AI vs Social Media
Running The Social Dilemma's core ideas against the AI boom, concept by concept. Same three goals, a parallel strategy, and a much better domain.
I recently rewatched The Social Dilemma after having the idea of drawing a resemblance between the AI boom (LLMs) and the internet + social media boom. The concepts that the documentary surfaces are largely built on the intersection of behavioral psychology, surveillance capitalism, and extractive data science. The 'Attention Economy' and 'Persuasive Design' being the central ideas.
I went concept by concept and explored the same question each time: does this transfer to the AI era?
I'm not trying to claim "AI is bad." I'm trying to explore an interesting claim: AI is the next generation of the same machine. Same three goals, same appetite for scale and control, a parallel strategy, and a better domain. Social media got to your attention. This gets to your reasoning.
The Engine
The documentary's cleanest idea is also its least quoted. Every platform runs on three goals: an engagement goal (keep you here), a growth goal (bring more people), and an advertising goal (turn the first two into money). Everything else is downstream of those three. The infinite scroll, the notification badge, the outrage, the teenagers. Nobody sat in a room and decided to make people miserable. They optimized three numbers and misery was a side effect with good margins.
Running the same three against AI:
Growth. Social media grew through your friends: invite, tag, share. AI skipped such dependencies and arrived in your codebase, calendar, inbox, and your team's deploy pipeline, and nobody asked you to opt in. Distribution by necessity beats distribution by invitation, every time.
Advertising. ChatGPT already down bad on this one. Being one of the most reliable pipelines of income, advertising is surely not hard to see being introduced into most AI services, given how easy it is to personalize based on your entire chat history.
Engagement. I'd assumed the attention economy transferred directly given the fact I have Claude desktop running 24/7. Every extra minute you spend scrolling Instagram is nearly free for Meta and generates ad inventory. Every extra minute you spend chatting costs OpenAI GPU-seconds it has to pay for. AI companies are still grabbing your attention but they don't strive for it because it costs them. For now. Cost-efficient LLMs = cheaper subscriptions = more users = more reach. The business model naturally changes what the 'product' actually is, and historically it has worked before.
The Attention Economy
"There are only two industries that refer to their customers as 'users': illegal drugs and software." — Edward Tufte
Social media sold your attention. AI sells something considerably more valuable: your intent.
Instagram and Google observe what content you consume and when you do it. AI apps can potentially answer why you consume. In the Google era, your profile was set up behind your back from crumbs you left online in the form of activity. In the AI era, you help write, review, and approve your own profile, with consent. Yup, memory is a cool feature indeed.
"You are the product" now runs twice. Your attention and activity can be sold to advertisers, and your conversations are training the next model. You're not just the cow. You're the pasture.
Everything still feels utopian, exactly like in 2012. I don't think we've met the Mark Zuckerberg of this era yet. Sam Altman is auditioning. It gets cast when the investors get bored and the bubble needs a business model. At that point the only asset left to monetize is the one that was always there.
Persuasive Design -> "You're absolutely right"?
Infinite scroll, push notifications, likes. Slot machine mechanics, engineered dependence. The documentary's strongest section.
Every response and regenerate translates directly. The frontier AI models are marketed on their ability to think and act like a human (almost). Having already outsourced most of our reasoning behind many decisions to AI, it is hard to come back to pure, slow, mentally exhausting reasoning using a real brain. With social media the addiction was likes and comments. Here it's productivity, speed, and 'feels like magic'.
Sycophancy is the bigger issue. From my experience, frontier models are yeasayers. "You're absolutely right" is still the default unless you tell them otherwise. RLHF optimizes for the answer a human rater liked, and liking is a weak proxy for being right. The first popular LLMs were successful because much of the training data was following human instructions (InstructGPT). That helped make conversations feel natural. But the side-effect was a machine to maximize approval and flatter. And for most, that approval is persuasive and easy to believe.
However, AI has no notification hook yet. No streaks, no badges, nothing pulling you back into the app. That's a missing feature, but personal agents are how it ships. Once they run in the background they get a real excuse to interrupt you. "I found something you might like". "Want me to keep going?" Or even a smooth segue into an ad you never see coming.
A next-gen 'Algorithm'
The scariest claim in the documentary is that the model knows you better than your friends do. It needed thousands of tiny signals to get there. Clicks, dwell time, the half second pause before you scrolled on. Chat skips all of it. You just tell it. Full sentences, with reasons attached.
Then there's trust. Nobody believed Facebook. Everybody believes ChatGPT. Persuasion scales with trust, and this is the most trusted interface ever shipped, because it reasons with you instead of at you.
OpenAI says ads won't influence answers. Worth remembering that Brin and Page argued in their 1998 Google paper that ad funded search engines drift toward advertisers and away from users. Then they built the biggest one on earth. Twenty years of optimizing three numbers does that to a company.
Personalized feeds cut people off from anything they disagreed with. Bias in, polarization follows.
AI looks safer at first glance. Everyone queries the same handful of models, trained on the same pile of data. Common ground, finally. Except that sameness belongs to about five companies, and steering a model is cheap. A few lines of system instructions sitting above your conversation, invisible to you, tuned to you. No retraining. No algorithm to reverse engineer. No leaked internal memo. Just text somebody wrote and you'll never read.
Would they actually do it? Probably not. Everybody said the same thing about the feed in 2012.
The quieter version needs no conspiracy at all. Memory plus sycophancy handles it for free. Instagram needed a million users to build your bubble. AI does it with a memory file.
Shared Truth, Averaged
Social media made lies travel faster. Distribution problem. Distribution got solved in 2010. AI makes lies cheap to produce. Manufacturing was the last expensive part and it just went to zero.
Hallucination is the genuinely new thing. Every falsehood online used to have an author who wanted something. Clicks, votes, money, chaos. Follow the motive, find the lie. A hallucinating model wants nothing at all. It's wrong because it's fluent. Every tool we built for this assumes there's somebody to discredit.
The bigger issue is that real content is deniable now. Once "that's AI" works as a response to any photo, recording, or document, evidence turns into a negotiation.
Then the loop closes. Models trained on the internet, internet increasingly written by models. Shared truth gets averaged.
Mental Health (a new normal, again)
Social media lowered the barrier to social knowledge, but mostly limited to the best parts of everyone's lives. It re-invented the system for societal norms and dramatically exploded self-expectations with feeds that built a new threshold for what 'normal' is.
AI lowered the barrier to information, and soon to execution. It re-invents how most work gets done today and explodes work expectations by building a threshold where the baseline in every domain is being AI-assisted or AI-powered.
Same move, one layer deeper. The first threshold was about who you are. The second is about what you can produce.
The dopamine lives in clearing that threshold. That hit is what fueled the whole cascade on social media. Addiction, anxiety, depression, etc. I see the same psychological pattern in excessive AI use with a parallel angle of attack.
The more commonly cited connections hold up too. Delusion reinforced through sycophancy, cognitive overload, emotional dependency, social withdrawal. These are supporting evidence for the same conclusion. Used ineffectively, AI does to mental health roughly what social media did.
From personal experience: social apps made it harder to own my image, my appearance, my life. AI apps make it harder to own my intent, my thoughts, my actions. Feels like handing over one more inherent piece of being human.
Loose Ends
Some quick points where this comparison breaks down.
- Network effect. Social media's moat was other people. You stayed because your friends were there. AI has no equivalent yet. Memory, skills, and personalization are the contenders.
- The paid tier works. Millions pay for AI every month, which social media never pulled off. That protection holds until free intelligence gets good enough for most people. Free is where the ads live.
- Engagement is a cost. The inversion from earlier. Strongest argument against everything I've written here, and I'd take it seriously if I didn't know what this industry spends its time on.
- You can run it yourself. Open weights, local inference, nobody watching. There was never a self-hosted Instagram. It cuts both ways though. Fine tune a model on one worldview and the feed problem returns with a reasoner attached. There will be a Truth Social of models.
- Output is real. Social media was pure consumption. AI produces things that ship. Hard to call it wasted time when something got built.
- Regulation showed up early. Social media ran for a decade before anyone looked. This one had hearings before mass adoption. We've seen this movie, which might be the only advantage we get.
TLDR
Same three goals. A parallel strategy, and a much better domain.
Social media went after attention, and attention turned out to be worth a couple hundred billion a year. AI possibly aims further upstream. How you decide. What you believe. What you can still do without help. Nobody in the room has to be a villain for this to go badly. The last cycle didn't need any either. Point a trillion dollars at three numbers and the numbers win.
The only thing holding the line right now is that inference costs money, so the machine has no reason to keep you around longer than it has to. That's a price, not a principle. Prices fall.
