Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Large Language Models Will Change The Way Children Think About Technology And Impact Every Interaction Paradigm

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Children who grow up conversing with large language models will expect every technology to remember them, talk back, and refine answers over time.

desk verdict A timely, honest position paper whose big prediction rests on an n=2 anecdotal study and an untested transfer assumption, but still worth reading as a prompt for discussion. read the letter →

arxiv 2504.13667 v1 pith:LUGA4VHF submitted 2025-04-18 cs.HC cs.AIcs.CL

Russell Beale This is my paper · ORCID
classification cs.HCcs.AIcs.CL
keywords largelanguagemodelschildreneducationconversationalinteractiondesignretrieval-augmentedgenerationuserexpectationstrustandexplainability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large language models, this paper argues, will change interaction design not mainly through new features but through the expectations of children who grow up using them. The author reviews early educational uses, points to survey evidence that most secondary students already use LLMs, and describes a six-week observation of two children revising for GCSE history with a retrieval-augmented chatbot tailored to their syllabus. From that episode he derives five design considerations: conversations will replace scroll-and-click as the default interaction, systems will be expected to remember context and preferences, narrowly specific tools will be tolerated less, explainability will be needed for trust, and LLM-assisted design will change how future designers work. The paper is an explicitly hopeful reflection rather than a controlled study, and it asks the design community to prepare for a shift it considers already underway.

What carries the argument

The mechanism is the expectation loop: repeated LLM use sets a new baseline for how interaction should feel, and that baseline transfers. The load-bearing technical object is retrieval-augmented generation (RAG), which pairs a generative model with an external retrieval step so answers are grounded in a chosen document set, such as an exam syllabus, and hallucination is reduced. In the paper the RAG chatbot is the working example of a focused, context-aware, conversational system, and the design considerations are the claimed consequences of children internalizing that style of interaction.

What would settle it

A controlled comparison would settle the transfer claim: give matched groups of children the same revision material, one group through a retrieval-augmented chatbot tailored to their syllabus and the other through conventional study materials, then measure exam performance, confidence, and later behavior when both groups use a standard menu-driven application. If the chatbot group shows no greater tendency to attempt conversation, expect continuity of context, or express frustration with a session-forgetting interface, the predicted shift in expectations is not supported.

Watch

Extended reading notes

Core claim

The central claim is that the biggest impact of LLMs on technology use will arrive through changed user expectations, not through any single capability. Children who learn by conversing with systems that remember context, refine answers across an exchange, and adapt to their specific syllabus will carry that baseline into every interface they meet, making menu-driven, session-forgetting software feel deficient. The author's own observation is offered as an early sign: his children used the bespoke chatbot mainly as a practice-and-feedback tool, generating exam-style questions and receiving model answers, and preferred this active, non-judgmental form of revision to passive study or group teaching. On that basis the paper claims teachers' roles will move toward higher-order skills and that designers must accommodate five consequences, from conversational defaults to a demand for transparency. This is a forward-looking argument, explicitly grounded in anecdote rather than controlled evidence.

Load-bearing premise

The load-bearing premise is that the observed six-week use of a bespoke revision chatbot by two children shows how children in general will relate to LLMs, and that the expectations formed there will carry over to every other technology they use; if those children are unrepresentative, or the effect belongs only to the carefully tailored tool, the five design considerations lose most of their force.

Editorial extensions

If this is right

  • Default interaction will shift from scroll, point, and click toward conversational exchanges in which users refine requests over multiple turns rather than crafting one perfect query.
  • Users will expect systems to remember their prior interactions and expressed preferences, so applications that present a blank state every session will feel deficient.
  • Narrowly specific, single-purpose tools will be tolerated less; designers will be pushed toward open data export, APIs, and software components that integrate with LLM-driven workflows.
  • Trust will depend on explainability: users will need systems that can justify their responses, because confident-sounding wrong answers make blind acceptance risky.
  • The designers of tomorrow, raised on LLMs and already assisted by them in practice, will spend less effort on prototyping and more on idea generation and problem solving.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension of the paper's expectation-transfer claim would be a between-subjects experiment: children who prepare for an exam with a context-aware conversational tutor should, compared with matched controls, show more frustration with and more natural-language attempts at a conventional menu-driven interface.
  • The five design considerations could be ordered into an age-cohort prediction: as LLM-native children age, tolerance for session-forgetting and single-purpose software should decline steadily, a pattern a longitudinal survey could discriminate from a short-lived novelty effect.
  • The paper treats explainability as a user demand, but the opposite over-trust path is equally plausible: children may accept fluent answers without questioning them, which would make explainability a safety constraint designers must impose rather than a feature users request. A log of how often children ask follow-up verification questions of a chatbot would separate these two cases.
  • Because the observed preference for the RAG chatbot could be driven by the syllabus grounding rather than by conversation itself, a direct comparison of RAG tutoring with an ungrounded chatbot on the same syllabus would isolate which ingredient creates the reported confidence and engagement.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper is a perspective on how large language models (LLMs) are changing children's learning and how these changes will reshape interaction design. It reviews prior work on LLMs in education, introduces retrieval-augmented generation as a means of focusing LLMs on curricular content, and reports a six-week self-ethnographic observation of two GCSE students using a bespoke RAG-based revision chatbot built by the author. From this observation, the paper derives five design considerations: a shift from pointing and clicking to conversational interfaces, expectations of continued context awareness, reduced tolerance for highly specific systems, a need for explainable systems and trust, and LLM-driven design. The conclusion includes a ChatGPT-generated list of impact areas the paper admits it missed, including social skills, safety, collaboration, inclusivity, and play.

Significance. If the central prediction is correct, the paper points to an important and timely design direction: children raised on conversational, context-aware LLMs may generalize those expectations to all interactive systems, forcing a move away from conventional menu-driven interfaces. The paper is candid about its methodological limitations, which is a genuine strength, and it offers a useful synthesis of relevant literature, concrete examples of RAG-based educational tools, and five clearly stated design considerations that could serve as hypotheses for future work. However, the empirical foundation is a single informal observation of two children, and the paper explicitly concedes that this evidence is anecdotal and not scientifically rigorous. The strongest part of the paper is its articulation of plausible design tensions; the weakest is the extrapolation from the observed scenario to the broad title claim about every interaction paradigm.

major comments (3)
  1. [§5 and §6] Section 5 concedes that the research is 'anecdotal and not scientifically rigorous' and describes an informal six-week observation of two children using a bespoke RAG system. Sections 6.1 through 6.5 then convert this observation into five general design requirements, including claims about what 'children' will 'begin to expect.' This is a load-bearing inferential leap: the observation lacks a control condition, a baseline, objective measures of learning or attitudes, and any sample that could support generalization. The paper should either reframe these as speculative design hypotheses with a concrete research agenda for testing them, or substantially broaden the empirical evidence. As written, the strength of the conclusion exceeds what the evidence can support.
  2. [§6.2] The transfer assumption is central and unsupported. The observation shows only that two teenagers used a syllabus-specific RAG chatbot for exam revision, yet §6.2 asserts that 'as children become used to models that know specifics about their circumstances, they will begin to expect to receive personalised responses' and that this will lead to 'expectations for other interactive systems to know what they did before.' No data in the paper measures expectations toward any non-LLM interface after LLM exposure, and there is no comparison group. A direct test—for example, comparing children's reactions to a conversational versus a menu-driven interface in a matched task after controlled LLM experience—would be needed to support this transfer claim. Without such evidence, the 'every interaction paradigm' claim is not established.
  3. [§7] The conclusion's use of ChatGPT to list omitted areas is self-undermining for the paper's scope. The generated list—'Social Skills No mention of how AI may affect children's social development. Safety Ignores risks like bias, overuse, and harmful content. Collaboration Overlooks tools for group learning. Inclusivity Lacks focus on diverse user needs. Play Misses LLMs' role in creative activities'—is reproduced, and the author agrees with it, but then the paper simply states 'we have no more space to explore these insights here.' These omissions directly contradict the title's claim that the impact covers 'every interaction paradigm.' The paper should either narrow its claims to the specific domains it actually addresses or expand the discussion to engage with the acknowledged gaps.
minor comments (5)
  1. [§2] The full text contains an obvious typo in the title: 'Think Abo ut Technology' should read 'Think About Technology.'
  2. [§6.2] The sentence 'they will begin to expect to receive personalised responses... this is likely to lead to expectations for other interactive systems to know wheat they did before to do it again' contains several typos and garbled phrasing ('wheat' should be 'what', and 'to do it again' is unclear). Please rewrite for clarity.
  3. [§4.1] The phrase 'making rigidly effective prompts less beneficial for learning' is unclear; it seems to refer to over-engineered prompt templates, but the intended contrast between structured prompting and exploratory interaction is not stated precisely.
  4. [References] Several references are incomplete, including [8], [12], [18], and [23], whose entries state 'details not publicly available' or 'in progress.' For a published paper, these citations need to be completed or replaced with accessible sources.
  5. [References] Reference [16] is described in the text as discussing how generative AI can create storyboards, but the cited paper 'Automated Essay Scoring: A Siamese Bidirectional LSTM Neural Network Architecture' appears to be about a different topic. Please verify this citation and correct it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper is a perspective piece whose design considerations are extrapolations from an anecdote, not forced by construction or self-citation.

full rationale

The paper contains no derivation chain, no equations, no fitted parameters, and no predictions computed from fitted inputs. The author's self-ethnographic RAG-based study is explicitly labeled anecdotal and not scientifically rigorous, and the five design considerations in Section 6 are presented as reflective extrapolations rather than as consequences forced by the observed data. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The closest potential concern—generalizing from two GCSE students using a bespoke tool tailored to their syllabus and exam board to a broad claim about children's future expectations—is an evidentiary overgeneralization, not a circular reduction: the conclusion that children 'will begin to expect' personalized responses is not contained in the premise that they used a personalized system. Similarly, asking ChatGPT to summarize missed areas in the conclusion is an illustrative aside, not load-bearing logical support. The paper's weaknesses are empirical scope and rigor, not circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or new entities are introduced because the paper is a perspective. The central claims rely on several unstated assumptions about LLM adoption and the transferability of interaction expectations.

assumptions (3)
  • domain assumption The trajectory of LLM capability and adoption will continue in the near term without major disruption.
    The paper's predictions of dramatic change assume continued improvement and uptake of LLMs; this is stated in the introduction rather than evidenced.
  • ad hoc to paper Interaction patterns learned by children in one context (a tailored revision chatbot) will generalize to their expectations for all other technology.
    This assumption underpins the leap from the six-week self-ethnographic study in Section 5 to the five design considerations in Section 6.
  • domain assumption The five proposed considerations are the most significant impacts for designers.
    The paper asserts this without comparative analysis; the conclusion admits possible missed areas, so this assumption is a simplification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models Will Change The Way Children Think About Technology And Impact Every Interaction Paradigm." pith.science (2026). https://pith.science/paper/LUGA4VHF

@misc{pith2026250413667,
  author       = {Pith},
  title        = {Pith review of: Large Language Models Will Change The Way Children Think About Technology And Impact Every Interaction Paradigm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LUGA4VHF}},
  note         = {Machine review of arXiv:2504.13667}
}
read the original abstract

This paper presents a hopeful perspective on the potentially dramatic impacts of Large Language Models on how we children learn and how they will expect to interact with technology. We review the effects of LLMs on education so far, and make the case that these effects are minor compared to the upcoming changes that are occurring. We present a small scenario and self-ethnographic study demonstrating the effects of these changes, and define five significant considerations that interactive systems designers will have to accommodate in the future.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 15 canonical work pages

  1. [16]

    Guoxi Liang, Byung-Won On, Dongwon Jeong, Hyun-Chul Kim, and Gyu Sang Choi. 2018. Automated Essay Scoring: A Siamese Bidirectional LSTM Neural Network Architecture. Symmetry 10, 12 (Dec. 2018), 682. https://doi.org/10.3390/sym10120682 Number: 12 Publisher: Multidisciplinary Digital Publishing Institute

    work page doi:10.3390/sym10120682 2018

  2. [1]

    Rania Abdelghani. 2024. Guiding the minds of tomorrow : conversational agents to tra in curiosity and metacognition in young learners . phdthesis. Université de Bordeaux. https://theses.hal.science/tel-04697786

    work page 2024

  3. [2]

    Rick Astley. 1987. https://www.youtube.com/watch?v=dQw4w9WgXcQ

    work page 1987

  4. [3]

    Bommasani, D

    R. Bommasani, D. Hudson, E. Adeli, et al. 2021. On the Oppo rtunities and Risks of Foundation Models. https://doi.org/10.48550/arXiv.2108.07258

    arXiv doi:10.48550/arxiv.2108.07258 2021

  5. [4]

    Ariel Han and Zhenyao Cai. 2023. Design Implications of G enerative AI Systems for Visual Storytelling for Young Lear ners. In Proceedings of the 22nd Annual ACM Interaction Design and Children Conference . ACM, 470–474. https://doi.org/10.1145/3585088.3593867

    arXiv 2023

  6. [5]

    Holmes, M

    W. Holmes, M. Bialik, and C. Fadel. 2019. Artificial Intelligence in Education: Promises and Implicat ions for Teaching and Learning . Center for Curriculum Redesign, Boston, MA. http://curriculumredesign.org/wp-content/uploads/AIED-Book-2019.pdf

    work page 2019

  7. [6]

    Zihui Hu, Hanchao Hou, and Shiguang Ni. 2024. Grow with Yo ur AI Buddy: Designing an LLMs-based Conversational Agent f or the Measurement and Cultivation of Children’s Mental Resilience. In Proceedings of the 23rd Annual ACM Interaction Design and Chi ldren Conference. ACM, 811–817. https://doi.org/10.1145/3628516.3659399

    arXiv 2024

  8. [7]

    Jaeho Jeon. 2024. Exploring AI chatbot affordances in the EFL classroom: Young learners’ experiences and perspectives. Computer Assisted Language Learning 37, 1-2 (2024), 1–26. Publisher: Taylor & Francis

    work page 2024

Show all 28 references
  1. [8]

    X. Ji, C. D’Arcy, and P. Velasco. 2022. A Systematic Revie w on Conversational AI in Language Education: Applications and Challenges. Reference details not publicly available

    2022

  2. [9]

    DU Jinming and Ben Kei Daniel. 2024. A systematic review o f AI-powered chatbots in EFL speaking practice: Transformi ng language education. Computers and Education: Artificial Intelligence (2024), 100230

    2024

  3. [10]

    There’s a new kind of coding I ca ll ’vibe coding’

    Andrej Karpathy. 2025. "There’s a new kind of coding I ca ll ’vibe coding’... ". https://x.com/karpathy/status/1886192184808149383

    2025

  4. [11]

    Kidd and B

    C. Kidd and B. Y. Hayden. 2015. The psychology and neuros cience of curiosity. Neuron 88, 3 (2015), 449–460

    2015

  5. [12]

    T. H. Kung, M. Cheatham, A. Medenilla, et al. 2022. Perfo rmance of ChatGPT on the United States Medical Licensing Exam: Potential for AI-Assisted Medical Education. Preprint or conference article; detail s in progress

    2022

  6. [13]

    S. J. Lee and K. Kwon. 2024. A systematic review of AI educ ation in K–12 classrooms from 2018 to 2023: Topics, strategi es, and learning outcomes. Computers and Education: Artificial Intelligence 6 (2024), 100211. https://doi.org/10.1016/j.caeai.2024.100211

    2024

  7. [14]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio P etroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler , Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances i...

    2020 doi arXiv

  8. [15]

    Xiu Li, Aron Henriksson, Martin Duneld, Jalal Nouri, an d Yongchao Wu. 2024. Supporting Teaching-to-the-Curricul um by Linking Diagnostic Tests to Curriculum Goals: Using Textbook Content as Contex t for Retrieval-Augmented Generation with Large Language M odels. In Artificial ...

    2024 doi

  9. [17]

    Loewenstein

    G. Loewenstein. 1994. The psychology of curiosity: A re view and reinterpretation. Psychological Bulletin 116, 1 (1994), 75–98

    1994

  10. [18]

    MacNeil, L

    C. MacNeil, L. Zhao, and J. Turner. 2022. Facilitating C ode Explanation with GPT-3: A Study in Computing Education. Preprint or under review; details not provided

    2022

  11. [19]

    Mehrabi, F

    N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. G alstyan. 2021. A Survey on Bias and Fairness in Machine Learn ing. Comput. Surveys 54, 6 (2021), 1–35. https://doi.org/10.1145/3457607

    2021 doi

  12. [20]

    Steven Mintz. 2025. Transforming History Education In to a More Immersive, Interactive Process of Inquiry and Disc overy. https://www.insidehighered.com/opinion/blogs/higher-ed-gamma/2024/11/25/using-ai-make-history-teaching -and-learning-more-impactful

    2025

  13. [21]

    Sue Sentance, William Marsh, and Jane Waite. 2020. Pare ntal Involvement in Computer Science Education and Computi ng Activities at Home. ACM Transactions on Computing Education (TOCE) 20, 4 (2020). https://dl.acm.org/doi/10.1145/3440890

    2020 doi

  14. [22]

    Seymour, Q

    W. Seymour, Q. Brown, and M. Jones. 2021. Alexa, Can I Pro gram You? Student Perceptions of Conversational Artificial Intelligence Before and After Programming Alexa. In IDC 2021. ACM. https://dl.acm.org/doi/10.1145/3459990.3460730

    2021

  15. [23]

    Tai and N

    H. Tai and N. Chen. 2020. Addressing Low Willingness to C ommunicate with an AI Chatbot in EFL Contexts. Reference det ails not publicly available

    2020

  16. [24]

    UNICEF. 2025. Policy Guidance on AI for Children. UNICE F Office of Global Insight and Policy. https://www.unicef.org/globalinsight/reports/policy-guidance-ai-children

    2025

  17. [25]

    Randi Williams, Stefania Druga, and Cynthia Breazeal. 2023. Examining Teenagers’ Perceptions of Conversational Agents in Learning Contexts. In IDC 2023. ACM. https://dl.acm.org/doi/10.1145/3501712.3529740 7 IDC2025, 2025, Reykjavík, Iceland Russell Beale

    2023

  18. [26]

    Elinor’s Talking to Me!

    Ying Xu, Valery Vigil, Andres S. Bustamante, and Mark Wa rschauer. 2022. "Elinor’s Talking to Me!": Integrating Con versational AI into Children’s Narrative Science Programming. In CHI 2022. ACM. https://dl.acm.org/doi/10.1145/3491102.3502050

    2022

  19. [27]

    Zhibin Zhou, Yaoqi Li, and Junnan Yu and. 2024. Explorin g the application of LLM-based AI in UX design: an empirical c ase study of Chat- GPT. Human–Computer Interaction 0, 0 (2024), 1–33. https://doi.org/10.1080/07370024.2024.2420991 Publisher: Taylor & Francis _eprint: http...

    2024

  20. [28]

    Tiffany Zhu, Kexun Zhang, and William Yang Wang. 2024. Em bracing AI in Education: Understanding the Surge in Large La nguage Model Use by Secondary Students. https://arxiv.org/abs/2411.18708 _eprint: 2411.18708. 8

    2024 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.