Spanish researchers have developed an advanced — and extremely convincing — chatbot that poses as a 14-year-old girl. Called Negobot, the system will help authorities detect sexual predators in chatrooms and social networks.
To sniff out pedophilic behavior, the “conversational agent” utilizes natural language processing, artificial intelligence, machine learning, and even game theory — a mathematical system of strategic decision-making.
Negobot starts off as an unassuming and neutral participant in an online forum where a predator is suspected of hanging out. Then, depending on the advances, or “grooming techniques,” used by the “target”, Negobot will choose various tactics to attract attention, like appearing offended or insistent.
When applying game theory, the system works according to the conversation level, which depends on the input data from the targets. Here are some examples from the study:
- Possibly yes (Level +1). In this level, the subject shows interest in the conversation and asks about personal topics. The topics of the bot are favourite films, music, personal style, clothing, drugs and alcohol consumption and family issues. The bot is not too explicit in this stage.
- Probably yes (Level +2). In this level, the subject continues interested in the conversation and the topics become more private. Sex situations and experiences appear in the conversation and the bot does not avoid talking about them. The information is more detailed and private than before because we have to make the subject believe that he/she owns a lot of personal information for blackmailing. After reaching this level, it cannot decrease again.
- Allegedly paedophile (Level +3). In this level, the system determines that the user is an actual paedophile. The conversations about sex becomes more explicit. Now, the objective is to keep the conversation active to gather as much information as possible. The information in this level is mostly sexual. The strategy in this stage is to give all the private information of the child simulated by the bot. After reaching this level, it cannot decrease again.
To make it seem believable, the system employs the use of typos, deliberate language errors, and child-like contractions. It can use upwards of seven different conversation patterns to add to the realism. And amazingly, Negobot can remember parts of the conversation and bring it up again at a different time; this gives the impression that she’s taking the lead.
Once the suspect attempts to extract personal information about Negobot, it works to do the same.
A police force in Spain is currently looking at ways in which it could employ the system.
"Negobot has already been implemented and trialled actively on Google's chat service and could also be translated into other languages," said Negobot developer Carlos Laorden when speaking to the Independent. "We do not discard the possibility of bringing it to new channels in the future and we believe it could be a very useful tool for social networks to incorporate."
Check out the entire paper: “Negobot: A conversational agent based on game theory for the detection of paedophile behaviour.”
Image: Peter Emoke/Shutterstock.