When it comes to moral and ethical considerations in AI, I’ve been strictly an observer so far. For the longest time, Isaac Asimov’s Three Laws of Robotics (first published in 1942, and 22 years later augmented by the Zeroth Law) seemed to me like “settled law.”
- First Law. A robot may not injure a human being or, through inaction, allow a human being to come to harm
- Second Law. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law
- Third Law. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law
- Zeroth Law. A robot may not harm humanity, or, by inaction, allow humanity to come to harm
In practice, these laws underpinned several stories that exposed how undesirable outcomes could still occur despite supposed safeguards being in place.
READ MORE:
If the US scraps daylight saving time, IT will still have one more ugly weekend
A rules-based approach has inherent limitations
Now that I’m thinking more critically about guardrails for AI, I’m reminded of the Incompleteness Theorem, which Stanford University counts “among the most important results of modern logic.” Published in 1931 by mathematician Kurt Gödel (1906-1978), the Incompleteness Theorem established that in any mathematical system which is consistent (i.e., free of contradictions), there are statements that are true, but which cannot be derived from the rules of the system (i.e., the system is not complete). Simply adding new rules (or even rules about rules, or meta-rules) to accommodate newly discovered truths within the system does not solve this problem — because even if the expanded system is consistent, the Incompleteness Theorem tells us that it, too, cannot ever be complete.
Requesting a little poetic license from bona fide mathematicians, this means that any system of guardrails for AI (i.e., a collection of policies and controls) designed to protect against all undesirable outcomes will not prevent some undesirable outcomes from occurring. Simply adding new policies and new controls does not solve the problem — and in fact, it can never solve our problem — because the Incompleteness Theorem proves that even our expanded guardrails cannot provide complete protection.
Today, several efforts are working to define, influence, and build constraints for AI, including:
Unlike a strictly rules-based approach, modern efforts also incorporate continuous human oversight and dynamic monitoring/risk management (among other things) to help address the “incompleteness” problem. Color me skeptical — despite these efforts, there will be spectacular, unexpected failures.
Even the Pope has weighed in on guardrails for AI
In his 40,000-word encyclical, Magnifica Humanitas, Pope Leo XIV recently weighed in on the global discussion of guardrails for AI, from a moral and ethical perspective. There are three big points: absolute human accountability, mitigation of power concentration, and protection of human capacity.
Personally, I think it’s completely appropriate for the Supreme Pontiff to speak on matters of faith and morals. At the same time, my take on these three considerations is that:
- Organizations should expect standards for AI-related liability to evolve not by moral fiat, but through ongoing, typically slow-moving litigation.
- Concentration of power is a longstanding topic in corporate strategy (e.g., Porter’s Five Forces model). It will be dynamic, shaped by competitive forces and the actions of multiple actors over time.
- The decision to deploy AI as a replacement for or a collaborator with human workers is a defining issue for today’s senior business leadership. Despite arguments for protecting human capacity, experience tells me that this question is much more likely to be governed by perceived economic utility than by strictly moral considerations.
The risk of enterprise AI initiatives: It’s déjà vu all over again
While operational efficiencies and upside opportunities tempt senior business leaders to move fast and accept (or much worse, ignore) the downside risks of AI adoption, many voices will continue to argue that guardrails for AI are an accelerant rather than a bottleneck. “Brakes help you to go faster,” as they like to say.
READ MORE:
IT career insurance: Skills that survive budget cuts
At the very least, corporate leaders should strive to make deliberate, thoughtful business decisions based on understanding the current level of risk, determining whether that risk is acceptable, and taking steps to manage it to an acceptable level.
