Concept Note · Kids Play Tech Lab
Generative Creepiness
TLDR: Generative creepiness is the tendency of GenAI tools to veer toward content that is creepy, inappropriate or otherwise unsettling for children.
generative creepiness (n.) — the tendency of generative AI tools to veer toward content, themes and/or imagery that is unsettling, age inappropriate, or otherwise seen as “creepy” by child users: diverging from what was asked for or intended by the user in ways that transgress developmental boundaries. Examples include the introduction of an uncanny character, a story that suddenly and unexpectedly veers into horror/frightening themes, the addition of a sexualized figure, or a taboo subject surfacing where it does not belong. The tendency is a property of how certain GenAI systems generate content — a product of their training data, inference errors, and the probabilistic nature of generation itself — not a clear or anticipated result of the user’s prompt(s) or intentions. It may surface in a first response or emerge as the interaction unfolds.
Why “creepy”?
“Creepy” is a term commonly used by children to describe unsettling, uncanny, unexpected and ambiguously threatening content and interactions, online, in the media, and in their everyday lives. Francis McAndrew and Sara Koehnke describe creepiness as a response to ambiguity about the presence of threat — the low-grade vigilance we feel when something might be wrong, but we cannot yet say what or even say for certain the feeling is warranted. Its tentative and unresolved nature distinguishes it from fear. For example, some of the children in the CAGD study used the term “creepy” to describe interactions, players or games that raised red flags or seemed iffy to them, but that were not decidedly dangerous or “actually” scary. Here too, ambiguity seemed to be a core component of “creepiness.”
Our finding tracks with previous research on children’s use and perceptions of digital technologies. Jason Yip and colleagues found that kids aged 7 to 11 used “creepy” to describe technologies that posed an unpredictable or ambiguous threat — one that they were concerned might cause physical harm or could potentially disrupt an important relationship.
Two sources of AI creepiness
In privacy scholarship, “creepy” already has a home. Omer Tene and Jules Polonetsky’s “theory of creepy,” and much of the work after it, use the word for the discomfort of being watched as we encounter targeted ads, data collection, and the sense of being followed across the digital (and sometimes nondigital) environment. This surveillance creepiness concerns how AI systems watch and gather data.
With the term generative creepiness, we are attempting to instead describe an emerging finding relating to what GenAI tools and systems make, i.e. generate and introduce, as children interact, converse and play with them.
Four registers
While still in the very preliminary stages of study, to date we have observed generative creepiness surfacing across at least four registers. While each of these registers can likely also be found across general-audience interactions with GenAI, our focus here is on expanding what features and impacts are considered when a child-centred lens is applied:
- Uncanny form — output that is near-human but off: the too-wide smile, the wrong number of fingers, the not-quite-right voice. This is the classic uncanny valley.
- Thematic veer — the system responds to a benign prompt by steering into frightening or violent or other mature themes. E.g., a story about two friends lost in the woods veers unprompted into a horror story.
- Sexualization / objectification — characters or scenarios rendered in sexualized, racialized, or otherwise biased ways that the child did not seek and cannot contextualize.
- Taboo intrusion — mature or distressing subject matter surfacing where a child did not invite it.
What it is not
Not merely “age-inappropriate content.” That regulatory category names the material; generative creepiness connects to a child’s experience of it and role (however inadvertent) in generating it, which might in turn amplify feelings of distress, guilt, inadequacy or complicity (e.g., if blame is internalized rather than delegated to the system and its programming).
Why it matters
Naming generative creepiness gives researchers and designers a specific “failure mode” to test for as they consider whether to introduce GenAI tools to a game or play space and (accordingly) develop guardrails and policies to ensure age-appropriate usage. It also gives regulators a new type of content risk to address that is anchored in child-centred research and terminology.
References
- McAndrew, F. T., & Koehnke, S. S. (2016). On the nature of creepiness. New Ideas in Psychology, 43, 10–15.
- Tene, O., & Polonetsky, J. (2013). A theory of creepy: Technology, privacy and shifting social norms. Yale Journal of Law & Technology, 16(1), 59–102.
- Yip, J. C., Sobel, K., Gao, X., Hishikawa, A. M., Lim, A., Meng, L., Ofiana, R. F., Park, J., & Hiniker, A. (2019). Laughing is scary, but farting is cute: A conceptual model of children’s perspectives of creepy technologies. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI ’19). ACM. https://doi.org/10.1145/3290605.3300303
Suggested citation: Grimes, S. M. (2026). Generative creepiness [Concept note]. Kids Play Tech Lab, McGill University.
