Via the New York Times, here's an interesting look at the political/cultural orientation of various AI models that are publicly available and widely used. It comes from a paper by David Rozado, who did the testing.
On the left are the results from base, untrained AI engines made by OpenAI and Meta. They're pretty centrist.
On the right are the results from various commercial AIs that have been fully trained. These are the products you actually use. They all tend toward lefty libertarianism, but some more than others:
The model furthest to the left is Google's Gemini, which is no surprise. Anthropic's Claude is the closest to centrist of the best-known models. Both GPT-4 and Twitter's Grok are close, but slightly more lefty.
Rozado also showed that it's fairly easy to train an AI to be reliably liberal or conservative. You just have them read the appropriate books and magazines.
Aside from Google, the major chatbot models aren't especially ideological, but they're definitely not conservative. Is that because of their Silicon Valley heritage or just because our written culture leans a bit to the left? Or something else? My own guess is that it's partly both of those plus one other thing: all the models are explicitly trained to not be racist/sexist/hateful/etc. Conservatives may not want to hear this, but like it or not, that pushes them away from the modern right wing.
Reality has a well-know liberal bias.
Known.
It is known.
"all the models are trained to not be racist/sexist/hateful/etc."
Interesting. On might train models to be racist obviously. But the very fact shows how far from intelligence AI is. It should be called FI or II (Fake or Imitation).
It is that they are trained to avoid sexism and racism but also because they are trained on things like neo-Keynesian economics which does a pretty good job of describing the real world as opposed to Austrian economics which does not. Presumably they are also trained on things like real physics which predicts things like climate change as opposed to lobbyist based physics which doesn't. As was pointed out, if you feed them a diet of things that are demonstrably not true, then the models would lean right.
all the models are explicitly trained to not be racist/sexist/hateful/etc.
They're trained to be logical and rational, and right wing ideology is neither.
The challenge, chat bots really struggle/can't discern fact from fiction. My point, if I were to create robust webpages, with lots of links etc, that claimed you cure COVID by eating chocolate, I would likely influence the AI large language model.
IF someone was skillful, well capitalized and motivated, it seems likely one could give an AI model all sort of political bents...
Not sure how you could get results to judge them on orientation if they've not been trained.
I don’t believe that Google trained its AI any differently. What they did do was append text to users’ queries such that all the questions were effectively asking for more diversity or what have you. Do the same with the other AIs and you’d likely receive similar results.
most likely they rate training sites for truthfulness/reliability and a lot of right wing sites would get demoted
just thought i'd mention that based on that chart a better characterization would be that chatbots lean libertarian left
wingnuts are mostly upset that they *really don't like authoritarianism
When you're dreaming you might not know that you're dreaming, but when you're awake you definitely know that you're awake.
There is nothing in an AI that can be awake.
So, does this mean not only do we need a return of the fairness doctrine, but it needs to be extended to AI/chatbots? …
Please do not use "training" They are "fed".
What conservatives display is "group loyalty at any and all costs" (to an extent which completely ignores any objective reality), which is not something that exists in a fundamental way in physics or mathematics.
"Group loyalty at all costs" can obviously have survival value in many cases, but I'd be surprised if a computer program demonstrated it, without having it intentionally added to the program by its makers.
Incidentally, the concept of "group loyalty" demands that there be an in-group and an out-group, and with human nature being what it is, the out-group is demonized, scorned, and marked for elimination.
https://www.dailykos.com/stories/2024/3/13/2229203/--How-Come-Everything-the-Republican-Party-Stands-for-Involves-Other-People-Dying?
There's also more conspiratorialism on the right at the moment. That could impact the political lean, especially if a lot of the conspiracies that more liberal people believe are less obviously partisan (like thinking germ theory isn't real.)
Kevin is a brilliant guy. I’ve gained a lot from reading him over the years. But he still seems to have a hard time coming to grips with certain modern realities:
“My own guess is that it's partly both of those plus one other thing: all the models are explicitly trained to not be racist/sexist/hateful/etc. Conservatives may not want to hear this, but like it or not, that pushes them away from the modern right wing.”
I grew up more or less thinking exactly the same thing. But that was then. Such statements now sound increasingly anachronistic.
Are there still bigots on the Right? Duh. Of course there are. And yes that’s a problem. But it’s just one part of the problem.
The Left is the other part of the problem. Ibram Kendi style “anti racism” is explicit racism against whites, Asians, Jews, or any other group perceived as white-adjacent. And racist antiracism is now the administrative and pedagogical norm at most colleges and universities in our country.
If you don’t find that terrifying, you’re not paying attention. Just look at the explosion of jaw droppingly vile antisemitism on campus. Where do you think that came from?
Also, the idea that Gemini AI was trained NOT to be racist is not just wrong. That’s a downright Orwellian perversion of reality. Gemini AI was very carefully designed to be racist against white people. It quite literally tries to erase them from history.
We have problems with extremist bigots on both the Left and the Right. Update your narratives accordingly.
"Just look at the explosion of jaw droppingly vile antisemitism on campus. Where do you think that came from? "
Largely from being 18. College students have a long history of being stupid and intolerant. Duh. Brain development does not complete until the mid to late 20's. College students have every right to be idiots. They're Poor Things, children in adult bodies.
MAGAs tooling around on their scooters, not so much. They have no excuse.
@ pjcamp1905:
From Michael Mandelbaum @ American Purpose:
“The responsibility for this dreadful state of affairs does not, however, fall entirely on the students. Most of them arrive on campus uninformed. It is the radical faculty who devote their classes to disseminating antisemitic teachings—and the feckless administrators who enable the demonization of Jews that these faculty members preach and the harassment of Jewish students that that preaching inspires—who are the 21st-century equivalents of the communist propagandists of the Cold War. Whatever else may be said of these people, they are not idiots–they know full well what they are doing; and when it comes to nurturing the values that institutions of higher education were established to promote and without which democracy cannot flourish, they are not useful. They are not even useless. They are extraordinarily pernicious.”
From David French @ the NYT:
“At the time that I read accounts of the violence at Berkeley, I was also reading two legal complaints about antisemitism on elite college campuses, one filed against Harvard and the other against the Massachusetts Institute of Technology. Both complaints make for horrifying reading, detailing a cascading series of antisemitic incidents, including acts of violence and physical intimidation. The moral and legal injury is compounded by blatant double standards on the part of the universities. As the complaint filed against Harvard states, ‘Harvard permits students and faculty to advocate, without consequence, the murder of Jews and the destruction of Israel, the only Jewish country in the world. Meanwhile, Harvard requires students to take a training class that warns that they will be disciplined if they engage in sizeism, fatphobia, racism, transphobia, or other disfavored behavior.’”
"‘Harvard permits students and faculty to advocate, without consequence, the murder of Jews"
Citation needed. And not of where this claim came from, but some examples of advocating for the murder of Jews (no, 'from the river to the sea' is not that)
Leo, it seems to have escaped your notice that Kevin Drum, and many of his readers, are quantitatively oriented. One of the quotes that pops up occasionally at the top of the page is 'Without data, you’re just another person with an opinion' — W. Edwards Deming. Ibram Kendi* and the alleged campus anti-semites** are a tiny and mostly inconsequential minority. If you have numbers to substantiate an ‘explosion’ of anti-semtism, or racism becoming a ‘norm’, show ‘em.
*I think I’m to the left of Democratic voters — I favor living-wage legislation, a wealth tax, universal health-care coverage, and similar — yet I know nothing about Ibram X. Kendi except his name.
**Mostly ‘guilty’ of chanting slogans; meanwhile, a six-year-old Palestinian boy was murdered, his mother attacked, three Palestinian college students shot. Have there been Congressional hearings on anti-Arab bigotry?
@ KenSchulz:
From the ADL:
“Since the Hamas massacre in Israel on Oct. 7, U.S. antisemitic incidents reached the highest number of incidents during any two-month period since ADL (the Anti-Defamation League) began tracking in 1979, according to preliminary data released today.”
From the WP:
“As one of a number of left-wing commentators who have been critical of mainstream anti-racism — and who believe the movement is little more than self-help for White people that runs interference for corporations and wealthy universities — I’ve watched the Kendi crisis unfold with a touch of schadenfreude. Yet though this public reckoning feels long overdue, I can’t help but also have a smidgen of empathy for the embattled anti-racism guru. Kendi was transformed from a respected historian — winner of the National Book Award for his 2016 tome, ‘Stamped From the Beginning,’ but hardly a household name — to the head sage of a global progressive movement in the wake of George Floyd’s murder. No longer a mere ambassador for academic anti-racism, Kendi became a brand.
The prospect of Kendi’s unraveling is not — or at least, is not only — the story of a huckster who was happy to cash in on America’s racial trauma, slapping his name on strange children’s books, including ‘Antiracist Baby’ and ‘Goodnight Racism,’ while raking in hundreds of dollars a minute to give short talks at American universities. Instead, the Kendi affair is yet another example of an age-old truism: White American elites on both sides of the political spectrum — academics, publishers, members of the media, corporate leaders — are always waiting in the wings to turn a shiny new Black intellectual into a mouthpiece for their political agenda.”
It is really unclear to me how you think your second quote answers a request for data. As to the first citation, I think the ADL counts anti-Isreali attitudes as anti-semitism and I find that completely wrong-headed. ie "there are reasons to be skeptical of the ADL’s methodology. Counting antisemitic incidents—and hate crimes in general—is an inexact science, which makes it hard to compare numbers over multiple years. Furthermore, the ADL has frequently included incidents of anti-Zionist political expression in its antisemitism tally, and has recently announced an intent to zealously oppose anti-Zionist politics, making it difficult to trust the organization’s judgment and motives as it relates to its reports." from https://jewishcurrents.org/the-adls-antisemitism-findings-explained almost a year ago. Since Oct 7 they have really doubled down on that point of view.
Do you have another source? I would frankly expect to see an increase in anto-jewish racism along with the increase I think I am seeing in anti palestinian racism, plenty of example from this blog's commenters. Personally, I unequivicolly condem both.
@ Coby Beck:
Read more carefully.
The quote about Kendi relates both to my original post and to the comment that KenSchulz makes on that topic.
I have also already posted commentary above referencing numerous lawsuits that document campus antisemitism in detail.
If you do in fact have any interest in and receptivity towards reality, the rest is up to you.
I have looked through your comments up-thread and don't see any references as you describe. If you have posted them elsewhere I don't think it is fair to expect me to go searching for them. And I would note that a reference to a lawsuit about something is not really a reference, but granted some lawsuits contain rigorous citations.
The ADL data for 2023 shows the following after stripping out anti-LGBTQ+ incidents (which they also track):
7338 incidents where the ideology is "Right Wing (White Supremacist)"
5 incidents where the ideology is "Right Wing (Other)"
1 incident where the ideology is "Right Wing (Anti-Government)"
1 incident where the ideology is "Left Wing"
1 incident where the ideology is "Islamist"
4431 incidents where the ideology is unknown.
You claim that “bigots on the Right” are “just one part of the problem,” but the data source you cite suggests that they are almost all of the problem.
I'd go further than that. If you are concerned with creating an LLM that doesn't lie or hallucinate, that is necessarily going to make it non-conservative.
Given the nature of most publically available training data, by default the models generally end up incredibly racist. Then the models are coerced to prevent that. From what I understand from people who would know, that takes a lot of work and the tendency is to over compensate a bit in order to make sure the model doesn't veer back to its uncorrected version.