Zeroth Law: A robot may not injure humanity or, through inaction, allow humanity to come to harm.
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.
Isaac Asimov introduced the First Law of Robotics (and the word ‘robotics’ itself) in one of his earliest robot stories, “Liar!” The First Law, which is baked into every robot’s positronic brain, reads, “A robot may not injure a human being or, through inaction, allow a human being to come to harm.” In the story, the robot Herbie (RB-34) follows a simplistic interpretation of the First Law, and tries to please his human handlers with lies that give them short-term happiness, at the cost of greater long-term suffering, and, arguably, real harm. Eventually caught in his web of lies, and faced with the inescapability of causing further harm, poor Herbie undergoes a sort of logic bomb and goes catatonic.
One obvious solution to prevent future Herbies would be to add a law compelling them to be truthful, or at least to stay silent rather than lie. Asimov did add more laws in later stories, but not a single one was about truth or any form of rationality. Asimov’s Laws of Robotics cover only two areas: interest and authority.
Asimov’s First and Third laws imply that a robot’s primary interest is the well-being of individual humans, while its secondary interest is its own well-being. Sandwiched between those two laws of interest is respect for authority: a robot must obey humans. Finding these three laws to be inadequate, Asimov added to the list of interests a Zeroth Law: “A robot may not injure humanity or, through inaction, allow humanity to come to harm.”
Authority would have to be prioritized—a robot could not be taking orders from any random passer-by in conflict with the owner’s direction. And interest… Asimov’s laws are simple, and acknowledge no trade-offs, like trolley problems—that’s the grist for his stories’ plots. The robots are *presumed* to be rational, so they have to perform all manner of entertaining contortions to follow the laws. But should Asimov have included some sort of law representing rationality? His laws would still have remained manageably countable on one hand.
“A robot must be rational” would make a nice start, but it would require detail—what would be some candidate rationality principles? Truth, facts, logic, probability, statistics, Bayesianism, game theory, avoiding bias… Presumably a robot, unlike humans, would have mathy things built in, but it could still lie. The most obvious rationality principle to include explicitly would be honesty.
It’s worth noting here that there are two kinds of rationality: epistemic and instrumental, i.e. knowing truth and achieving goals. Honesty lies at the intersection of rationality and interest. Dishonesty can further some ends, but it interferes with the collective pursuit of truth. Dishonesty sounds like a tempting but risky tool in the toolbox, and that is indeed how we treat it throughout human culture. From Aesop’s fable “The Boy Who Cried Wolf”, to “Pinocchio”, to Christianity’s Deceiver and Father of Lies, we look at dishonesty as something sinful. Lying evokes disgust. Indeed, in the modern science-fiction series The Three-Body Problem, an alien race considers the very ability to lie so abhorrent that it marks the human race for extinction.
Dishonesty contaminates and spoils the epistemic environment, what psychologist Steven Pinker calls the rationality commons. More than half a century ago, philosopher Hannah Arendt commented on how lying contaminates the information environment and corrupts people:
"This constant lying is not aimed at making the people believe a lie, but at ensuring that no one believes anything anymore. A people that can no longer distinguish between truth and lies cannot distinguish between right and wrong. And such a people, deprived of the power to think and judge, is, without knowing and willing it, completely subjected to the rule of lies. With such a people, you can do whatever you want." ~Hannah Arendt
Today one hears constant complaints about dishonesty in the mass media, often motivated by political tribalism, government disinformation campaigns, or paternalism. In what some call the post-truth era, one infamous source of disinformation is the Russian government, employing Soviet-era techniques amplified by modern AI and bot farms. In the US, over-enthusiastic “fact-checking” of social media created the opposite problem: suppressing heretical information. Even in the ivory tower, supposedly immune to such influences, academic research has become increasingly suspect. Paternalism is the top-down version of the robot Herbie’s bottom-up lying: a noble lie for the simple folk who can’t handle the truth. But such white lies are often self-serving; as Thomas Sowell said, “When you want to help people, you tell them the truth. When you want to help yourself, you tell them what they want to hear.”
This is all hard to solve for humans, but certainly we can try to embed honesty in our AIs. AI hallucinations trouble us, but nowhere near as much as the prospect of AI dishonesty, notably scheming: an AI’s feigning alignment with moral principles. So it is no surprise that Anthropic’s Claude AI, famous for its Constitutional AI, includes honesty in its fundamental tripartite principle: Be helpful, honest, and harmless. Its other two components correspond roughly to Asimov’s, and though one might quibble with the formulation, alliteration makes it memorable. Beyond honesty is explainability: Make the AI’s thought processes transparent and traceable, so that honesty is not a matter of choice.
Even with rationality principles embedded, an AI, like a lone human, is limited by available information, both for training and reference. And that’s where we share an interest: increasing the quality of the knowledge commons. This is the goal of the ASIMOV Protocol. We cannot force truthfulness, but we can support and incentivize it.
The maxim “Honesty is the best policy” dates back more than four centuries, but Americans most associate it with Benjamin Franklin. If we accept it literally, honesty should be the first law of robotics. Franklin might agree with that: he was once beaten at chess by the Mechanical Turk, a chess-playing automaton that turned out to be a hoax.
But should honesty or some other element of rationality have been the first law, ahead of a principle of interest or respect for authority? One can argue for any sequencing, and such a discussion would be interesting. The primacy of one law over another is a common topic in Asimov’s stories. But it is simpler to take three fundamental principles—rationality, respect for authority, and locus of interest—as an inseparable triad, all necessary for action.
Are laws of robotics or constitutional AI at all realistic? Political scientist and modern-warfare specialist Peter W. Singer (not to be confused with moral philosopher Peter Singer), says Isaac Asimov’s Laws of Robotics are wrong. But let’s be charitable: few would argue that the Ten Commandments or Buddhism’s precepts are perfect and complete, but they have proved durable and provided many with useful guidance. As with Asimov’s laws, one can analyze them for deeper structures.
While one can criticize such formulations, at the end of the day, they are essential. When we have real people, and now real robots, interacting in the real world, we need real rules.
—--------
Benjamin Franklin and Isaac Asimov were both polymath scientists who both made famous contributions to ethics. They were both prolific writers, also of autobiographies; The Autobiography of Benjamin Franklin and I, Asimov are both interesting and inspiring. However, their similarities diverged with chess: Asimov played poorly, but Franklin was fairly good, and once played against a faux robot–the Mechanical Turk.