Facebook negotiation chatbots create secret language

Two AI-powered chatbots which were de signed to learn how to negotiate developed a form of communication that appeared to be gibberish to human observers. Facebook's Artificial Intelligence Research (FAIR) group developed two chatbots, name d Alice and Bob, that were intended to be capable of negotiation and used techniques like self-play and dialogue rollouts to improve their performance. The chatbots created an efficient shorthand language to communicate with each other that appeared to be gibberish but which in fact was a result of not being incentivi s ed to stick to English-language rules. The behaviour is a well-established phenomenon where AI agents develop shorthands or "codewords" to maximise efficiency. The episode was seen to highlight the complexities of developing AI systems able to understand and use natural language effectively, as well as the challenges in creating AI that can negotiate and reason strategically. It was also seen to demonstrate the need for careful consideration of communication protocols and standards as AI systems become more advanced and potentially develop their own efficient ways of interacting. System 🤖 End-to-end negotiator system code (GitHub) 🔗 Documents 📃 System research paper Facebook. Deal or no deal? Training AI bots to negotiate Operator: Developer: Meta/Facebook Country: Global Sector: Business/professional services Purpose: Simulate negotiations Technology: Machine learni ng Issue: Incentivisation

Date it happened

2017-01-01

Product, system or model

End-to-end negotiator

This is a record of what was reported, not a finding that anyone broke the law. If it names your organisation and you believe it is wrong, the corrections process is free and open to everyone.