Finding the Bad Guys with AI: Connecting the Dots In Data

Christian Berg and Mats Kvarnström of Paliscope, a company of about 40 people that builds software for investigators in law enforcement and government agencies, showed how AI helps turn large, mixed data sets into evidence. Their software extracts text and metadata from documents, images, video, and audio, with OCR, speech-to-text, object, face, and licence-plate recognition, translation, and NLP, and links the results, so that investigators can search a video for what is said in it, or see who appears together with whom. Because the data must stay on a single computer with no cloud connection, they use small models rather than large language models.

Kvarnström showed chat clustering: sentence embeddings trained on one forum of 3,000 users, which in a real grooming case grouped about ten handles used by the same offender, set the victims apart, and filtered out spam bots. CLIP embeddings let investigators sort hundreds of thousands of images and search them with a sentence. Berg described open-source monitoring of the Wagner Group on Telegram, identifying types of tank from images, and finding where photos were taken from skylines and backgrounds. The insight still comes from the investigator, and the results have to be explained in court.

In the questions, they discussed bias in policing and the need for explainable methods.