Translation doesn't need grammar; policy solutions don't need root causes.

An excerpt from a letter published in the Financial Times:

Translation is one of the many positive revolutions brought about by artificial intelligence, but, surprisingly it has gone largely unnoticed. While machine translation has made constant but slow progress since the first Georgetown-IBM experiment in 1954 the shift by Google by 2016 from structural rule-based systems relying on grammatical structures to statistical models and neural machine translation was a major breakthrough. Automated translation now takes into account the context to solve the ambiguity of some words or phrases, and to grasp the tone as well as the style in order to find the right vocabulary and the most suited idiomatic forms. Being trained on vast datasets, it is also very unlikely that it will get stuck on a rare or technical expression that a human translator will in turn struggle to find in a paperback dictionary. How AI is overcoming Europe's language barriers, Philippe Huberdeau, 'Financial Times', 30 April 2025

To me, this has application in the world of policy. Our current policymaking system relies on being able to trace the causes of our social problems, to do so continously, and then to try to deal with them. This identification-based model works well when cause and effect can be reliably identified, and don't change much over time, nor differ much over space.

In today's complex societies, it isn't good enough. There are too many variables; and feedback loops mean that implementation of policies changes how people and systems respond. The current model cannot address such complexities, so we get fossilised policies that also tend to be one-size-fits-all. 

I don't propose that we try to use AI techniques to solve policymaking goals such as reducing crime, improving health or, at the global level, eliminating war. Trying to condense all of human history into the context according to which effective policies could be made is impossible. Reducing the human experience into terms that machines can understand can't be done. Partly because it's too great, partly because it's too subjective and partly because it keeps expanding. 

But what we can learn from the application of AI to language translation is that we don't need to fully understand everything about how a language works - its grammar - to translate it effectively. And this is where the Social Policy Bond concept, with its focus on outcomes, enters the picture. Investors in the bonds don't necessarily need to identify and deal with any supposed 'root causes' of crime, war or any of our other social or environmental pathologies in order to address them effectively. Trying to do so would be a Herculean task and so can readily be invoked as an excuse to do nothing. What we can do, and what a Social Policy Bond regime would do, is give incentives to investors to look for the best ways of solving our problems on a continuous bases. That might, in some cases, mean looking for root causes. But it might simply mean researching, trialling and refining many diverse potential solutions and adopting those that are most promising - constantly. Achieving our social goals requires diverse, adaptive approaches. The current policymaking system delivers neither. 

For more on this theme see here and here.