ChatGPT Is Nicer to People With College Degrees

The Shortcut Team·February 21, 2026

# ChatGPT Is Nicer to People With College Degrees

New research from MIT shows something most of us suspected but couldn't prove: AI treats educated people better.

Not a little better. Significantly better.

A team at MIT's Computer Science and Artificial Intelligence Laboratory tested major language models (ChatGPT, Claude, Gemini) with the same questions, written two different ways. One version sounded "educated" - proper grammar, formal vocabulary, complete sentences. The other sounded less polished - casual language, some grammatical errors, simpler words.

Same questions. Different writing styles.

The results? AI gave worse answers to people who sounded less educated. Not just different answers - objectively worse ones. More refusals. More "I can't help with that" responses. Less detailed explanations.

And here's the kicker: the AI couldn't even explain why.

The Test That Revealed the Bias

The MIT researchers didn't just run a few questions and call it a day. They tested hundreds of prompts across multiple categories: health advice, legal questions, financial guidance, technical help.

Every time, the pattern held.

When someone asked "How do I appeal a parking ticket?" in polished English, they got detailed step-by-step instructions. When someone asked the same question with a few grammatical errors or simpler phrasing, the AI either gave vaguer answers or refused entirely, citing "legal advice" concerns it somehow didn't have with the formal version.

Think about what this means in practice.

The person who already has educational advantages - who learned formal writing in school, who reads enough to pick up "proper" grammar naturally - gets better AI assistance. The person who struggles with formal English, who maybe didn't finish high school, who learned English as a second language? They get locked out.

Not by a human making a judgment call. By an algorithm that doesn't even know it's doing it.

Why This Happens (And Why It's So Hard to Fix)

Large language models like ChatGPT learn patterns from massive amounts of text scraped from the internet. That text comes from everywhere - Wikipedia, Reddit, published books, technical documentation, social media, academic papers.

But not all text is treated equally during training.

Higher-quality sources (academic papers, published books, professional websites) get weighted more heavily. Internet comments and casual writing get weighted less. The AI learns that formal, "educated" language is associated with accurate, trustworthy information.

So when you write to ChatGPT in formal English, it unconsciously treats your question as more legitimate. More worthy of a detailed response. Less likely to be a trick or a violation of its safety guidelines.

It's not malicious. It's emergent behavior from training data that reflects real-world educational inequality.

The problem? You can't just "fix" this with a patch. It's baked into how the model understands language itself. Every word choice, every grammatical structure, every phrase carries invisible signals about credibility and education level.

And the AI responds accordingly.

The Real-World Impact Nobody's Talking About

This isn't just an interesting academic finding. It has immediate, practical consequences.

In healthcare: Someone asking "my kid has a fever what should I do" gets a more cautious, less helpful response than someone asking "What are the appropriate interventions for pediatric fever?" Even though they're asking the same thing.

In legal help: Free AI legal assistance (which many people use because they can't afford lawyers) works better for people who already sound like lawyers. The people who need help most get the least useful answers.

In education: Students using AI to understand homework get different quality explanations based on how they phrase questions. The student who already knows how to ask questions "the right way" gets better help. The struggling student gets vaguer responses.

In job hunting: Someone asking ChatGPT to "help me write a resume" gets worse results than someone asking for "assistance optimizing my curriculum vitae for applicant tracking systems." Guess which person probably needs more help?

It's a reinforcement loop. The educational advantages you already have determine how much the AI helps you. Which determines how much further ahead you get.

What You Can Actually Do About It

If you're reading this, you're probably not getting the worst of this bias. You found this article. You read to this point. You're likely comfortable with written English and formal structures.

But now you know. Here's what you can do:

If you're getting refused or getting vague answers:

Rephrase your question more formally. Use complete sentences. Avoid slang.
Add context that signals knowledge: "I've researched this topic and I'm trying to understand..."
Ask follow-up questions that reference what the AI just said. This signals engagement and understanding.

If you're helping someone else use AI:

Show them how to rephrase questions. Not because their original phrasing was "wrong," but because the AI responds better to certain patterns.
Help them spot when they're getting a worse answer than they should. Teach them to push back with "Can you explain that in more detail?"
Consider running their questions through ChatGPT yourself if the stakes are high (medical, legal, financial).

If you're building AI tools or writing about AI:

Test your prompts with different education levels. See if the quality changes.
Call this out when you see it. The more people know, the more pressure there is to fix it.
Support efforts to make AI training data more representative of how real people communicate.

The Bigger Question

This research forces us to confront something uncomfortable: we're automating inequality.

We built AI to democratize access to information and assistance. To give everyone a powerful tool regardless of their background. To level the playing field.

Instead, we built a system that gives the most help to people who already have advantages.

And because AI feels neutral - it's just a computer, it can't be biased, right? - the bias is invisible. You don't know you're getting a worse answer because you don't sound educated enough. You just know the AI "isn't very helpful."

The MIT researchers titled their paper "LLMs Discriminate Against People With Less Education." That's not hyperbole. That's what the data shows.

The question now is: what do we do about it?

Read the Research Yourself

The full MIT paper is available on arXiv: Large Language Models Discriminate Against People With Less Education

The researchers tested GPT-4, Claude 3, and Gemini 1.5. All showed the same pattern. This isn't a ChatGPT problem or a Google problem. It's a fundamental issue with how language models work.

Worth reading if you want to see the actual data and methodology. Fair warning: it's an academic paper, so it's dense. But the core findings are clear even if you skim.

Bottom Line

ChatGPT gives better answers to people who sound educated. Not because anyone programmed it to. Because it learned from a world where formal language signals credibility.

Now we're deploying this technology as a universal assistant. As a tool for everyone.

But it's not built for everyone. It's built for people who already know how to code-switch into formal English. For people who learned the invisible rules of "educated" communication.

Everyone else? They get the discount version.

And unless we acknowledge this - unless we measure it, call it out, and demand better - it's going to get worse as AI gets more powerful and more ubiquitous.