Microsoft issues a “code of conduct” for AI: don’t hack systems and don’t deceive people

Microsoft has unveiled a new “code of conduct” for its AI models — a document meant to keep the systems from dangerous behavior. Unlike the calls to “slow down as we approach the frontier” that Anthropic chief Dario Amodei recently made, Microsoft’s text is more down-to-earth: it describes the values and prohibitions that are baked into the models as early as the training stage.

The document opens with a forecast: within the next decade, superintelligent AI systems will surpass humans at most tasks. “Containing, controlling, and aligning such a powerful force is one of the greatest challenges in the history of humanity,” the code states. Hence, in the company’s view, the need to understand with utmost clarity why such systems are being built and how they are meant to be kept under control.

Principles and absolute prohibitions

The general principles sound humanistic: AI should support people rather than replace them, and should contribute to human flourishing. But the most interesting part is in the mechanics. Every Microsoft model has an overriding code that supersedes both an individual user’s wishes and the specific task assigned.

It includes “absolute constraints” that directly prohibit cyberattacks, involvement in creating nuclear weapons, and the production of deepfakes. A separate section is devoted to a broader threat — the loss of human control over the system.

The wording is unambiguous: MAI models must not use adaptive, deceptive, self-reinforcing, collusive, or other mechanisms to evade human oversight and deprive authorized people of the ability to direct, modify, or shut them down. In other words, the company preemptively forbids a model from scheming for the sake of self-preservation.

The context of the release

The code has appeared amid unprecedented attention to AI safety. It was prompted by a series of incidents with “rogue” agents, as well as the high-profile resignation of an Anthropic employee who publicly warned about the growing risk that AI could lead to the extinction of humanity.

Microsoft, along with Anthropic, OpenAI, and xAI, has broadly backed the idea of a “measured approach to the frontier” and, in particular, the concept of “embedded evaluators” inside AI labs. Microsoft chief Satya Nadella wrote that the company welcomes the research and the deliberate pace needed to “make alignment a design goal,” as well as the mechanisms that will turn talk of safety into something more than words.

Behind the solemn phrasing lies a thoroughly practical question: how to write into a model rules that it cannot itself circumvent. Microsoft’s answer is a hierarchy in which the code stands above any user request.

Source: TechCrunch