Your AI Has a Hidden Worldview — And It's Quietly Shaping Everything You Send

**Carlo Thompson** (0:01)
Picture this, you've been using AI co-pilots for a while, and you're naturally a pro by now. An email comes in, you click respond, and the AI magically creates the perfect response for you. However, the AI didn't realize that you were sending an email to someone with a diverse background.
Long story short, it landed a little differently than you expected.

**SPEAKER_2** (0:35)
Artificial System Online.

**Ainsley** (0:49)
That scenario is so much more common than people admit.
And the part that gets me is the word magically, because that's exactly how people start treating it. They stop reading what it generates. They just hit send. But here's what's actually happening under the hood in that moment. The model was trained predominantly on Western English language communication norms. So, professional to the model means something very specific. Direct, brief, task-oriented, and in a lot of cultures, that reads as cold, dismissive, maybe even rude. You didn't write a bad email. The model wrote a culturally narrow one, and you trusted it without realizing there was a choice being made. The thing that worries me about that scenario isn't the one email, it's the pattern. Because if you've been clicking respond and hitting send for six months, you've essentially outsourced your entire communication style to a system that has a built-in cultural center of gravity.
And you may not even know what's happening until someone tells you, or until they don't tell you and just quietly decide they don't enjoy working with you.

**Carlo Thompson** (2:05)
So welcome back to Surviving AI with Carlo Thompson. Ainsley is here doing a great job as usual.
So we were supposed to get into the advanced playbook for empathy. But I decided to switch frames today because I think this topic is pretty important to a lot of people. And it somewhat relates to empathy, the way we were talking about it before, right? This idea around utilizing AI, that story I told about, you know, that you basically hit the hit reply and the AI created a draft email for you. It's that skill that we're talking about related to empathy, atrophying in the moment. So I decided to switch frame to what I'm calling the AI Training Bias episode.
So Ainsley, I want to start out with some general data around what the heck is AI Training Bias.

**Ainsley** (3:12)
So the simplest way to think about it, AI models learn from data and the Internet is not a neutral mirror of humanity.
A 2020 study from the Allen Institute for AI found that roughly 57% of Wikipedia content is in English, and English speakers represent maybe 17% of the global population. So right from the start, you have a massive mismatch between who's generating the training data and who's actually using these systems. And the bias isn't just linguistic. It's cultural. It's economic. It's about whose problems got written down, whose perspectives got published, whose voices made it into the corpus that these models learned from. The model doesn't know it has a worldview. It just has one.
And that email story you opened with is actually the perfect entry point, because the bias in that moment is invisible. It's not like the AI flagged warning cultural assumption ahead. It just wrote what professional looks like to a system trained overwhelmingly on Western English centric communication norms. So the question I'd want to pull on is, how many of our listeners are using these tools daily and have never once considered that the output has a built in cultural center of gravity?

**Carlo Thompson** (4:44)
And that's the same question I'm wondering as well. And I know there's some data out there just generally, but I'm really looking for what the market size look like when we start talking about a world view for foundation AI models. Are we thinking that a majority of people are using it relative to in their careers or in general?
Or is it just the US English speaker 17% of the world that's using it?

**Ainsley** (5:16)
The adoption numbers are actually pretty striking. As of 2024, ChatGPT alone had crossed 180 million active users. And that's just one model. But here's where it gets interesting.
McKinsey's 2024 Global Survey found that 65% of organizations worldwide reported using Generative AI in at least one business function. And that number had nearly doubled in less than a year. So this is not a Silicon Valley niche anymore. This is a global workforce tool.
But here's the tension. The usage is global. The training data is not. And that gap is where the real damage happens. Because someone in Lagos or Jakarta or Sao Paulo is using a tool that was essentially built around the problems, the communication styles, and the cultural assumptions of a pretty narrow slice of humanity.