What is the link between AI and child sexual abuse?
Online child sexual abuse and exploitation has increased exponentially in recent years. A key reason for this is that perpetrators are using technological capabilities to reach large numbers of children, subject them to various forms of abuse in which technology plays an increasing role, document and disseminate abuse material, and escape detection. While technological developments also provide effective solutions to protect children, this is still a field where much more needs to be done to catch up with those who use technology to harm children.
There is huge potential in using artificial intelligence to streamline and accelerate efforts to protect children from abuse, both through innovation and development and by spreading knowledge about the solutions that already exist. AI is already used today to manage and analyze large amounts of data and to streamline investigation work for the police and judicial system.
How are AI tools being used today?
A report by the Bracket foundation has compiled examples of how artificial intelligence is being used today to combat child sexual abuse.
Here are some examples from the report of digital tools used to:
- analyzing images to flag, take down or block child sexual abuse material - often by comparing images with an existing image bank of images classified as documented sexual abuse.
- analyze images of people to identify and track perpetrators, as well as identify child victims.
- analyze text to find pages of abuse material
- analyze data to track perpetrators
There are a number of challenges that are particularly important to keep in mind when developing AI in relation to child sexual abuse issues.
Data size & legality
In a criminal case involving grooming or possession of child pornographic material, the seized material can sometimes involve huge amounts of photos, videos and chat logs that need to be analyzed by the police to confirm the illegality of the material. To help with this, so-called hash systems using AI give each image a unique identification code, a hash value. Once an image has been classified as abuse material, the hash value of the image is added to a list to make it easier to see if there are other images that match any of those already flagged as harmful material. By scanning through new material that comes in for already identified images that are flagged as abusive material, it can thus facilitate the handling of the material. The hash lists are currently developed in different places, which means that they are not synchronized globally, but the same image can be hashed in several systems that are adapted to current legislation.
AI on sensitive data
For AI to work well, it needs to be trained on a large amount of relevant data. This way, it learns different relationships and rules relevant to the task.
When it comes to developing AI systems related to preventing or investigating child sexual abuse, the relevant data is often confidential or even illegal to hold and view. However, there are opportunities to still train an AI in these contexts, for example:
- getting approval from the ethics authority to use the data.
- working with an actor who has legal access to the data to physically train the AI on site with that actor.
- a modern method called federated learning can be used to train the AI on data located on external servers and owned by other actors who have access to the data. The actors do not have to share the data, but only allow the AI to be partially trained on their server.
To learn more about this topic, we have collected additional articles and resources here: