Robot Ethics: Why “Friendly AI” Won’t Work

Thomas Blake, Bernd Carsten Stahl and N.B. Fairweather


Artificial agents, from unembodied web bots to robots, are becoming a real part of our world. In time, we may be able to create beings with human-level intelligence. As with any agent that can act upon the world, we should expect such beings to act well. But how can we expect good behavior from such an alien being? The Singularity Institute has proposed the concept of “friendly AI” to answer this question. I will show that their analysis of the issue is insufficient and propose another approach to creating ethical intelligent artificial agents.

The Singularity Institute wrote Creating Friendly AI (CFAI) based on the concepts from General Intelligence and Seed AI (GISAI). The goal of GISAI was to describe, in the abstract, how to create an artificial general intelligence, bearing in mind the failures of past attempts at AI. In general, the method adopted by GISAI was to create a “seed AI”, a mind which may not have human intelligence but instead has the ability to improve itself. Such an AI would be able to gradually reach human intelligence (and better-than-human intelligence) by way of small self-improvements.

CFAI is an attempt to determine how to ensure that a seed AI treats us well. According to the Singularity Institute, there is little reason for us to assume that an AI would share our values and goals. Our values and goals come largely from our evolutionary history, and a being that does not share that history should not be expected to share those goals. CFAI suggests that engineers who create AI should attempt to imbue them with ethics.

CFAI advocates making “friendly” AI, beings whose goals and values include, for instance, being sympathetic towards humanity. However, this approach will not work either. An AI that can change its own programming might pursue some of its goals and values by changing others. If this is truly an alien being, then many of its goals and values might conflict with artificially inserted directives like “feel sympathetic towards humanity”. When this happens, the goals which do not mesh with the AI’s nature are likely to be the first to go.

As I have argued elsewhere, ones values come from ones nature. If Robby the Robot has a nature that is alien to us, we should not expect what is good for Robby to be good for us. If we want Robby to behave “ethically”, that is to behave in a way that would be good for a human, then we must ensure that Robby’s nature necessitates an ethics that is in line with ours. If we want Robby to respect and protect humans, then Robby should be a social being that we treat as an equal.

It follows that we should attempt to create artificial agents with a nature similar to ours. Then, an ethics like ours would follow from its nature, and the artificial agent should have no more trouble being good than we do. Of course, this is no guarantee that our creations will be good, but we shouldn’t expect more of ourselves than we do of God.


Blake, Thomas. “Technological Transcendence: Why It’s Okay that the Future Doesn’t Need Us”. Proceedings of ETHICOMP2007, Tokyo, Japan.

Singularity Institute (orig. Eliezer Yudkowsky). “Creating Friendly AI”. online at

Singularity Institute (orig. Eliezer Yudkowsky). “General Intelligence and Seed AI”. online at