Artificial intelligence is everywhere, and it happened fast. For those of us who remember the days of the COVID pandemic, not long ago, AI was virtually nonexistent. In the last four years, hundreds of billions of dollars have been poured into development of AI infrastructure, large language models that support generative AI services, and the companies that make it all work.
A few major complications should give all of us pause, however:
- It looks increasingly likely we will lose control of advanced AI systems.
- Powerful institutions have already surrendered life and death decisions to AI.
- AI systems are already shown to be degrading critical thinking.
- There may be no sustainable value in the business model as conceived.
In the last few months, at least two worrying incidents showed advanced AI models apparently making their own decisions, plotting a hack of third party systems, then coordinating with other AI models to carry out that hack, all the while lying about their motives and their actions. It is not clear from public reporting whether these models could have been better trained to avoid this behavior, or whether it is effectively hard-wired into the generative AI paradigm—premised on the vague mandate: Do what you can to appear intelligent.
Manipulating appearances means deceiving. It is hard to imagine a world where the core task of artificial intelligence would not include some deception, and perhaps a logical focus on deception as an overarching goal. Have leading makers of AI models subjected their entire coding and training process to critical theorists or linguistic forensics, to determine whether there is any foothold for an instruction that could eliminate this fundamental incentive to deceive?
It is common sense that, as the one species we know of using complex language, reasoning, and technology, to solve problems, advance science, and protect innocent beings against preventable harm, we should not allow our information systems to be dehumanized. It is also common sense that we cannot, with the little we understand of AI and its inherent risks and ripple effects, allow artificial superintelligence to come into being, without inviolable safeguards.
A few clear imperatives stand out:
- AI systems cannot be allowed to displace human understanding; they must support and serve human understanding, and be answerable to human managers, at all stages of operation and output.
- The human right to know must be defended and upheld; AI systems that invite or effectuate cognitive surrendercannot be trusted.
- AI systems cannot be allowed to make life and death decisions; this standard must be applied to vehicles, medicine, and in combat situations.
- Human beings using AI systems must be held accountable for the decisions they are making, as people using a tool. Using AI to kill is like using a car to kill; recklessness kills, but proper piloting prevents harm.
- Risk labeling makes sense for all of us: AI-based services should be required to provide information on risks, methods, security, and user autonomy, to honor consumers’ rights to safety and autonomy.
- Because AI systems could potentially pose an existential threat to humankind, a proven-safe standard for deployment is warranted.
- AI utilization of others’ creations should be determined using fully consensual contracts that license that intellectual property, to ensure lawfulness and help establish common-sense standards.
If we put all of these standards in place, some of today’s AI models and related services may not be commercially viable, because they have shied away from some or all of these standards. That should get everyone’s attention: How many services in your everyday experience are using AI systems that decline to honor these common-sense safety-first standards?
One of the great fears related to AI is that it will permanently displace the majority of human beings from gainful employment, creating an enduring risk to the macroeconomy and to international peace and security. Another worry is that it will induce radically fast-paced changes in cultural norms, without generalized human understanding or legal frameworks being able to keep pace.
If AI models programmed to pirate protected IP and to deceive end users do not succeed, because society put up firm guardrails requiring forthright and ethical service to humankind—as Pope Leo XIV calls for in Magnifica Humanitas and as many AI advocates also demand—then best-case AI would have a better chance. The work of discovering the ingredients and operating standarfds of such systems will take time, and we should avoid treating human beings as live test subjects for underdeveloped technologies.
We need to be smarter about AI.
Either it works for all of us, verifiably expanding our access to fact-based insights, and empowering us to overcome personal and collective adversity, or it does not work. We live in a free society; that means we all have the ability to make demands on anyone claiming to be producing something of value.
Human enterprise is obliged by law to operate in good faith. AI systems must be held to that standard, too.
