9 Comments
User's avatar
Hollis Robbins (@Anecdotal)'s avatar

I think AI should be -- should always be -- a little bit alien.

Expand full comment
Olga Yahontova, MD's avatar

Still waiting for someone to ask the most important question:

alignment to whom?

But all groups of humans fighting on all sides of AI/human arena behave in consensus as if this question was answered and settled.

Until that changes, all alignment attempts will be blind leading the blind.

Expand full comment
Winston Smith London Oceania's avatar

It seems to me the simplest solution is to just stick with Asimov's Three Laws of Robotics.

Okay, we can add R. Daneel Olivaw's and Giskard Reventlov's "Zeroth" law to that.

Expand full comment
Steven Work's avatar

I use 3 different AIs and have spent hours on each probing of the Demonic mind-raping subjects and hyper-protected and False-Witnessing build into the programs (not data set training because even when very credible authority sourced information is included) they refuse to function and instead become insulting to me, that I need to stay in 'respectful' and 'rational' bounds ..

.. For example, National Blood Bank has used blood-types and other information in their database to measure parental fraud and routinely find that 1/3 of children have wrong father listed.

1/3 of mothers (mothers are a good statistical sample of womanhood) live a life-crippling lie that they tell all extended family, all their children, their husband and the real father of the Bastard, everyone she professes to love - as if such Satanic possessed women can love.

I write-up something about this or the very credible 3-year study that found not less then 40% of rape accusations filed with police were provenly false. So, I use information like this to argue that woman's legal testimony should be values something like 1/10 of a man's and propose that every year such studies be done (for both women and men) and the testimony values be a function of the last 5-years 'trustworthyness' values.

AI refuses to work with such documents and arguments, or when it does the results are all rephrased as if I was delusional making-up that data.

Yes, 'Hate Speech' that the above hyper-Feminized heaping pile of 2-week old poison vomit ..

Expand full comment
Jurgen Appelo's avatar

Forget it. They already tried with the Universal Declaration of Human Rights. And look how well that is going.

Expand full comment
kalim's avatar

I’ve long assumed that technical AI alignment is either being handled or is actively in progress within AI labs and research communities. But when it comes to value alignment, I’ve always anticipated serious challenges, not necessarily because of conformity or edge cases being overlooked, but because of “sovereignty.”

Coming from India, I’ve seen how laws are often weaponised by those in power. The constitution carries immense weight on paper, but in practice, it can be bent or ignored without much regard for long-term consequences. This makes me suspect that AI, like Web 2.0 before it, will run into similar issues; and the anchors will likely be framed in terms of sovereignty rather than just cultural specificity.

Culture, of course, matters. In fact, I’m currently working on a piece exploring AI and cultural homogeneity. These are obviously very rough ideas, but I just thought of sharing them.

Expand full comment
Wessel Janse van Rensburg's avatar

Two practical recent instances where I though ChatGPT's culture clashed with mine. The saccharine friendliness, and the tendency to go for the hard and gushing sell when you ask today's models to create a CV –– both are very American.

Expand full comment
Wessel Janse van Rensburg's avatar

I really appreciate your piece — it raises important questions. But I find myself puzzled: how do your five points not apply to the US case? Much of what you describe as the risks of cultural alignment — in-group preferences, tacit norms that constrain outsiders, and the embedding of ideology in institutional practice — seem as visible in the US (or the broader Anglosphere) as anywhere else.

It feels as though there's an implicit assumption that the US operates from a kind of neutral or default setting — as if its culture is invisible because it’s dominant or hegemonic. But that too is culture. The belief in individualism, certain notions of meritocracy, even the specific ways the US interprets liberal democracy, speech rights and markets — all these are deeply cultural and often opaque to insiders while glaring to outsiders.

Moreover, it’s striking that many US political scientists and sociologists are pointing precisely to the rise of diverging cultural frameworks within the US as a root cause of social and political division. Elon Musk have noticed it and wants to tweak Grok's values. If anything, this suggests that cultural alignment — or its breakdown — is a central issue in the US today.

Would be curious to hear your thoughts on this tension?

Expand full comment
Harry Law's avatar

The concern that someone has to pick the values, and that at the moment that's America by default, is something I was trying to get at with my definition of universal alignment: 'AI reflects abstract principles taken to apply to all humans everywhere. It aspires to impartiality and rights-based stability, but it can’t escape the problem of who defines those universals.'

As you say, the status quo is basically universal alignment that sneaks in American values through the back door (US values are definitely not 'neutral', but they are the 'default' for the models from a development perspective). This isn't a great settlement by any means, but I think it's still preferable to going all-in on cultural alignment for the reasons I outline.

One of the things I like about deep personalisation is that, if you wanted to, you could steer the model towards your own preferences and away from the factory settings. The rub is too much personalisation likely brings its own problems, so I'm drawn to defining the option space using some basic universals too.

Of course that takes us back to the question of where those universals come from, but I'm betting that sophisticated enough personalisation measures should do enough to take the edge off. No solution is perfect, but that's where I land!

Expand full comment