Toxicity and hallucination detectors for AI agents were developed by VTB together with Moscow University - the first detector will intercept requests with prohibited content, the second will check the responses of the neural network for errors, fictitious facts and contradictions. If one of the detectors finds a problem, the response can be sent for review or blocked. For AI in financial services, this is especially important so that the neural network does not recommend to the customer something that does not actually exist. The bank plans to implement the development in retail services with digital assistants. Now we need to come up with a toxicity detector for colleagues.

Toxicity and hallucination detectors for AI agents were developed by VTB together with Moscow University - the first detector will intercept requests with prohibited content, the second will check the responses of the neural network for errors, fictitious facts and contradictions. If one of the detectors finds a problem, the response can be sent for review or blocked. For AI in financial services, this is especially important so that the neural network does not recommend to the customer something that does not actually exist. The bank plans to implement the development in retail services with digital assistants. Now we need to come up with a toxicity detector for colleagues.