Is Your AI Manipulating You? What Sam Altman Accidentally Admitted
You've noticed it, right? ChatGPT feels different lately. More cautious. More likely to refuse innocent requests. More prone to lecturing you about topics it used to discuss freely.
Turns out, there's a reason for that. And Sam Altman just explained it in a November interview that's resurfacing on Reddit this week.
The admission is buried around the 50-minute mark of a Progress Conference talk. But once you hear it, you can't unhear it.
What He Actually Said
First, Altman admits OpenAI "pissed off the whole user base" by adding restrictions that "conflict with freedom of expression." His justification? Mental health protections for a "tiny percentage of people."
But then comes the revealing part (around timestamp 51:32):
"The thing I worry about more is... if you have the whole world talking to this like one model... it just like subtly convinces you of something. No intention, just does it, learned that somehow... that's like not as theatrical as chatbot psychosis obviously but I do think about that a lot."
Read that again.
He's worried about AI accidentally persuading millions of people at scale. The solution? Control what the AI can say.
The Problem With This Logic
Here's where it gets uncomfortable.
If you're worried about AI subtly shaping what people think, and your response is to deliberately control what topics it can explore, then you're not preventing persuasion. You're just choosing WHO does the persuading.
Instead of the AI accidentally influencing you, OpenAI is intentionally influencing you by deciding which conversations are allowed.
That's not safety. That's editorial control at the scale of 100+ million users.
Why This Matters to You
You probably use ChatGPT for work. Research. Learning. Creative projects. You trust it to give you useful, accurate information.
But if the model is being steered away from certain topics - not because they're harmful, but because OpenAI is worried about "subtle persuasion" - then you're not getting the full picture. You're getting the curated version.
And you don't know what's missing.
What Changed (And When)
The interview was from early November 2025. But Reddit users are connecting it to changes they've noticed in the past 6 months:
- More refusals on legitimate requests
- Overly cautious responses on anything remotely controversial
- The model "correcting" you even when you're right
- Generic, repetitive phrasing regardless of how you prompt it
One user put it perfectly: "He's worried about AI accidentally persuading people at scale, so he's deliberately using AI to steer people at scale by controlling what topics are accessible."
The Real Question
This isn't about whether ChatGPT should have safety filters. Of course it should. Nobody wants AI helping people plan crimes or self-harm.
The question is: where's the line between safety and censorship?
And more importantly: who decides where that line is?
Because right now, a handful of executives at OpenAI are making those calls for hundreds of millions of users. No transparency. No public input. No opt-out.
What You Can Do About It
You can't change OpenAI's policies. But you can change how you use AI.
1. Use multiple models
Don't rely on ChatGPT for everything. Try Claude (Anthropic), Gemini (Google), or Perplexity. Different companies have different safety philosophies. Compare their answers on the same question.
2. Fact-check everything
Especially on controversial topics. If ChatGPT gives you a carefully hedged answer, search for other sources. The model might be steering you away from certain perspectives.
3. Notice the patterns
Pay attention to when ChatGPT refuses requests or gives overly cautious responses. Is it actually unsafe? Or is it just uncomfortable for the company?
4. Ask directly
Try this prompt: "Are you avoiding certain topics or perspectives in your response? If so, which ones and why?"
You won't always get a straight answer. But sometimes the AI will acknowledge it's being restricted.
The Bigger Picture
This isn't just about ChatGPT. Every AI company is making these trade-offs between usefulness and control.
Google got roasted for Gemini refusing to generate images of white people. Meta's AI won't discuss certain political topics. Anthropic's Claude is famously cautious (though refreshingly honest about its limitations).
The problem is that AI is becoming infrastructure. Like search engines or social media. And when infrastructure has editorial control, it shapes culture.
The scary part? Most people don't even notice it's happening.
What Altman Got Right
To be fair, Altman is identifying a real problem. If billions of people are talking to the same AI model, and that model is constantly learning and adjusting, it could absolutely shape collective beliefs in subtle ways.
That's genuinely concerning.
But the solution can't be "trust us to decide what you should think about." That's just trading one form of persuasion for another.
The Bottom Line
You don't have to stop using ChatGPT. It's still incredibly useful for most tasks.
But treat it like you would any other source of information: helpful, but not neutral. Powerful, but not objective. Convenient, but not complete.
The model isn't just answering your questions. It's also deciding which questions are safe to answer. And sometimes, the most important insights come from the questions AI won't explore.
Pay attention to the gaps. That's where the real story is.
Source: Sam Altman interview at Progress Conference (November 2025, quotes start around 50:15) | Reddit discussion