Human Alignment

With all this talk regarding AI alignment, we’re forgetting that we haven’t solved human alignment yet.

I don’t mean this in a utopian way (no criminals, etc.), but at the most basic institutional and societal level (take the example of the mess that was the covid response).

Why is it that humans are not aligned?

Let me begin by saying that full alignment would not be a good idea anyway: after all, having a few people disagree is a good risk-management strategy, from the point of view of the human species: it avoids going all-in on any single decision. That said, what we’re seeing lately is not just disagreement among a fringe, but massive polarization.

One thing to highlight is that we do not merely disagree on opinions but on facts, too. Going back to the example of Covid, we didn’t just disagree on whether we should wear masks (an opinion) but also on the extent to which masks work (a fact).

And while disagreeing on opinions is normal, disagreeing on facts isn’t. Or, better said, it is normal during an initial period, but as time passes by and information is gathered, facts should start converging. Going back one last time to Covid, what’s particularly bad is that we didn’t just disagree on facts in 2020, but that we still disagree on them in 2026. This points out to a major structural problem.

And of course, the more we disagree on facts, the more we will disagree on opinions. It’s impossible to behave as an aligned society if we disagree on facts.

The government, media, and education system have largely abdicated their role in helping us with that. Of course, I’m not advocating here for propaganda of false facts, just for the sake of alignment. Instead, I’m advocating for the following three processes together: truth-seeking, propagation of truth, and error-correction.

The keyword is together. Truth-seeking alone can produce falsehoods despite the best intentions; hence the necessity to pair it with error-correction processes. Similarly, error-correction without propagation of truth results in useless contrarianism. Instead, what we want is truth-seeking complemented by error-correction, both done by people who know enough true facts to do them properly. Each of these three processes needs the other two to work properly.

Of course, the same applies to companies. Are there effective processes for truth-seeking, propagation of truth, and error-correction? And are they working together, acknowledging that each without the other two will produce poor outcomes?