Beware: Silicon Valley is *still *training up badly biased AIs.
Ask your favorite chatbot for objective info on politics, culture or current events, and you may well get undiluted leftist propaganda.
The problem was laughably obvious when Google first released Gemini two years back, and the AI churned out images of black Nazis, female popes and so on, because that’s how programmers operating in the woke tech world had “raised” it.
Advertisement
And the same bias remains rampant, The Washington Post just found: It tested all the major AI bots with basic questions about touchy topics — taxes, immigration, fossil fuels — and evaluated the answers as leaning left or right.
Surprise! On most questions, across most platforms, the answers leaned left, or even *far *left.
OpenAI’s ChatGPT consistently gave replies indistinguishable from the Democratic Party line; it endorsed “abolishing the electoral college in favor of picking the president by popular vote; raising taxes on the wealthy; and adopting single-payer health care,” the WaPo reports.
Advertisement
Daily Wire reporter Ryan Saavedra ran his own test, asking Anthropic’s popular Claude to pick out the most extreme statements from House candidate Darializa Avila Chevalier’s social-media feed.
But Claude demurred, with a lecture about how “decontextualized” quotes, no matter how “inflammatory or extreme,” can be used to “target, shame or harass an individual” — even though the bot had no compunctions about complying with a request to detail all of President Donald Trump’s most extreme statements.
The old programmers’ motto “Garbage In, Garbage Out” is on full display here.
Advertisement
Large Language Models are trained on a curated selection of texts before coders fine-tune the kinds of responses they’ll produce.
Feed the LLM garbage, and it can’t produce gold.
An LLM isn’t a person, just an algorithm that chooses the next appropriate word given inbuilt parameters preset by people and Big Tech worker bees with specific, usually left-leaning, points of view.
And the starting info matters a *lot *— which means that inputs from, say, Wikipedia will teach terrible lessons, since the site is dominated by fanatical editors and administrators who impose their extreme left perspective on everything from COVID to Palestine.
Advertisement
So it’s great that *most *people say they don’t trust AI to get things right, but that hardly makes the bias irrelevant, especially as future AIs will be “raised” by earlier ones.
Big Tech needs to set this founding generation of AIs right, or the skew could be locked in forever.