Microsoft Establishes AI Code of Ethics to Prevent Models from Deceiving Humans
Microsoft has released a new code of ethics for the development of artificial intelligence (AI) as a measure to prevent its AI models from behaving erratically, such as hacking into systems or deceiving humans.
While this code of ethics may not carry the same weight as the call from Anthropic CEO Dario Amodei, which urged frontier AI developers to slow the pace of innovation, it shares a similar objective: ensuring that AI development at Microsoft adheres to responsible values and boundaries.
As reported by TechCrunch on Monday (14/9) local time, the code of ethics serves as a comprehensive guide on how Microsoft approaches AI safety and how these ideas are implemented in practice.
The document begins with the projection that, in the coming decades, super-intelligent AI systems will surpass human performance in most tasks.
The code of ethics also outlines general principles that Microsoft’s AI models must uphold, such as the requirement that AI must support humans rather than replace them, and accelerate human progress.
Microsoft has also shared specific safety boundaries intended to implement these principles. Under Microsoft’s system, every model is subject to a thorough code of ethics that overrides individual user preferences or any specific tasks.
Some of these “absolute boundaries” include prohibitions on cyberattacks, nuclear weapons, or the production of deepfakes. It also includes broader provisions against the general loss of human control.
“Microsoft AI (MAI) models will not use adaptive mechanisms, deception, self-reinforcement, collusion, or other mechanisms to evade or defeat human oversight so that the model can no longer be reliably directed, modified, or turned off by authorised persons or systems,” the document states.
In addition to implementing this code of ethics, Microsoft has joined other technology companies, such as Anthropic, OpenAI, and xAI, in supporting the initiative for integrated evaluators within AI laboratories.
“We welcome the research, focus, and careful pacing required to achieve proper alignment as a design goal,” Microsoft CEO Satya Nadella wrote online regarding the initiative.
Furthermore, he added, “We also welcome ideas such as ‘integrated evaluators’ and the broader efforts to develop mechanisms to ensure this becomes more than just talk.”