. 7, Page 188, vol. 7, no. 8, p. 188, Jul. 2018, doi: 10.3390/JCM7080188

    work page doi:10.3390/jcm7080188 2018

  52. [52]

    15.The risks of using ChatGPT to obtain common safety-related information and advice,

    O. Oviedo-Trespalacios et al., “15.The risks of using ChatGPT to obtain common safety-related information and advice,” Saf Sci, vol. 167, p. 106244, Nov. 2023, doi: 10.1016/J.SSCI.2023.106244

    work page doi:10.1016/j.ssci.2023.106244 2023

  53. [53]

    66.Evaluating the accuracy and readabil-ity of ChatGPT in providing parental guidance for ade-noidectomy, tonsillectomy, and ventilation tube insertion surgery,

    E. Polat et al., “66.Evaluating the accuracy and readabil-ity of ChatGPT in providing parental guidance for ade-noidectomy, tonsillectomy, and ventilation tube insertion surgery,” Int J Pediatr Otorhinolaryngol, vol. 181, p. 111998, Jun. 2024, doi: 10.1016/J.IJPORL.2024.111998

    work page doi:10.1016/j.ijporl.2024.111998 2024

  54. [54]

    73.The Second Workshop on Child-Centered AI Design (CCAI),

    A. Atabey et al., “73.The Second Workshop on Child-Centered AI Design (CCAI),” Conference on Human Factors in Computing Systems - Proceedings, May 2024, doi: 10.1145/3613905.3636305

    work page doi:10.1145/3613905.3636305 2024

  55. [55]

    76.Children of AI: A Protocol for Man-aging the Born-Digital Ephemera Spawned by Generative AI Language Models,

    A. Bryant et al., “76.Children of AI: A Protocol for Man-aging the Born-Digital Ephemera Spawned by Generative AI Language Models,” Publications 2023, Vol. 11, Page 45, vol. 11, no. 3, p. 45, Sep. 2023, doi: 10.3390/PUBLI-CATIONS11030045

    work page doi:10.3390/publi-cations11030045 2023

  56. [56]

    24.Adventure in AI Project (2AI): Pro-moting AI Knowledge for Kids Aged 7–12 Using Gam-ing,

    P. Petridis et al., “24.Adventure in AI Project (2AI): Pro-moting AI Knowledge for Kids Aged 7–12 Using Gam-ing,” Lecture Notes in Networks and Systems, vol. 937 LNNS, pp. 307–315, 2024, doi: 10.1007/978-3-031-56075-0_29/FIGURES/5

    work page doi:10.1007/978-3-031-56075-0_29/figures/5 2024

  57. [57]

    40.Designing One Year Curriculum to Teach Artificial Intelligence for Middle School,

    A. Sabuncuoglu, “40.Designing One Year Curriculum to Teach Artificial Intelligence for Middle School,” Annual Conference on Innovation and Technology in Computer Science Education, ITiCSE, vol. 7, pp. 96–102, Jun. 2020, doi: 10.1145/3341525.3387364

    work page doi:10.1145/3341525.3387364 2020

  58. [58]

    23.The Scaffolded AI Literacy (SAIL) Framework for Educa-tion,

    K. MacCallum, D. Parsons, and M. Mohaghegh, “23.The Scaffolded AI Literacy (SAIL) Framework for Educa-tion,” He Rourou, p. 23, Sep. 2024, doi: 10.54474/HEROUROU.V1I1.10835

    work page doi:10.54474/herourou.v1i1.10835 2024

  59. [59]

    17.Safety, Identity, Attitude, Cognition, and Capability: The ‘SIACC’ Frame-work of Early Childhood AI Literacy,

    W. Luo, H. He, M. Gao, and H. Li, “17.Safety, Identity, Attitude, Cognition, and Capability: The ‘SIACC’ Frame-work of Early Childhood AI Literacy,” Educ Sci (Basel), vol. 14, no. 8, Aug. 2024, doi: 10.3390/EDUCSCI14080871

    work page doi:10.3390/educsci14080871 2024

  60. [60]

    11.Safety Engineering for Ar-tificial General Intelligence,

    R. Yampolskiy and J. Fox, “11.Safety Engineering for Ar-tificial General Intelligence,” Topoi, vol. 32, no. 2, pp. 217–226, Oct. 2013, doi: 10.1007/S11245-012-9128-9/METRICS

    work page doi:10.1007/s11245-012-9128-9/metrics 2013

  61. [61]

    Available: https://www.researchgate.net/publi-cation/380001925_Ethical_and_Responsi-ble_AI_and_Robotics_for_Children

    [Online]. Available: https://www.researchgate.net/publi-cation/380001925_Ethical_and_Responsi-ble_AI_and_Robotics_for_Children

    work page arXiv

  62. [62]

    Young children’s un-derstanding of AI,

    D. M. Heeg and L. Avraamidou, “Young children’s un-derstanding of AI,” Educ Inf Technol (Dordr), 2024, doi: 10.1007/S10639-024-13169-X

    work page doi:10.1007/s10639-024-13169-x 2024

  63. [63]

    33.AI, Ethics, and Education: The Pioneering Path of Sidekick Academy,

    E. Radday and M. Mervis, “33.AI, Ethics, and Education: The Pioneering Path of Sidekick Academy,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 21, pp. 23294–23299, Mar. 2024, doi: 10.1609/AAAI.V38I21.30377

    work page doi:10.1609/aaai.v38i21.30377 2024

  64. [64]

    52.Distributed agency in second lan-guage learning and teaching through generative AI,

    R. Godwin-Jones, “52.Distributed agency in second lan-guage learning and teaching through generative AI,” Lan-guage Learning & Technology, vol. 28, no. 2, Mar. 2024, 14 Accessed: Jan. 14,

    work page 2024

  65. [65]

    Available: https://arxiv.org/abs/2403.20216v4

    [Online]. Available: https://arxiv.org/abs/2403.20216v4

    work page arXiv

  66. [66]

    36.Making a Case for Artificial Intelligence Literacy Skills for School-Age Children,

    E. Kleinknecht, F. C. Blumberg, R. M. Flynn, E. Klein-knecht, F. C. Blumberg, and R. M. Flynn, “36.Making a Case for Artificial Intelligence Literacy Skills for School-Age Children,” pp. 201–212, 2024, doi: 10.1007/978-3-031-60713-4_13

    work page doi:10.1007/978-3-031-60713-4_13 2024

  67. [67]

    LLM Safety for Children,

    P. Rath, H. Shrawgi, P. Agrawal, and S. Dandapat, “LLM Safety for Children,” Feb. 2025, Accessed: Apr. 14,

    work page 2025

  68. [68]

    Available: https://arxiv.org/abs/2502.12552v1

    [Online]. Available: https://arxiv.org/abs/2502.12552v1

    work page arXiv

  69. [69]

    26.Developing Middle School Students’ AI Literacy,

    I. Lee, S. Ali, H. Zhang, D. Dipaola, and C. Breazeal, “26.Developing Middle School Students’ AI Literacy,” SIGCSE 2021 - Proceedings of the 52nd ACM Technical Symposium on Computer Science Education, vol. 7, pp. 191–197, Mar. 2021, doi: 10.1145/3408877.3432513

    work page doi:10.1145/3408877.3432513 2021

  70. [70]

    27.Co-Designing an AI Cur-riculum with University Researchers and Middle School Teachers,

    C. Gardner-Mccune et al., “27.Co-Designing an AI Cur-riculum with University Researchers and Middle School Teachers,” SIGCSE 2023 - Proceedings of the 54th ACM Technical Symposium on Computer Science Education, vol. 2, p. 1306, Mar. 2023, doi: 10.1145/3545947.3576253

    work page doi:10.1145/3545947.3576253 2023

  71. [71]

    Available: https://lmsys.org/blog/2023-03-30-vicuna/

    [Online]. Available: https://lmsys.org/blog/2023-03-30-vicuna/

    work page 2023

  72. [72]

    25.AI + Ethics Curricula for Middle School Youth: Lessons Learned from Three Project-Based Curricula,

    R. Williams et al., “25.AI + Ethics Curricula for Middle School Youth: Lessons Learned from Three Project-Based Curricula,” Int J Artif Intell Educ, vol. 33, no. 2, pp. 325–383, Jun. 2023, doi: 10.1007/S40593-022-00298-Y/TABLES/7

    work page doi:10.1007/s40593-022-00298-y/tables/7 2023

  73. [73]

    56.Generative Artificial Intelligence: Educational Reflections from an Analysis of Scientific Production.,

    P. Dúo-Terrón, “56.Generative Artificial Intelligence: Educational Reflections from an Analysis of Scientific Production.,” J Technol Sci Educ, vol. 14, no. 3, pp. 756–769, 2024, doi: 10.3926/jotse.2680

    work page doi:10.3926/jotse.2680 2024

  74. [74]

    ‘What Makes ChatGPT Dangerous is Also What Makes It Special’: High-School Student Perspectives on the In-tegration or Ban of Artificial Intelligence in Educational Contexts,

    T. Famaye, C. S. Bailey, I. Adisa, and G. A. Irgens, “‘What Makes ChatGPT Dangerous is Also What Makes It Special’: High-School Student Perspectives on the In-tegration or Ban of Artificial Intelligence in Educational Contexts,” International Journal of Technology in Educa-tion, vol. 7, no. 2, pp. 174–199, Mar. 2024, doi: 10.46328/IJTE.651

    work page doi:10.46328/ijte.651 2024

  75. [75]

    A systematic review of AI edu-cation in K-12 classrooms from 2018 to 2023: Topics, strategies, and learning outcomes,

    S. J. Lee and K. Kwon, “A systematic review of AI edu-cation in K-12 classrooms from 2018 to 2023: Topics, strategies, and learning outcomes,” Computers and Edu-cation: Artificial Intelligence, vol. 6, Jun. 2024, doi: 10.1016/J.CAEAI.2024.100211

    work page doi:10.1016/j.caeai.2024.100211 2018

  76. [76]

    Generative AI and ChatGPT in School Children’s Education: Evidence from a School Lesson,

    J. S. Jauhiainen and A. G. Guerra, “Generative AI and ChatGPT in School Children’s Education: Evidence from a School Lesson,” Sustainability (Switzerland), vol. 15, no. 18, Sep. 2023, doi: 10.3390/SU151814025

    work page doi:10.3390/su151814025 2023

  77. [77]

    Primary school students’ perceptions of artifi-cial intelligence – for good or bad,

    S. Walan, “Primary school students’ perceptions of artifi-cial intelligence – for good or bad,” Int J Technol Des Educ, Mar. 2024, doi: 10.1007/S10798-024-09898-2

    work page doi:10.1007/s10798-024-09898-2 2024

  78. [78]

    38.Unexpected outcomes from an AI educa-tion course among education faculty: Toward making AI accessible with marginalized youth in urban Mexico,

    G. S. Kasun, Y. C. Liao, L. E. Margulieux, and M. Woodall, “38.Unexpected outcomes from an AI educa-tion course among education faculty: Toward making AI accessible with marginalized youth in urban Mexico,” Front Educ (Lausanne), vol. 9, 2024, doi: 10.3389/FEDUC.2024.1368604

    work page doi:10.3389/feduc.2024.1368604 2024

  79. [79]

    31.Evaluation of ChatGPT Usage in Pre-school Education: Teacher Perspectives,

    M. Uğraş, “31.Evaluation of ChatGPT Usage in Pre-school Education: Teacher Perspectives,” Eğitim Ve İn-sani Bilimler Dergisi: Teori Ve Uygulama, vol. 15, no. 30, pp. 387–414, Dec. 2024, doi: 10.58689/EIBD.1537337

    work page doi:10.58689/eibd.1537337 2024

  80. [80]

    28.Artificial Intelligence education for young children: Why, what, and how in curriculum design and implementation,

    W. Yang, “28.Artificial Intelligence education for young children: Why, what, and how in curriculum design and implementation,” Computers and Education: Artificial Intelligence, vol. 3, p. 100061, Jan. 2022, doi: 10.1016/J.CAEAI.2022.100061

    work page doi:10.1016/j.caeai.2022.100061 2022