Group 15 · Research pitch

Deepfake spreading and prevention among minors

How do explicit non-consensual deepfake images spread among adolescents, and how could a school-based program reduce it?

Background

What an AI deepfake actually is

  • An explicit image of a real person who never posed
  • Made with diffusion models, the AI behind image generators
  • Free apps + a few public photos: anyone can be targeted
  • Realistic, and you often can't tell what's real

noise → image

Background

Philosophical implications

  • Epistemic uncertainty (Fallis 2020)
  • Disinformation and trust in the political system (Fallis 2020, Cavedon-Taylor 2025)
  • Normative (De Ruiter 2021):
    • Permissible uses
    • Consent
    • Identity
  • Criminality (Petrini & Belén Gauna 2024)

Background

Statistics

  • Children are more likely to use generative AI
  • Most parents are oblivious to children's use
  • 40% of US students and 29% of teachers know someone depicted recently
  • Social collusion and bullying lead to anxiety, depression and PTSD
  • Deepfake CSAM as harmful as real CSAM for victims' mental health

deepfakes online

Background

Psychological factors

Adolescents' involvement influenced by:

  • Peer pressure
  • Desire for social status
  • Social norms
  • Moral disengagement theory
  • Bystander effect

Chasapis et al., 2025; Mandau, 2020; Bussey et al., 2015; Meter & Bauman, 2018; Barlińska et al., 2013

→ perceive the forwarding of intimate images as a joke, gossip, harmless entertainment, rather than IMAGE-BASED SEXUAL ABUSE

Sciacca et al., 2023

Objective

What we aim to achieve

  • Giving insight into how deepfakes spread amongst minors
  • Educating minors on deepfakes to make them less prone to harm
  • Identifying the most effective strategies for reducing the spread of technology-facilitated sexual violence

Scientific relevance & novelty

Image Based Sexual Abuse

  • 550% annual growth in explicit deepfake content since 2019
  • 98% of online deepfakes are pornographic
  • IBSA consequences: anxiety, depression, shame, social exclusion, reputational damage

School-based prevention programs reduce non-consensual sexting behaviours, immediately after implementation and several months later.

The problem

Why we model the spread, not detect fakes

  • A detector is out of date the day you ship it
  • Training one needs the very images we want to stop
  • The harm is the spread, not the file existing
  • Can't test fixes on real kids, so we simulate it

The model

A simulated school network

~45s - AI-made explicit image of a real person who never posed - Same tech family as the text-to-image art tools - Access changed: free app, seconds, a few public photos - Realistic, unlimited, can't tell what's real anymore - HANDOVER: that doubt -> [Yord], philosophy

Context = >20 school-related cases of AI-generated sexualized images of ado girls since 2023 Causes: - peer pressure - social norms - desire for social status - moral disengagement = enables individuals to justify harmful behavior by minimizing its consequences -> diffusing responsibility - bystander behavior -> do not report, share Boys: status enhancement Girls: victim-blaming Ado: do not perceive sexual violence when are not victims + > likely to engage in harmful behavior when rewarded Harmful content = individual decisions + collective beliefs Research has focused on techno detections and legal implications, but not on psychological drivers of spreading

~30s - Detector = obvious move, two problems: outdated on ship day; training needs the very illegal images - Real damage = how fast it rips through a school, not the file - Can't test on real kids -> simulate - TRANSITION: here's what that simulation is

~50s, slow down: this is the slide the room remembers - No program: most of the year reached - Random 30%: barely moves, 64% - Same 30%, most-connected: collapses, under 10% - Same effort, completely different reach - Search on top (who to educate, given a budget) = genuine AI problem, part of the proposal - Honesty line: numbers illustrative; calibration + search = what the grant funds