Help Frontline Responders Find More Kids, and Faster
Summary
In 2024, Thorn committed to build the next generation of machine learning/artificial intelligence models by the end of 2027 to classify child sexual abuse and exploitation material at a more granular level, enabling frontline responders to more quickly locate victims and remove them from harm. With greater prioritization and triage, this next generation of Child Sexual Abuse Material classifiers will enable stakeholders across jurisdictions to collaborate more effectively, hold perpetrators accountable, and make removal of abuse content across the open web more efficient. Working with data from National Center for Missing and Exploited Children and other partners, this model will be developed, tested, and deployed, ultimately equipping 1,150 victim identification specialists and content moderators globally, as well as law enforcement entities and technology platforms worldwide, with next-generation tools to combat child exploitation.
Approach
Thorn, as one the largest teams dedicated to solving online child sexual abuse with technology, commits to building the next generation of ML/AI models to classify CSAM/CSE at a more granular level by 2027 and to equip the current 1150 VID specialists, from 634 agencies located in 34 countries, (along with new users) with the technology. With more efficient technology, frontline responders can find more kids faster, with less risk to their mental health and resiliency. With greater prioritization and triage, Thorn’s next generation of classifiers has the potential to surface crucial information to support investigators in locating the child and enable stakeholders across jurisdictions to collaborate more effectively. Ultimately, this will significantly reduce the time it takes to find a victim and remove them from harm.
To implement this project, Thorn will evaluate, iterate (via alpha and beta testing) and deploy an initial set of classifiers they have already built to a production setting. At the same time, they will begin or continue to build a secondary set of classifiers, which will also be evaluated, iterated, and deployed. The first set of classifiers includes Photorealism (e.g. on a scale from fully cartoon/animated to indistinguishable from a picture taken by a camera) ; Text of Interest (e.g. location, access to children) ; Sexual Act Occurring; Full Image Age Estimation (e.g. infant/toddler vs. pre-pubescent vs. post-pubescent) ; Nudity (e.g. the level of nudity/genitalia displayed) ; Self-Generated (images taken by children of themselves) ; Similarity Matching (across multiple images/scenes, connect those that depict the same child across various abuse settings) .
To date, their data science team has collectively developed, deployed, and maintained over 20+ production models for CSAM/CSE detection and VID. With a critical need to move classifiers beyond “Is this CSAM ” to “What else is happening in this image/video ” Thorn is building the next generation of classifiers that will grow in value and impact. Providing the world’s limited victim identification resources with intelligence that makes their investigations more efficient and effective means more children are removed from harm, and more perpetrators are prosecuted.
Action Plan
Thorn’s classifiers will be trained on data from NCMEC and other trusted VID partners. They anticipate developing new types of classifiers as the project and landscape progress. Currently, they are planning to build the following next-generation classifiers: Doppelgangers, CSA Multilingual Text, Gen AI Prompt, and Location Detection (image-based) . The action plan splits the work into three phases: alpha testing (proving the models are technically feasible) , beta testing (iterating with select partners) , and deployment (general availability to VID specialists and content moderators) .
In 2025, Thorn will continue to deploy its CSAM classifier at scale in major law enforcement agencies and forensic software providers. Scaling the core CSAM classifier is essential to enable the more refined classifiers to be efficient and effective. They will also complete alpha testing stages for 50% of the next suite of classifiers, begin beta testing, and complete product design for the next generation of classifiers.
In 2026, they will complete alpha stages for the other 50% of the next suite of classifiers and 50% of betas. They will then transition to general availability for law enforcement agencies, forensic providers, and industry offerings.
In 2027, they will complete the other 50% of betas for the rest of the classifiers and transition to general availability (for those classifiers where the alpha phase proved technically feasible) for law enforcement agencies, forensic providers, and industry offerings.
Background
Child sexual abuse material (CSAM) and exploitation (CSE) are a global crisis that is escalating at lightning speed and unprecedented scale. In 2023, more than 104 million images and videos of child sexual abuse were reported to the National Center for Missing and Exploited Children (NCMEC) — up from 20 million in 2017. The full scale of online child sexual abuse and imagery is hard to quantify, but discovered material is revealing horrifying trends toward more violent and extreme, with younger victims – some preverbal. Investigators report that these images depict some of the worst crimes humanity has ever seen. Emerging large-scale threats (including online grooming, sextortion, and generative AI) mean that any child with internet access is at risk – regardless of their physical proximity to an abuser.
New CSAM is produced and uploaded to online platforms every day, but many online platforms don’t proactively detect it, while others can only find existing CSAM. This means that content may not be detected by law enforcement until months or years after being recorded, leaving the child at risk for continued abuse and revictimization via widespread sharing across the web. Finding new and unknown CSAM often relies on manual processes, with frontline responders sifting through millions of files documenting abuse at great cost to their own mental health, including vicarious trauma through exposure. Not only is the process arduous and time-consuming, but the welfare of the front-line responders is challenged both by exposure to graphic images and by the overwhelming amount of data through which they must sift just to find the data that will most help them identify and find victims and perpetrators. Human intervention in this crisis is not enough – the field of responders is small, with high turnover and inadequate equipment to do the job.
Classifiers and other AI technologies can help rapidly triage and prioritize content. By pairing automation and humans-in-the-loop, the next generation of CSAM classifiers can achieve what it would take hundreds of people with limitless hours to do – reducing content exposure and supporting the wellness and resiliency of human investigators so that they can find more kids faster. Currently, the ecosystem is in the early phases of adopting machine learning/artificial intelligence (ML/AI) technologies to address CSAM. Victim Identification (VID) specialists who have deployed it view AI capabilities as indispensable and are the same ones requesting this next generation of classifiers to streamline their work, resulting in greater efficiencies.
Progress Update
Partnership Opportunities
Thorn is seeking to raise $8.1 million to implement this work of which $2.5 million is already committed., As a leader in child safety, Thorn provides best practices information to help the world better understand the issue and equip those on the frontlines with solutions to defend children from sexual abuse and exploitation in the digital age. Every platform with an upload button or messaging capabilities is at risk of hosting CSAM or interactions that could lead to CSE. Thorn is committed to equipping platforms with tools and expert guidance to mitigate these risks.