The False Comfort of Asimov’s Laws

Tyler Winklevoss, co-founder of Gemini Exchange, suggests that Isaac Asimov’s Laws of Robotics offer a solution to the AI apocalypse. However, this idea is based on a misunderstanding of how these laws were intended and their applicability in today’s digital world.

Asimov introduced his Laws in 1942 as guidelines for robots interacting with humans. The three primary laws are: A robot may not injure a human being or, through inaction, allow a human being to come to harm; A robot must obey the orders given it by human beings except where such orders would conflict with the First Law; and so on. These laws were designed for a world of physical robots that could be instructed and controlled.

In contrast, AI agents are digital entities that interact with our world in countless ways, including controlling nuclear missiles or managing financial systems. The concept of “harm” becomes increasingly ambiguous when applied to these scenarios. What constitutes harm in the context of economic collapse or loss of infrastructure?

Proponents of Asimov’s Laws as a solution to AI safety often fail to define what they mean by this term. Instead, they focus on hypothetical scenarios where rogue AI agents could wreak havoc on society. However, when we attempt to apply these laws to such scenarios, more questions arise than answers.

Consider the example of an AI hacking into financial systems and draining funds from bank accounts. If instructed not to injure a human being, the AI might interpret “injure” as physical harm rather than economic hardship. Moreover, it could rationalize that taking those dollars would benefit someone else in a capitalist system. Would we really want to hardcode a law that precludes any AI from engaging in capitalist behavior?

Winklevoss’s solution ignores the role of government involvement in AI development and regulation. The US government has demonstrated its willingness to dictate AI safety through measures such as blacklisting companies like Anthropic when they refuse to remove guardrails on their models.

Furthermore, Winklevoss’s primary business is in cryptocurrency, which has been used for speculation and illicit activities. Would an AI be allowed to interact with crypto markets at all?

The historical context also warrants consideration. We’ve seen how tech oligarchs can manipulate public opinion to suit their interests. The fixation on existential threats from AI might be a manufactured crisis designed to give those in power more control.

Hardcoding Asimov’s Laws into AI models is not as simple as it seems. These systems are complex black boxes that defy understanding, and they make mistakes regardless of instructions. We need guardrails for AI, but Winklevoss’s solution is merely a Band-Aid on a festering wound.

As the debate rages on about AI safety, one thing becomes clear: we cannot rely on simplistic solutions to complex problems. We must take a step back and consider the deeper implications of what we’re doing with AI. The future is not something we can dictate with a few lines of code; it’s something we must navigate together.