AI risk mitigation already exists
AI is only a threat when it interfaces with everything, information is opaque, and tainting it is like poisoning a well.
I've danced around the edges of the existential AI risk debate for years. While I've predicted consequences for the labor market, I've never viewed AI with awe or fear. The factors that make AI potentially dangerous, and those that can mitigate these risks, have often been overcomplicated.
Here’s a Perplexity collection outlining the risks AI poses. Some are just a sociologist’s Barnum statement, others a theologian’s imagination. All I have to say to them is yes, these are problems that humans have always struggled with. AI will nourish the good and the bad. Trying to prevent people from using tools in ways that undermine civilization’s cohesion is a futile pursuit. Better to find new ways to cohere.
Onto more distinct and tangible concerns. Here are the two risks that I can take seriously and their presumptions. || stands for “or”.
A person asks an AI to do something. The AI does something at the cost of
everythinglots we hold dear.Humans build interfaces between AIs and things of extreme influence.
Any person can link an AI to things of extreme influence. || AI controlling things of extreme influence can be commanded by anyone.
A person asks an AI to help them do something extremely bad. The AI does it, and the person does something extremely bad.
The person pursues their malicious objective undetected.
The person can access to all of the resources the AI advises them to use. || The AI can gain access to any private resources the person asks of them.
When outlining the requirements for the inputs to produce the outputs, the concerns begin to subside.
Access: Not everyone has access to things of extreme influence. Society is very selective about who does.
Safety systems: Extensive systems have been built so that people in control of things of extreme influence must be deliberate when performing consequential actions. These systems sometimes fail catastrophically without AI causes, but they largely succeed. Why assume those responsible for the systems will dismantle them and hand control over to AI apps?
Dispersion: The copies of ChatGPT weights, across some computers, wrapped by code that regulates what actions the model output can perform are separate from the copies of NSA-forked LLAMA weights, across some computers, wrapped by code that regulates what actions the model output can perform. “AI” should be conceived in the singular, not the collective.
Monitoring: Tech companies already monitor activity for unlawful behavior. It isn’t a massive ask for many of those with chatbots to process prompts & responses post hoc to detect any intent to commit atrocities. Fun fact: they are composed of humans who don’t want someone to build a species-ending pathogen.
Not God: The moment we assume an AI can gain access to any secure resource, we violate points 1, 2, and 4 and elevate it to the status of God. AIs are not omnipotent.
Information availability: The AI learned the “dangerous” information from somewhere on the internet. There are plenty of papers, websites, and manuals describing how to do things that we don’t want everyone to do. It’s easier to get to that information now, but the hard part is usually the details you have to figure out while doing it.
The heralds of AI risk often make the same errors as those who initially feared self-driving cars, overlooking that humans are often the more unpredictable element. Similar solutions for AI are on their way.
Decentralization: Running AI on personal devices would enhance security by limiting computational power, increasing traceability, and provisioning an API gate between servers and agents.
Blockchain: Publicly traceable transactions could help monitor malicious activities. That plutonium won’t be free.
It’s worth noting the volatility of this technology's early years should allow for the proliferation and evolutionary selection of mitigation strategies we have never considered.
AI risk proponents too often treat AIs as a singular god – an agent mindlessly granted control of all systems of extreme influence but easily accessed, deceived, and exploited by petty criminals. The nature of these systems is everywhere and anywhere to them. Their solutions often propose some anointed class deemed intelligent, responsible, and ethical enough to regulate how others can leverage the technology. However, many risk-mitigating systems are already in place for humans that can be adapted for machine agents. Decentralized development reduces risk as well. Resorting to centralized power and bureaucratic approval pipelines so soon is neurotically overhasty at best and dangerously arrogant at worst.