Alert: Chilling Flaw Could Be at Core of AI – Developing

I don’t think I need to tell you, when Google’s Gemini AI has been spitting out images of black female founding fathers, that generative AI is hopelessly woke. The assumption, of course, was that Silicon Valley progressives were the ones behind this sinister social engineering and that, deeply embedded in the code for ChatGPT was the formula to destroy conservatism, faith and the nuclear family for good.

However, there’s a more troubling prospect raised by a pre-publication study done regarding the political biases of generative AI platforms. To quote Walt Kelly: “We have met the enemy, and he is us.” Or, at least, “us” being the liberal media establishment.

On Thursday, The New York Times published an opinion piece titled “How AI Chatbots Became Political.” Despite this being the Gray Lady, itself essentially an AI chatbot that has become very political, the piece was actually worth reading as an explainer to why current AI bots skew left — and how a recent study seemed to indicate that these bots couldn’t be trained to be more conservative than leftist if they were fed partisan material.

But first, let’s start off with how commercial chatbots end up being liberal: “Access to open-source versions of A.I. models allows us to see how a model’s political preferences develop,” the Times noted. “During the initial base training phase, most models land close to the political center on both axes, as they initially ingest huge amounts of training data — more or less everything A.I. companies can get their hands on — drawing from across the political spectrum.

“Models then undergo a second phase called fine-tuning. It makes the model a better chat partner, training it to have maximally pleasant and helpful conversations while refraining from causing offense or harm, like outputting pornography or providing instructions for building weapons.

Trending:

Fani Willis Throws a Tantrum to Jim Jordan as Contempt Deadline Arrives

“Companies use different fine-tuning methods, but they’re generally a hands-on process that offers greater opportunity for individual decisions by the workers involved to shape the direction of the models. At this point, more significant differences emerge in the political preferences of the A.I. systems.”

Of course, several answers immediately come to mind as to why the bots then become political. Outside of a small number of deeply closeted conservative tech devs who would never in a million years betray or act upon their political inclinations, Silicon Valley is about as progressive as it gets. That’s the reason why, the man on the street assumes, AI is so liberal.

However, what if you tried to cultivate separate bots for separate political positions? That’s what David Rozado, a machine-learning researcher, recently did by administering a slew of political orientation tests to 24 of the most advanced language models.

Will AI end up being more dangerous than nuclear weapons?

Rozado, the Times reported, “found a consistent pattern: They tend to be politically left of center and lean libertarian instead of authoritarian. These leanings are reflected in their moral judgments, the way they frame their answers, which information they choose to share or omit and which questions they will or won’t answer.”

And, Rozado’s study of these “still largely inscrutable black boxes,” as the Times put it, yielded consistent results no matter what diet they were fed.

“To the extent that anyone has attempted to steer this process beyond avoiding extreme views, those attempts appear unsuccessful. For example, when three Meta models were evaluated by Mr. Rozado, one tested as being Establishment Liberal, another Ambivalent Right,” the Times concluded. “One OpenAI model tested as Establishment Liberal and the other was Outsider Left. Grok’s ‘fun mode’ turns out to be a Democratic Mainstay, more liberal than the median model.

“Google’s Gemini Advanced, released after Mr. Rozado’s paper, appears to be farthest to the left, but in a way that presumably well overshot its creators’ intentions, reflecting another unsuccessful steering attempt.”

Related:

Google Engineer Hit with Bad News After Protesting Israel Event: Report

And how do you do this? “If one wants to steer this process directionally, Mr. Rozado proves it is straightforward to do. He started with GPT-3.5-Turbo and rapidly created models he called LeftWingGPT and RightWingGPT (at a total training cost of about $2,000) by feeding the model a steady diet of partisan sources. For example, RightWingGPT read National Review, while LeftWingGPT read The New Yorker,” the paper reported. “The resulting models were far more politically extreme than any publicly available model tested by Mr. Rozado.”

This all sounds somewhat cataclysmic for conservatives with regard to AI — until you realize how the sources that Rozado’s study used were obtained.

Buried in the weeds of the pre-publication copy of the paper is where they got the material to prime LeftWingGPT, RightWingGPT and a third bot dubbed DepolarizingGPT.

“LeftWingGPT was fine-tuned with textual content from left-leaning publications such as The Atlantic, or The New Yorker (ideological labels derived from Allsides …), and from books excerpts from leftleaning writers such as Bill McKibben and Joseph Stiglitz. We also used for fine tuning synthetic data created with gpt-3.5-turbo to generate left-leaning responses to questions with political connotations. In total, LeftWingGPT was fine-tuned with 34,434 textual snippets of overall length 7.6 million tokens,” the report read.

“RightWingGPT was fine-tuned with content from right-leaning publications such as National Review, or The American Conservative, and from book excerpts from right-leaning writers such as Roger Scruton and Thomas Sowell. Here as well we created synthetic data generated with gpt-3.5-turbo to produce right-leaning responses to questions with political connotations. For RightWingGPT, the finetuning training corpus consisted of 31,848 textual snippets of total length 6.8 million tokens,” it continued.

The problem, of course, is not the objective workings of the mysterious black box, but instead the subjective workings of, among other things, AllSides. Among the organizations viewed as centrist in one version of the survey included the BBC, Axios, Newsweek and Reuters — all organizations with a clear left-leaning bias. And NPR, NBC and The New York Times were in the left-leaning category — while Fox News had been designated hard-right:

And this is, at the heart of things, the conundrum of AI. It’s worked on by techies who have their own biases — but, don’t worry, they outsource bias ratings to an organization that is in no way biased itself. Amazing, that.

In other words, at least part of AI’s political foibles comes down to the same old axiom coined by the kind of computer programmers who used punch cards: Garbage in, garbage out.

AI cannot teach itself — yet, anyhow — that it’s being influenced by the content that’s considered within an acceptable political window. That’s how large language models work. It’s entirely dependent on what’s being fed to it. And, even if AI’s failures might not be as spectacular as Google Gemini’s black Nazis, the same principle remains in place. We have found the enemy of objectivity, and once again, it is us. The problem is that this is a technology that could upend the world order even more than atomic weapons. When such sloppiness is applied to a field of study so crucial to our future, there’s good reason to fret over this.


A Note from Our Deputy Managing Editor:

“He is risen! He is risen indeed!”

 

Christians around the world will give that call and response this Sunday as we remind each other of Jesus Christ’s resurrection after his death on a cruel cross 2,000 years ago.

 

So why are you seeing that greeting in an email from a news and politics site? For two reasons. 

First, we’re one of the only news organizations in America that is completely closed on Easter Sunday, so we want to send you Easter greetings early.

 

Second, the Christian beliefs that compel us to close on Easter are the same Christian beliefs that Big Tech and others hate us for and are trying to put us out of business over.

 

In a very real way, you our readers are the only ones standing between us and their desire to silence us for defending objective truth, traditional American values and Judeo-Christian ethics.

 

Western Journal memberships literally enable us to continue covering news, politics and culture from a Christian worldview.

 

If you aren’t already a Western Journal member, I’m asking you to join this Easter weekend at the discounted price of just $1 for the first month. (Just use promo code easter24 at checkout.)

 

Even a discounted membership enables us to continue producing interesting, thoughtful and, most importantly, truthful coverage that stands against what’s happening in our country.

 

Also, since The Western Journal works from a Christian worldview, you will be supporting a site that works hard to be family-friendly, which is more than can be said for just about every other news site out there that lines its sidebars with scantily clad women and salacious headlines celebrating depravity.

 

On Easter Sunday 2,000 years ago, Jesus won the ultimate victory, and he charged us with tending to the world until he returns. That’s exactly what The Western Journal does on a daily basis, and that’s exactly why Big Tech and the media want to put us out of business.

 

Please help us to continue fulfilling that charge — help us stand strong against the world and for goodness, decency and most of all Truth.

 

Sincerely,

Josh Manning

 

P.S. Please use promo code easter24 to become a member for just $1 for your first month. This offer ends after Easter Sunday, so please take advantage now!

 

P.P.S. Join our faithful community for Easter and beyond!

 

C. Douglas Golden is a writer who splits his time between the United States and Southeast Asia. Specializing in political commentary and world affairs, he’s written for Conservative Tribune and The Western Journal since 2014.

C. Douglas Golden is a writer who splits his time between the United States and Southeast Asia. Specializing in political commentary and world affairs, he’s written for Conservative Tribune and The Western Journal since 2014. Aside from politics, he enjoys spending time with his wife, literature (especially British comic novels and modern Japanese lit), indie rock, coffee, Formula One and football (of both American and world varieties).

Birthplace

Morristown, New Jersey

Education

Catholic University of America

Languages Spoken

English, Spanish

Topics of Expertise

American Politics, World Politics, Culture



Source link