AI-made social stories and cartoons: early evidence, risks and safeguards

Background: why AI-made stories, and why caution

A good social story is written for one child and one moment: it describes what will happen, who will be there and what the child might do, in a patient, literal voice (Gray & Garand, 1993). Writing one well takes time and skill, which is why large language models (LLMs, the systems behind chatbots such as ChatGPT, Gemini and Claude) and image generators are attractive: they can draft a story, illustrate it with a likeness of the child, add a voice and even animate it in minutes.

The same speed creates risks. A model can state things that are not true (often called "hallucination") (Ji et al., 2023), drift into a directive or frightening tone, or draw pictures that do not match the words. For a child who is already anxious about a haircut or a new classroom, an inaccurate or unsettling story could do more harm than none. The evidence on AI-made social stories is very new and thin. As of October 2026 we found no randomised trial measuring whether AI-generated social stories or cartoons on their own change children's behaviour or wellbeing. The nearest is a controlled study of EMooly, a tablet game for 24 autistic children and their caregivers that combined generative-AI personalised social stories with augmented-reality activities and caregiver participation; it reported better emotion recognition than a baseline condition, but the stories' own contribution cannot be separated from the rest of the package (Lyu et al., 2024).

How the research has been done

Four kinds of study exist. Benchmark studies from computer science generate large numbers of stories and score them against a rubric, using trained human raters or another LLM as judge. Design and deployment studies from human-computer interaction build a tool and test it for days or weeks with parents or practitioners, mostly measuring usability and self-reported change. Perception experiments ask professionals or parents to rate AI content. A few case reports describe single families. Around these sit two older literatures: reviews of AI tools in autism and special education, and developmental studies of how children trust and learn from machines. None of this is outcome evidence of the kind discussed in our reviews of social stories for autistic children and other children and digital formats.

What the studies found

Quality of AI-written social stories

The largest benchmark, SS-GEN, used GPT-4o to generate over 5,000 social stories and scored them against criteria adapted from Gray's guidance: clear structure, describing more than directing, and "situational safety" (appropriate perspective, positive tone, accurate and exact wording) (Feng et al., 2025). Only 74.3% of generated stories survived filtering and manual screening. On a sample of 100, GPT-4o and GPT-3.5 nearly always used a positive, patient tone (98% and 97%), but only 59% and 56% met the rule to describe more than direct. Small models used without training met the safety criterion in only 6–10% of stories, against 80–83% after fine-tuning on the screened stories. Three trained psychology students rated 50 stories; GPT-4's ratings correlated only modestly with theirs (r = 0.34 to 0.54). No autistic children, parents or teachers took part. The study also used an older version of Gray's ratio (more than two descriptive sentences per coaching sentence), looser than the current 10.3 criteria of at least three (Gray, 2021).

General studies of LLM fiction point the same way. LLM-written children's stories still fell short of the quality and nuance of real ones (Bhandari & Brennan, 2023), and 11,800 stories generated for 236 countries converged on one plot, with conflict sanitised and tension downplayed, a form of narrative bias (Rettberg & Wigers, 2025).

Tools tested with parents and practitioners

In AutiHero, 16 parents of autistic children aged 7 to 12 used a GPT-4o-based app for two weeks to create illustrated, branching social narratives (Lee et al., 2025). They made 218 stories and edited the text of 27% after generation, revising wording that did not sound like a children's book; one parent also regenerated pictures that did not match the characters, and five found the structure repetitive. Parents perceived positive change in 76% of target behaviours and reported more confidence, but these are self-reports from a short, uncontrolled study, and stories were not formally checked against Gray's criteria.

AdaptED Stories, a practitioner tool using Gemini for text and GPT-4o for images made from a student's photo, scored well for usability with seven practitioners using fictional student profiles (Enkhjargal et al., 2026). Participants and authors noted Western-looking visuals, generic facial expressions, inconsistent backgrounds and stories lacking a clear problem and solution. No children took part.

How professionals judge AI stories depends partly on who they think wrote them. In an experiment with 227 education and therapy professionals in Israel, all stories were written by Claude but labelled as either AI or a speech therapist. Labels made no overall difference, but AI-labelled stories were rated better for encouraging positive behaviour and "human" stories were preferred for reducing challenging behaviour (Zaklas & Gazit, 2026).

AI pictures and personalised picture books

Two case reports describe toddlers whose caregivers read a personalised picture book, made with DALL·E 3 to show the real dentist, procedure and child, in the week before a first dental visit. Caregivers' anticipatory anxiety fell (mean 3.5 to 2.5 on a four-point scale), children were rated highly cooperative, and views on the reading burden were mixed (Ishitani et al., 2026). Two cases cannot show an effect.

Image quality is a real concern. Text-to-image models produce demographic stereotypes even from ordinary prompts, and neither counter-prompts nor built-in guardrails prevented them (Bianchi et al., 2023). In a study of 13 families with children aged 4 to 8, most parents accepted AI pictures if the text was human-written and an expert reviewed the images; children noticed when a picture's emotion clashed with the text more often than parents did (Jin & Yuan, 2025). Children older than about nine find very human-like robots creepier than machine-like ones, while younger children do not (Brink et al., 2019), so near-realistic faces may unsettle older children. AI-generated children's videos, mostly music and stories, are becoming common on YouTube, alongside tutorials on how to monetise them (Goes Mintz & Bozzi, 2026).

AI tools in autism and special education

Reviews are early and cautious. A scoping review of generative AI in autism care found only 10 empirical studies and listed small samples, data bias, limited validation and hallucinations among the problems (Sohn et al., 2025). A systematic review of AI-supported autism interventions (mainly robots, virtual reality and wearables) found 14 studies, mostly with fewer than 20, mainly male, participants (Kuca et al., 2026). Of 11 studies of AI tools for pupils with learning disabilities, all reported positive results but none was at low risk of bias (Paglialunga & Melogno, 2025).

When parents ask chatbots about autism, answers are often accurate but imperfect: ChatGPT-4 was largely correct and clear but gave little actionable advice and inaccurate references and links (McFayden et al., 2024); three LLMs made errors clustered on local, genetic and financial questions and missed readability benchmarks (Akhund et al., 2026). A purpose-built agent's answers to 400 parent questions were rated safe, empathetic and useful by experienced evaluators (Brown et al., 2026), a reminder that quality depends on design and testing.

Children's trust in and learning from AI

Children learn from machines, and they often trust them. With age, children aged 4 to 8 increasingly trusted a voice assistant for facts and a person for personal information (Girouard-Hallam & Danovitch, 2022). In an RCT with 117 children aged 3 to 6, a conversational agent asking questions during reading improved story comprehension as much as a human partner (Xu et al., 2022). Children aged 4 to 6 preferred a voice assistant over a peer or a teacher for science explanations (Haber et al., 2026). Older children become sceptical after a robot makes errors (Flanagan et al., 2024), but autistic children aged 4 to 7 showed less selective trust between accurate and inaccurate informants than their peers (Chen et al., 2026). Primary-school children tend to overestimate smart speakers' intelligence and misunderstand their privacy (Andries & Robertson, 2023). Errors in AI content may therefore be absorbed, not questioned.

Risks and when they are greatest

• Inaccuracy. Models can invent steps, people or facts (Ji et al., 2023). In a story meant to make a real visit predictable, a wrong detail defeats the purpose.
• Tone and content. Unscreened output often breaks the rules that keep stories calm and descriptive (Feng et al., 2025), and LLMs can respond poorly to children's disclosures, an "empathy gap" (Kurian, 2025).
• Bias and sameness. Images and plots default to narrow norms (Bianchi et al., 2023); (Rettberg & Wigers, 2025), a problem when a child should recognise their own family and culture.
• Mismatched or uncanny visuals. Children notice emotional mismatches (Jin & Yuan, 2025), and older children may find near-human faces creepy (Brink et al., 2019).
• Privacy. Personalisation needs names, photos and details of a child's difficulties. Human Rights Watch found 170 photos of Brazilian children, some with names, in a dataset used to train image tools, after checking a tiny fraction of it (Human Rights Watch, 2024).
• Over-reliance. Because children trust machines (Haber et al., 2026), AI stories should support, not replace, a trusted adult.

Risk is greatest when output goes to a child unchecked, when stories touch medical or safety facts, for children who trust informants less selectively (Chen et al., 2026), and when identifiable photos are uploaded to services that keep them.

What this means in practice

What the rules say

UNICEF's updated guidance sets ten requirements for child-centred AI, including safety, protection of children's data and privacy, non-discrimination, and transparency, explainability and accountability (UNICEF Innocenti, 2025). In the UK, the ICO's Age appropriate design code applies to online services likely to be accessed by under-18s and sets 15 standards, including the child's best interests, transparency, default settings and data minimisation (Information Commissioner's Office, 2020). In the US, the COPPA Rule protects children under 13; its definition of personal information includes photos, videos and audio files containing a child's image or voice, and 2025 amendments require separate parental consent before disclosing children's data to third parties and limit how long data may be kept (Federal Trade Commission, 2025). In the EU, Article 50 of the AI Act requires providers of generative systems to mark outputs as AI-generated in machine-readable form and deployers to disclose deep fakes, with a lighter duty for evidently artistic or fictional works (Regulation (EU) 2024/1689); the Commission states these duties apply from 2 August 2026 (European Commission, n.d.). None of this is legal advice.

A safeguard checklist

• A person is the author. Treat AI output as a draft; parents in the largest home study edited over a quarter of stories (Lee et al., 2025).
• Check it against Gray's criteria. Mostly descriptive sentences, at most one sentence that coaches the child, a first- or third-person voice and literally accurate words (Gray, 2021).
• Check every fact. Confirm steps with someone who knows the real setting, such as the dentist or class teacher (Ji et al., 2023).
• Nothing frightening. Look at each picture against its sentence for emotion, safety and accuracy (Jin & Yuan, 2025).
• Check representation. The child, family and setting should look like their own (Bianchi et al., 2023).
• Share as little as possible. Avoid surnames and school names; upload photos only where storage and deletion are clear (Information Commissioner's Office, 2020); (Human Rights Watch, 2024).
• Say that AI was used. Parents preferred one clear notice to labels on every picture (Jin & Yuan, 2025).
• Watch together. An adult should introduce the story, talk about it and notice the child's reaction (Kurian, 2025).

What this means for short personalised cartoons

Cala Luna, which publishes this site, uses AI tools for drawing, voices and animation under human direction, and a person checks every cartoon. This workflow, and our personalised cartoons, have not been evaluated in any study.

What plausibly transfers is the case for human review: most of the story and image studies above found errors or mismatches that a careful adult would catch. What does not transfer is any evidence of benefit; the research on video modelling and animated versus filmed models tested human-made materials. Personalised likenesses also raise privacy questions that a plain-drawn story does not, discussed further in our review of personalisation.

Limits of the evidence

• No randomised trial has measured children's outcomes after AI-made social stories or cartoons on their own; the one controlled study tested a multi-component game (Lyu et al., 2024).
• Most studies are preprints or conference papers, often by tool developers, with small samples and short follow-up (Lee et al., 2025); (Enkhjargal et al., 2026).
• Story quality is judged by students, LLMs or parents, rarely by autistic people or experienced story writers (Feng et al., 2025).
• Models change every few months, so findings about one version may not hold for the next.
• A review of AI tools for learning disabilities found no study at low risk of bias (Paglialunga & Melogno, 2025).

Open questions for research

1. In a randomised trial, do AI-drafted, human-edited social stories achieve the same goal attainment as stories written entirely by trained staff, at lower cost?
2. How often do current models produce factual errors about common situations (dental visits, haircuts, school routines), and does a structured checklist catch them?
3. Do children respond differently to a cartoon version of themselves than to a generic character, and does likeness raise or lower anxiety?
4. At what age, and for which children, do AI-generated faces become unsettling?
5. Does an AI disclosure change parents' trust or children's engagement?
6. Can useful personalisation be achieved without uploading photos, for example from a description or a parent-chosen avatar?
7. Do autistic children's patterns of trust in technology make them more likely to absorb errors in AI stories?

Frequently asked questions

References

1. Gray, C. A., & Garand, J. D. (1993). Social stories: Improving responses of students with autism with accurate social information. Focus on Autistic Behavior, 8(1), 1–10. doi.org/10.1177/108835769300800101
2. Gray, C. (2021). Social Stories 10.3 criteria (with S. Timmins & C. Faherty). Carol Gray Social Stories. carolgraysocialstories.com/wp-content/uploads/2021/10/Social-Stories-10.3-Criteria.pdf
3. Ji, Z., Lee, N., Frieske, R., et al. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38. doi.org/10.1145/3571730
4. Feng, Y., Song, M., Wang, J., et al. (2025). SS-GEN: A social story generation framework with large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 39(2), 1300–1308. doi.org/10.1609/aaai.v39i2.32119
5. Bhandari, P., & Brennan, H. (2023). Trustworthiness of children stories generated by large language models. In Proceedings of the 16th International Natural Language Generation Conference (pp. 352–361). doi.org/10.18653/v1/2023.inlg-main.24
6. Rettberg, J. W., & Wigers, H. (2025). AI-generated stories favour stability over change: Homogeneity and cultural stereotyping in narratives generated by gpt-4o-mini [version 1]. Open Research Europe. doi.org/10.12688/openreseurope.20576.1
7. Lee, J., Lee, K., Hwang, I., Park, S., & Kim, Y.-H. (2025). AutiHero: Engaging parents in creating personalized, multi-path social narratives for autistic children (arXiv:2509.17608) [Preprint]. arXiv. arxiv.org/abs/2509.17608
8. Enkhjargal, B., Lalwani, H., & Salam, H. (2026). AI-assisted social story intervention for special education: The design of AdaptED Stories (arXiv:2609.24245) [Preprint]. arXiv. arxiv.org/abs/2609.24245
9. Zaklas, N., & Gazit, T. (2026). The experience of using AI-generated social stories as a behavioral intervention tool for children. Computers & Education, 254, 105723. doi.org/10.1016/j.compedu.2026.105723
10. Ishitani, H., Tanaka, Y., & Ono, Y. (2026). Effect of personalized AI-generated picture books on anxiety during first dental visits in young children: Two case reports. Case Reports in Dentistry, 2026(1). doi.org/10.1155/crid/9947164
11. Bianchi, F., Kalluri, P., Durmus, E., et al. (2023). Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 1493–1504). doi.org/10.1145/3593013.3594095
12. Jin, Q., & Yuan, Y. (2025). "They all look mad with each other": Understanding the needs and preferences of children and parents in AI-generated images for stories. International Journal of Child-Computer Interaction, 46, 100787. doi.org/10.1016/j.ijcci.2025.100787
13. Brink, K. A., Gray, K., & Wellman, H. M. (2019). Creepiness creeps in: Uncanny valley feelings are acquired in childhood. Child Development, 90(4), 1202–1214. doi.org/10.1111/cdev.12999 · PubMed
14. Goes Mintz, A., & Bozzi, J. (2026). Slop for kids: An exploratory study of AI-generated videos for children on YouTube. AoIR Selected Papers of Internet Research. doi.org/10.5210/spir.v2024i0.15158
15. Sohn, J.-S., Lee, E., Kim, J.-J., Oh, H.-K., & Kim, E. (2025). Implementation of generative AI for the assessment and treatment of autism spectrum disorders: A scoping review. Frontiers in Psychiatry, 16, 1628216. doi.org/10.3389/fpsyt.2025.1628216 · PubMed
16. Kuca, J., Stencel, M., Pilarski, B., Florek, S., & Pudlo, R. (2026). Artificial intelligence-supported therapeutic interventions for autism spectrum disorder: A systematic review. Frontiers in Psychiatry, 17, 1833853. doi.org/10.3389/fpsyt.2026.1833853 · PubMed
17. Paglialunga, A., & Melogno, S. (2025). The effectiveness of artificial intelligence-based interventions for students with learning disabilities: A systematic review. Brain Sciences, 15(8), 806. doi.org/10.3390/brainsci15080806 · PubMed
18. McFayden, T. C., Bristol, S., Putnam, O., & Harrop, C. (2024). ChatGPT: Artificial intelligence as a potential tool for parents seeking information about autism. Cyberpsychology, Behavior, and Social Networking, 27(2), 135–148. doi.org/10.1089/cyber.2023.0202 · PubMed
19. Akhund, S. A., Alsaleh, A., Kazi, N. H., Rajpal, B., Naz, S., & Arain, S. A. (2026). LLMs in medical education for autism caregivers: A comparative evaluation of accuracy, readability, actionability, and neurodiversity-affirming language. Healthcare, 14(14), 2137. doi.org/10.3390/healthcare14142137 · PubMed
20. Brown, F. J., Muscat, I. S., Quinn, L., Best, L., & Cooper, P. (2026). Validation of the safety, empathy and utility of a large language model conversation agent for parents of autistic and neurodivergent children. Scientific Reports, 16(1). doi.org/10.1038/s41598-026-44254-5 · PubMed
21. Girouard-Hallam, L. N., & Danovitch, J. H. (2022). Children's trust in and learning from voice assistants. Developmental Psychology, 58(4), 646–661. doi.org/10.1037/dev0001318 · PubMed
22. Xu, Y., Aubele, J., Vigil, V., Bustamante, A. S., Kim, Y.-S., & Warschauer, M. (2022). Dialogue with a conversational agent promotes children's story comprehension via enhancing engagement. Child Development, 93(2), e149–e167. doi.org/10.1111/cdev.13708 · PubMed
23. Haber, A. S., Kumar, S. C., Swenson, M., Bode, K., & Ruel, E. (2026). How do children evaluate scientific explanations provided by digital voice assistants, teachers, and peers? Behavioral Sciences, 16(5), 661. doi.org/10.3390/bs16050661 · PubMed
24. Flanagan, T., Georgiou, N. C., Scassellati, B., & Kushnir, T. (2024). School-age children are more skeptical of inaccurate robots than adults. Cognition, 249, 105814. doi.org/10.1016/j.cognition.2024.105814 · PubMed
25. Chen, Z., Zheng, J., Gao, Y., Fang, J., Wang, Y., Chen, H., & Wang, T. (2026). Do children with autism spectrum disorders show selective trust in social robots? Journal of Autism and Developmental Disorders, 56(10), 4175–4182. doi.org/10.1007/s10803-024-06474-4 · PubMed
26. Andries, V., & Robertson, J. (2023). Alexa doesn't have that many feelings: Children's understanding of AI through interactions with smart speakers in their homes. Computers and Education: Artificial Intelligence, 5, 100176. doi.org/10.1016/j.caeai.2023.100176
27. Kurian, N. (2025). 'No, Alexa, no!': Designing child-safe AI and protecting children from the risks of the 'empathy gap' in large language models. Learning, Media and Technology, 50(4), 621–634. doi.org/10.1080/17439884.2024.2367052
28. Human Rights Watch. (2024, June 10). Brazil: Children's personal photos misused to power AI tools. www.hrw.org/news/2024/06/10/brazil-childrens-personal-photos-misused-power-ai-tools
29. UNICEF Innocenti. (2025). Guidance on AI and children (Version 3.0): Recommendations for AI policies and systems that uphold child rights. UNICEF. www.unicef.org/innocenti/reports/policy-guidance-ai-children
30. Information Commissioner's Office. (2020). Age appropriate design: A code of practice for online services. ICO. ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/childrens-information/childrens-code-guidance-and-resources/age-appropriate-design-a-code-of-practice-for-online-services/
31. Federal Trade Commission. (2025, April 22). Children's Online Privacy Protection Rule [Final rule amending 16 CFR Part 312]. Federal Register. www.federalregister.gov/documents/2025/04/22/2025-05904/childrens-online-privacy-protection-rule
32. European Parliament & Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
33. European Commission. (n.d.). Transparency obligations under Article 50 of the AI Act [Frequently asked questions]. Retrieved 24 October 2026. digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
34. Lyu, Y., Liu, D., An, P., Tong, X., Zhang, H., Katsuragawa, K., & Zhao, J. (2024). EMooly: Supporting autistic children in collaborative social-emotional learning with caregiver participation through interactive AI-infused and AR activities. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 8(4), 1–36. doi.org/10.1145/3699738

Spotted an error, or know a study we should add? Email hello@calalunakids.com. Every study we draw on is in the research library; research terms are explained in our methods glossary.