Good morning, everyone.
My name is Lloyd Richardson. I'm the director of technology at the Canadian Centre for Child Protection. We operate Cybertip.ca, which is Canada's national tip line for reporting the online sexual exploitation of children, as well as Project Arachnid, which is a global tool designed to curb the distribution of child sexual abuse material on the Internet.
Our organization has witnessed first-hand the impact AI is having on children, particularly girls, across our country.
The development of AI has definitely delivered some societal benefits, but it has also exposed some significant safety, regulatory and ethical gaps that disproportionately affect women and girls. I think everyone here is aware of the existence of deepfake apps and websites that market the ability to nudify—creating non-consensual intimate images of people—in exchange for money or tokens. These websites offer nudification services advertised with gendered language in the interface for their customers. For example, one particular website has an action button in the interface labelled “reveal her”. This leaves very little doubt as to who this targets and what the intent is.
Additionally, all of the sample imagery on these websites features women as models for the deepfake features. While the language on these sites speaks of women, we know it's not just adult women. It's also adolescent and younger girls, based on the images we have reported to us by victims and what we encounter online.
While interventions such as removal notices can be made with websites and apps that offer deepfake or nudification services, a more challenging issue is the open-weight AI models that we see that can be freely downloaded from the Internet. On their own, some of these AI models have safeguards to prevent the creation of sexualized material. However, they can be paired with low-rank adaptations, often referred to in short as LoRAs. These LoRAs are essentially plug-ins for AI models that add new features or allow the model to create content related to a specific individual, an object or an artistic style.
Within communities of individuals with a sexual interest in children, we see requests made for the creation of these LoRAs for specific victims or survivors of child sexual abuse. This horrific use of the technology allows offenders to generate new abuse imagery depicting previously identified and rescued victims. The same technology allows someone to create child sexual abuse material of any child they're able to download, capture a picture of or take a picture of in public.
We've catalogued and issued removal notices on 100 of these unique LoRAs, and I can tell you that the overwhelming majority of the individuals represented in them are prepubescent girls.
The harms associated with these services and offender communities are reflected in what we see reported to us through the tip line or what we've detected through the tip line. Back in 2023, we assessed about 4,000 deepfake images. In 2024, it was 9,000. In 2025, we went up to 61,000. This year, we're on track for about 100,000 that we've tagged. This only represents the imagery we're able to label, and as you can see, the problem is only getting worse.
Lastly, we also see harm in the datasets used to train some of these large models. Back in 2023, we assisted Stanford research in identifying over 1,000 known child sexual abuse images in the public LAION-5B dataset. This dataset has been used to train popular models for most image generation tools out there as it's one of the largest. Once again, the overwhelming majority of these victims were girls.
In conclusion, whether it's a commercial deepfake service, freely available AI models or the datasets used to train them, we're seeing a consistent lack of guardrails in the development and deployment of generative AI. Across each of these examples, the impact falls disproportionately on women and girls.
Thank you.
