Do You Speak ChatGPTese? Beyond Writing, AI Is Also Flattening The Way We Talk

HAMBURG — ChatGPT is changing the way we write, and that’s no big surprise. It has been proven in the realm of academic writing. Certain words that ChatGPT tends to use very frequently have also appeared far more often in scientific publications since the software became available. The question now is: Is our spoken language changing alongside our writing?

Scientists at the Max Planck Institute for Human Development, led by Hiromu Yakura, Levin Brinkmann, and Ezequiel Lopez-Lopez, have now provided evidence that it is. They analyzed 360,000 YouTube videos and 771,000 podcast episodes recorded before and after ChatGPT’s release to track the use of terms like “delve,” “examine” and “explore,” all of which appear more than average in ChatGPT outputs. We spoke to Yakura and Lopez-Lopez about their findings.

ZEIT: Have you noticed yourselves speaking differently since you began using ChatGPT and other AI tools?
HIROMU YAKURA: Yes, that’s exactly why I launched this research. English is my second language, but I use it often, and I really value the support ChatGPT gives me. When I read that scientists had started using the phrase “delve into” more often in their writing since ChatGPT became available, because the AI frequently uses that term, I realized I was doing the same thing, both in writing and in speech.

What was your reaction when you noticed?
HY: I was a bit embarrassed, honestly, because it was clear that I’d picked up language from an AI model. But as a researcher, I wanted to find out whether it was just me or if others were experiencing the same shift.

You analyzed academic lectures on YouTube and a wide range of podcasts — more than 700,000 hours in total. What did you discover?
HY: Since ChatGPT became available to the public, the use of the word “delve” has increased by 50% every year. Of course, ChatGPT didn’t invent the word, but before the model’s release, it didn’t seem to play such a big role. We also saw a sharp rise in other words that ChatGPT frequently uses, like “comprehend,” “underscore,” and “swift.”

You focused on the vocabulary used by scientists, who, to be fair, already have their own way of speaking. Why that group?
EZEQUIEL LOPEZ-LOPEZ: Because many scientists are not native English speakers. They often have to explain very complex ideas to one another in English, and AI helps with that. Our profession is especially exposed to this kind of influence.

EY: But we also looked at business podcasts and shows on educational topics, and we saw the same effect there. Even religious podcasts now use the word “delve” far more than before. The only area where we didn’t observe a change was in sports podcasts.

The training methods and datasets used by ChatGPT are completely opaque.

Why is that?
EY: ChatGPT is not equally used by all social groups. But there’s probably a kind of secondary reinforcement at play: lots of people have friends working in academia or business who use ChatGPT regularly. That means the language change reaches them too, indirectly. ChatGPT is not just changing the behavior of individuals: it’s influencing us as a group.

Do we know why ChatGPT tends to favor words like “delve” and “comprehend,” or punctuation like the em dash?
EY: We wish we could answer that, but the truth is, we don’t really know why these particular words are preferred. One theory is that early versions of ChatGPT were trained using feedback from English-speaking workers in Africa, especially Kenya. The patterns we’re noticing might simply reflect common phrases in Kenyan English. But the training methods and datasets used by ChatGPT are completely opaque, and that’s part of the problem.

📢 New Preprint!
Now it's known ChatGPT overuses words like 'delve' and 'adept.' This raises the possibility that through the use of ChatGPT, our language can also be infected. To explore this, we transcribed and analyzed 300k YouTube videos, as in this excerpt from our dataset. pic.twitter.com/K8SBQCCjVC

— Hiromu Yakura (@hiromu1996) September 4, 2024

Human language constantly evolves. Isn’t it normal for ChatGPT to leave some sort of trace?
ELL: That’s true. Take the pandemic, for instance: It introduced new terms like “lockdown” into everyday language. Even therapeutic terms are now used much more casually. But what’s striking about the changes ChatGPT is causing is that they involve structural words (adjectives, adverbs and connectors) not nouns with a specific meaning. These changes are more about the framework of language than the content.

So what sets the language shifts caused by chatbots apart from other types of change?
EY: The mechanism is different. Usually, a new noun enters the language through the media. But this is more subtle. ChatGPT’s frequent use of words like “delve” or “underscore” may be an unintended side effect of its training. Still, it raises the possibility that foreign or even hostile institutions could deliberately manipulate language models and thereby shape the language used by millions of people, and possibly steer public discourse.

ELL: What makes this different is that large language models function autonomously, we don’t know how they work, and they centralize a huge amount of power in the hands of just a few companies.

The next generation of AI gets trained on our current language.

EY: On top of that, the next generation of AI gets trained on our current language. If that language becomes more uniform because people copy the AI, then future training data will be poorer in quality. That could further accelerate the homogenization of language. We already know that diversity in training data is crucial for AI systems to function well. Until now, people thought that AI would become stale and less creative only if it interacted solely with other machines. But our study shows the same effect can occur when AI interacts with humans.

But is that really so dangerous? If the systems stop working, couldn’t we just stop using them and get our language back?
EY: Sure, that could happen. But the danger is we may not even notice the change. We’re too close to it.

Couldn’t the models just adapt and avoid using “delve” and the other overused words now that it’s been flagged?
EY: OpenAI is already trying to reduce how often those words show up. But it might also go the other way: People could start using those words less just to avoid sounding like ChatGPT. That could create a feedback loop that brings linguistic diversity back.

So what is it exactly that worries you?
ELL: What’s troubling isn’t just that humans adapt their behavior when they use AI. It’s that the power to influence this behavior lies with a handful of companies. And we don’t yet know how far-reaching the consequences could be. So far, we’ve only measured how often certain words are used. But language reflects culture and values. If the language changes, those could change, too.

But now that we know this, couldn’t AI help us promote diversity by encouraging us to reflect on how we speak?
ELL: It’s not that simple. By design, ChatGPT favors the average. That makes it harder to foster diversity. These systems are based on statistics: ChatGPT always tries to give the most likely correct answer. So anything that’s less statistically likely tends to get filtered out. And what’s left is just the most probable. That’s the core of the homogenization process. And people are becoming less active in contributing to language diversity, too. Many no longer use forums where ChatGPT once gathered training data. Instead of asking each other, they go straight to ChatGPT.

Without variation, we stop coming up with new ideas or solutions. That’s why we’re raising the alarm with our study.

So one day the machine might be left talking to itself.
EY: Providers like OpenAI are not prioritizing diverse outputs, even if they could. Technically, it is possible to program AI to offer less probable alternatives too. Maybe we’ll need incentives or institutional requirements to ensure that cultural and linguistic diversity is preserved as AI models evolve.

Since Darwin, we’ve known that variation is essential for evolution. It’s what allows us to find new paths. If AI flattens out variation, what does that mean for our cultural development?
EY: Exactly. Without variation, we stop coming up with new ideas or solutions. That’s why we’re raising the alarm with our study. But AI also has potential upsides. It can lower the barrier for people to create something new, and that can boost creativity — especially for beginners.

Back to communication. Researchers at Meta recently showed that two chatbots developed their own efficient language when allowed to talk to each other. It was based on human language, but no longer intelligible to us. Could chatbots help us create a more efficient way to speak?
EY: In a way, they already have. ChatGPT taught me the word “delve.” I didn’t know it before, and that created a gap in my ability to express myself. Now I find it really useful. So in that sense, yes, we can thank ChatGPT. If efficiency were all that mattered, maybe everyone should start using “delve.”

But what makes “delve” so efficient, exactly?
EY: I’m not a linguist, so I can’t give a strict definition of efficient communication. But I think one reason “delve” has caught on is that it captures what people want to say, sometimes better than they could before. In that way, ChatGPT helps us rediscover practical, underused words. But language is more than a tool for getting ideas across. It also reflects social structures. It sends signals about who we are.

When Siri was introduced on the iPhone over a decade ago, some researchers feared we’d all turn into little dictators because we kept shouting commands at it. But our speech habits stayed mostly the same. Isn’t the concern about AI altering how we communicate maybe a bit overblown?
EY: Unlike with Siri, we now actually have data to measure the effects. And this is qualitatively different: With ChatGPT, we’re having full conversations on every topic, sharing emotions too. These new AI voices sound far more human than Siri or Alexa ever did. And we’re forming deeper connections with them.

Do you still use “delve”?
EY: Yes, probably once a week.

LL: What’s driving me crazy now is the em dash. It’s suddenly everywhere. That might be the new “delve.”