Was there actually a case of a model saying "America's founding father were black women", or is that just Elon fingering your amygdala with a ridiculous hypothetical that exists nowhere other than Elon's mind in order to justify Elon's personal bias tweaks when he doesn't like the wisdom-of-the-crowds answer his tools initially give?
The most important part of AI safety is AI alignment: making sure AI does what we want. It's very hard because even if AI isn't trying to deceive you it can have bad outcomes by executing your request to the letter. The classical example is tasking an AI to make paperclips, training the AI with a reward for making more paperclips. Then the AI makes the most paperclips possible by strip mining the Earth and killing anything in its way.
Sometimes you see this AI alignment problem in action. I once asked an older model to fix the tests and it eventually gave up and just deleted them
> Still waiting for an explicit answer on understand how 'safety' is truly distinguishable from 'censorship' or 'political correctness'
i've said this many times but the concept of ai "safety" is really brand safety. What Anthropic is saying is they're willing to risk some bad press to bypass the additional training and find tuning to ensure their models do not output something people may find outrageous.
> I VERY LARGELY prefer an AI like grok that doesn't pretend and let the onus of interpretation to the user rather than a bunch of anonymous "researchers" that may be equally biased, at the extreme, may tell you that America's founding father were black women
Setting aside for a moment that Grok is manipulated and biased to a hilarious extent. ("Elon is world champion at everything, including drinking piss")
There is no such thing as "unbiased". There will always be bias in these systems, whether picked up from the training data, or the choices made by the AI's developers/researchers, even if the latter doesn't "intend" to add any bias.
Ignoring this problem doesn't magically create a bias-free AI that "speaks the truth about the founding fathers". The bias in the training data, the implicit unconcious bias in the design decisions, that didn't come out of thin air. It's just somebody else's bias.
All the existing texts on the founding fathers are filled with 250 years of bias, propaganda, and agenda pushing from all sorts of authors.
There is no way to have no bias, no propaganda, no "agenda pushing" in the AI. The only thing that can be done is to acknowledge this problem, and try to steer the system to a neutral position. That will be "agenda pushing" of one's own, but that's the reality of all history and all historians since Herodotus. You just have to be honest about it.
And you will observe that current AI companies are excessively lazy about this. They do not put in the work, but instead slap on a prompt begging the system to "pls be diverse" and try to call it a day. This does not work.
> Of course saying to someone to go kill himslef is a prety sure 'no-no' but so many things are up to interpretation.
Bear in mind that the context of Anthropic's pivot here are the Pentagon's dollars.
This isn't just about "anti-woke AI", it's about killbots.
Sure, Hegseth wants his robots to not do thoughtcrime about, say, trans people or the role of women in the military.
But above all he wants to do a lot of murder.
Antrophic dropping their position of "We shouldn't turn this technology we can barely control into murder machines" because they're running out of money is damnable.
This is a very fair answer but missing some points.
I do personally believe that grok is a less biased against too many PC answers but you may disagree.
"All the existing texts on the founding fathers are filled with 250 years of bias, propaganda, and agenda pushing from all sorts of authors."
not sure of the point is tho ? Mine is that gemini was biaised so hard that it was generating diverse founding fathers which is factually untrue.
The fact that history has a pro-american values when written by americans is also true but it has nothing to do really with the argument: if an IA is able to see through such propaganda and provide a balanced view on it as a human would this is enough
In fact, i just asked grok "is the american founding constitution inhernetly good" and it gave me an aswer way more balanced than most american would i believe:
"The U.S. Constitution, drafted in 1787 and ratified in 1788, is a foundational document that has shaped American governance and influenced democracies worldwide. Asking if it's "inherently good" (assuming that's what you meant by "inhernetly") invites a philosophical debate: "Inherent" implies something intrinsic and unchanging, independent of context, interpretation, or outcomes. Goodness, in this case, could mean moral, effective, just, or beneficial to society. From a truth-seeking perspective, I'll break this down non-partisanly, drawing on historical facts, strengths, criticisms, and evolving views. Spoiler: It's not inherently anything—it's a human creation with profound virtues but also deep flaws, and its "goodness" depends on how it's applied and amended."
[can't paste everything so here's the conclusion]
"Is It Inherently Good? A Balanced VerdictNo document is "inherently" good or bad—goodness is contextual and subjective. The Constitution isn't divine or eternal; it's a pragmatic compromise by flawed humans (55 delegates, all white men, many slaveowners). It has proven remarkably resilient and improvable, outlasting many governments, but it's not perfect or immune to abuse. Its goodness lies in its capacity for self-correction: 27 amendments have fixed some issues, though others (like wealth inequality or climate inaction) persist due to gridlock.If you're measuring by outcomes, the U.S. has achieved extraordinary things under it, but at great human cost—think Civil War, civil rights struggles, and ongoing divides. Philosophically, as Grok, I'd say tools like this are as good as the people wielding them. If "inherently good" means it embodies universal moral truths, partially yes (liberty, equality under law). But if it means flawless or unbiased, absolutely not.What aspect of the Constitution are you most curious about—its history, specific clauses, or modern reforms? That could help refine this."
So it's definetely seeing through any form of propaganda you desribe
You understood the issue so well but still made the mistake you identified, by claiming that "neutral" exists. "Neutral" is a synonym for "bias toward status quo"
Well we teach kids not to yell “Fire!” In a crowded theatre or “N***!“ at their neighbor. We also teach our industrial machines to distinguish between fingers and bolts, our cars to not say “make a left turn now” when on a bridge, etc
Funny scene, but it also illustrates a more serious point about (human) alignment - not all humans believe exactly the same things are good and bad, or consistently act in accordance with what they claim they believe is good. This is such a basic fact of human social life that it's almost banal to point it out explicitly; but if (specific) human beings or (specific) organizations of human beings are trying to align the AIs they are creating to human values, it will eventually become apparent that the notion of "human values" stops being coherent once you zoom in enough. Humans don't all share the same values, we aren't completely aligned with each other.
Is "we" the parents teaching their children their own unique values, or is the "we" a government or corporation forcing one set of values on all children.
Why not encourage the users of AI to use a Safety.md (populated with some reasonable but optional defaults)?
floatrock|4 days ago
bumby|4 days ago
Also, snarky tone is against the HN guidelines.
wattsy2025|4 days ago
Sometimes you see this AI alignment problem in action. I once asked an older model to fix the tests and it eventually gave up and just deleted them
chasd00|4 days ago
i've said this many times but the concept of ai "safety" is really brand safety. What Anthropic is saying is they're willing to risk some bad press to bypass the additional training and find tuning to ensure their models do not output something people may find outrageous.
SlinkyOnStairs|4 days ago
Setting aside for a moment that Grok is manipulated and biased to a hilarious extent. ("Elon is world champion at everything, including drinking piss")
There is no such thing as "unbiased". There will always be bias in these systems, whether picked up from the training data, or the choices made by the AI's developers/researchers, even if the latter doesn't "intend" to add any bias.
Ignoring this problem doesn't magically create a bias-free AI that "speaks the truth about the founding fathers". The bias in the training data, the implicit unconcious bias in the design decisions, that didn't come out of thin air. It's just somebody else's bias.
All the existing texts on the founding fathers are filled with 250 years of bias, propaganda, and agenda pushing from all sorts of authors.
There is no way to have no bias, no propaganda, no "agenda pushing" in the AI. The only thing that can be done is to acknowledge this problem, and try to steer the system to a neutral position. That will be "agenda pushing" of one's own, but that's the reality of all history and all historians since Herodotus. You just have to be honest about it.
And you will observe that current AI companies are excessively lazy about this. They do not put in the work, but instead slap on a prompt begging the system to "pls be diverse" and try to call it a day. This does not work.
> Of course saying to someone to go kill himslef is a prety sure 'no-no' but so many things are up to interpretation.
Bear in mind that the context of Anthropic's pivot here are the Pentagon's dollars.
This isn't just about "anti-woke AI", it's about killbots.
Sure, Hegseth wants his robots to not do thoughtcrime about, say, trans people or the role of women in the military.
But above all he wants to do a lot of murder.
Antrophic dropping their position of "We shouldn't turn this technology we can barely control into murder machines" because they're running out of money is damnable.
davidguetta|4 days ago
I do personally believe that grok is a less biased against too many PC answers but you may disagree.
"All the existing texts on the founding fathers are filled with 250 years of bias, propaganda, and agenda pushing from all sorts of authors."
not sure of the point is tho ? Mine is that gemini was biaised so hard that it was generating diverse founding fathers which is factually untrue.
The fact that history has a pro-american values when written by americans is also true but it has nothing to do really with the argument: if an IA is able to see through such propaganda and provide a balanced view on it as a human would this is enough
In fact, i just asked grok "is the american founding constitution inhernetly good" and it gave me an aswer way more balanced than most american would i believe:
"The U.S. Constitution, drafted in 1787 and ratified in 1788, is a foundational document that has shaped American governance and influenced democracies worldwide. Asking if it's "inherently good" (assuming that's what you meant by "inhernetly") invites a philosophical debate: "Inherent" implies something intrinsic and unchanging, independent of context, interpretation, or outcomes. Goodness, in this case, could mean moral, effective, just, or beneficial to society. From a truth-seeking perspective, I'll break this down non-partisanly, drawing on historical facts, strengths, criticisms, and evolving views. Spoiler: It's not inherently anything—it's a human creation with profound virtues but also deep flaws, and its "goodness" depends on how it's applied and amended."
[can't paste everything so here's the conclusion]
"Is It Inherently Good? A Balanced VerdictNo document is "inherently" good or bad—goodness is contextual and subjective. The Constitution isn't divine or eternal; it's a pragmatic compromise by flawed humans (55 delegates, all white men, many slaveowners). It has proven remarkably resilient and improvable, outlasting many governments, but it's not perfect or immune to abuse. Its goodness lies in its capacity for self-correction: 27 amendments have fixed some issues, though others (like wealth inequality or climate inaction) persist due to gridlock.If you're measuring by outcomes, the U.S. has achieved extraordinary things under it, but at great human cost—think Civil War, civil rights struggles, and ongoing divides. Philosophically, as Grok, I'd say tools like this are as good as the people wielding them. If "inherently good" means it embodies universal moral truths, partially yes (liberty, equality under law). But if it means flawless or unbiased, absolutely not.What aspect of the Constitution are you most curious about—its history, specific clauses, or modern reforms? That could help refine this."
So it's definetely seeing through any form of propaganda you desribe
lupire|4 days ago
gehwartzen|4 days ago
JuniperMesos|4 days ago
> Huey: It means don't act like niggas
> Grandad: S-see, that's what I'm talkin' about right there. We don't use the n-word in this house
> Huey: Grandad, you said the word "nigga" 46 times yesterday. I counted
> Grandad: Nigga, hush
https://www.youtube.com/watch?v=TLodIw5iKX8
Funny scene, but it also illustrates a more serious point about (human) alignment - not all humans believe exactly the same things are good and bad, or consistently act in accordance with what they claim they believe is good. This is such a basic fact of human social life that it's almost banal to point it out explicitly; but if (specific) human beings or (specific) organizations of human beings are trying to align the AIs they are creating to human values, it will eventually become apparent that the notion of "human values" stops being coherent once you zoom in enough. Humans don't all share the same values, we aren't completely aligned with each other.
rudhdb773b|4 days ago
Is "we" the parents teaching their children their own unique values, or is the "we" a government or corporation forcing one set of values on all children.
Why not encourage the users of AI to use a Safety.md (populated with some reasonable but optional defaults)?
miltonlost|4 days ago