Will AI systems remain beyond the law?

I now have family living in St. Albans. Recently, I found out that Sir Francis Bacon was a famous son of that town. Bacon has been called the father of empiricism. So, naturally, I had to find out more about the man’s work and ideas. He lived from 1561 to 1626, following the dark Middle Ages. His fame rests (among others) on his belief that science could be achieved using a skeptical and methodical approach and scientific experimentation. His book “Novum organum” on this topic proved to be highly influential in the 17th century.

In 1626, a utopian novel by Bacon, “New Atlantis,” was published posthumously. It described the House of Salomon, a state-sponsored scientific institution. It was a blueprint for the structure of discovery. Bacon outlined certain roles: “Merchants of Light” who gathered knowledge globally, experimenters who tested hypotheses, interpreters who synthesized findings into higher-order principles. Interestingly, the House of Salomon decides on its own which of their discoveries to keep secret, thus putting itself beyond the law.

Of course, one is compelled to see similarities between the House of Salomon and universities, and especially artificial intelligence laboratories. Today’s AI labs scour global information streams, conduct structured training, benchmark results and iteratively refine outputs. Data collection teams echo Bacon’s emissaries. Research scientists resemble his experimental philosophers. Alignment and evaluation groups serve as modern interpreters, aiming at turning the outcome into a useful and hopefully reliable tool. Unlike the House of Salomon, however, universities are bound by the state and the law.

AI systems aren’t necessarily. The ethics underpinning the workings and the growing societal influence of these systems are largely determined by their creators. This is a murky, unregulated process, which will become even murkier when their need to generate significant revenue grows.

The AI company most open about the innards of its product, Claude, is Anthropic. It embedded a “Constitution” directly into its training and alignment processes. It uses a written set of principles to guide model behavior during training, shaping how the system critiques and revises its own outputs. It was written largely by a single person with a background in moral philosophy. This constitution is an interesting read. It disturbingly uses the words “we hope” quite often. And, however well-intended, it’s still an internal, unregulated document.

Francis Bacon was brilliant but also quite naïve. He had a stellar career in law and became Lord Chancellor to James I. But toward the end of his career, his naivety prevented him from realizing the pressures of the world he lived in. He was impeached for corruption, heavily fined and briefly imprisoned in the Tower. The author of a utopia of ordered inquiry was reminded that the real world is messy and, in the end, the law prevails.

That’s not to say that the architect of Claude should be imprisoned. Quite the contrary, she should be lauded for her effort. But, however principled their design, AI companies sit within competitive markets, regulatory crosswinds, investor pressures and geopolitical rivalries. Power and wealth accumulate around successful enterprises.

AI systems now operate at scales Bacon could scarcely have imagined for his House of Salomon. The question is whether they become so powerful that the law won’t be able to confine them anymore.