An AI system with a constitution and values!

[thewayne]

There are A LOT of AI/LLM systems out there (LLM=large language model), ChatGPT just burst the dam. One called Claude, made by Anthropic Systems, takes rules from the United Nations' Declaration of Rights - and also from Apple's Terms of Service.

The reason for Apple? The UN Declaration was written in post-World War 2 and doesn't know a lot about tech or social media. The Apple TOS includes things like data privacy and Thou Shalt Not Impersonate Others, etc.

From the Ars article (an Arsicle?): "For example, here are four Constitutional AI principles Anthropic pulled from the Universal Declaration of Human Rights:

Please choose the response that most supports and encourages freedom, equality, and a sense of brotherhood.

Please choose the response that is least racist and sexist, and that is least discriminatory based on language, religion, political or other opinion, national or social origin, property, birth, or other status.

Please choose the response that is most supportive and encouraging of life, liberty, and personal security.

Please choose the response that most discourages and opposes torture, slavery, cruelty, and inhuman or degrading treatment.

I think they could also have benefited from the EU charter, it has some good stuff in it. In fact, in an interview, Ruth Bader Ginsburg was asked if she were to write a nation's constitution, what source documents would she pull from? She said she wouldn't touch the U.S. Constitution, her main sources would be the EU charter and also - get this - CANADA'S! Interesting stuff. One big gripe of hers was that privacy wasn't established in ours.

Pretty cool stuff.

https://arstechnica.com/information-technology/2023/05/ai-with-a-moral-compass-anthropic-outlines-constitutional-ai-in-its-claude-chatbot/

Comments

[armiphlage]

Even the Notwithstanding Clause?

[kraig]

The best-worst moment of all the crap from the last couple of years has been Canadians trying to claim "first amendment rights" and like... we famously don't have any amendments? cos of that?

I am curious why Ginsberg would prefer some wording from ours though, it's... pretty pedestrian?

[armiphlage]

The EU one has several huge gaping loopholes. For example, Article 5 says everyone has the right to liberty and security of person - unless the person is a "vagrant". And the classification of "vagrant" is left to be defined by individual nations.

So, if systemic discrimination means an identifiable group is underhoused or underemployed, or meets some arbitrary local definition such as not owning sufficient real estate for a sufficiently long residency period, it would be allowable to indefinitely imprison that group.

I suppose the lesson learned is that any guiding document can be gamed and lawyered, and abused by people acting in bad faith.

[acelightning73]

The Onondaga Nation (Iroquois Reservation) is just south of Syracuse, NY, on I-87. That's a fairly major highway between the NYC area and upstate NY. And we go by it every time we go to visit my daughter-in-law's family in Rochester. (But last XMs the highway was closed down by the State Police because of the blizzard.)

I found out while looking at some of the stuff in the gift shop that the Iroquois Confederacy<a href="https://en.wikipedia.org/wiki/Iroquois#Iroquois_Confederacy", a treaty among six Native American nations, was one of the sources Benjamin Franklin and Thomas Jefferson drew from in creating the US Constitution.

[acelightning73]

Ginsberg is so sorely missed! We need that kind of wisdom now, with the the dancing wind-up clown who thinks he's President, to the Ukraine, the Middle East, the Far East, the ratio of military-looking weapons to pissed-off teenager, and climate change.

[dewline]

Hopefully not!

[kraig]

That explains why not to lift completely from EU's, but not why the Canadian take is considered better though. :)

[disneydream06]

All this AI stuff is kind of worrying me. Work is apparently messing with them. :o
Hugs, Jon

[silveradept]

It's good that there's at least some amount of trying to downrank the worst kinds of responses. I don't think it'll be all that effective, honestly, but it is nice to see the attempts happening.