We've spent a trillion dollars teaching AI to be the worst kind of student, then wondered why nobody wants to consume what it produces.
Part I: The Output Problem (Or: Why Everyone Hates AI Content)
YouTube viewers can smell AI narration from the first sentence. If the visuals are AI-generated too? Instant click-away. Corporations pump out AI advertisements and watch them get ratio'd into oblivion by comments. The feedback is everywhere, loud and consistent: people hate AI slop.
Yet creators, producers, and corporations keep shoving everything into the AI machine anyway. They think we're gullible. They think we won't notice. They think if they just produce enough of it, we'll eventually accept it.
The CEOs and VCs? They're "seeing bright." The metrics look good on paper. Cost per video is down 90%. Production velocity is up 10x. But there's a problem: the people actually consuming the content are leaving.
This isn't a technical problem. The models can generate coherent sentences, realistic images, and smooth voiceovers. The problem is that nobody wants it. And this disconnect—between what the spreadsheets say and what humans actually value—is the foundation of a trillion-dollar delusion.
Part II: The Economics Are Broken (The B2B2C Catastrophe)
Let's trace the money:
- AI Startup sells tools to Business (B2B) ✅
- Business uses AI to make content for Consumers (B2C) ✅
- Consumers reject the output and leave ❌
The entire AI content economy is built on a chain where the final link—the actual human on the receiving end—is broken. But the money keeps flowing because the people making purchase decisions are insulated from the consumer response by 2-3 layers of "adoption metrics."
The AI startup shows revenue growth (B2B sales!).
The business shows cost savings (efficiency!).
The VCs see "AI adoption" (line go up!).
The actual humans click away, disengage, leave negative feedback.
It's like if a factory produced a product nobody wants, but everyone in the supply chain from raw materials to retail could show positive metrics—except the actual customer never buys it.
The Free Tier Rotation Problem
Here's the other side of the broken economics. Most AI tools offer "generous" free tiers, hoping users will hit limits and convert to paid subscriptions.
But users just... rotate:
- ChatGPT free → Claude free → Gemini free → Perplexity free → back to ChatGPT (it reset!)
- Bounce between 10 different AI wrapper apps
- Never pay a cent
So the AI SaaS company is:
- Paying OpenAI/Anthropic for API calls (real cost)
- Getting $0 revenue from free tier hoppers
- Burning VC money to subsidize free users
- Praying for conversion that never comes
The conversion rate is abysmal because why would anyone pay when there are 50 competitors offering the same thing for free, and you can just rotate between them?
The Real Unit Economics
Let's do the napkin math on what an AI subscription should cost without VC subsidies:
1. Your Compute Cost: ~$7.50/month
(Based on 50 queries/day, 1.5M tokens/month at $5/1M tokens)
2. The Free Rider Tax: ~$25/month
(For every paying user, 50-100 free users exist. You're subsidizing their compute at ~$0.50 each)
3. The CAPEX Mortgage: ~$416/month
(OpenAI/Anthropic spend ~$50B/year on infrastructure ÷ ~10M paying users)
TRUE FAIR PRICE: ~$448.50/month
CURRENT PRICE: $20/month
VC SUBSIDY: 95%
It's a trillion-dollar gaslighting campaign where VCs subsidize losses to create the illusion of a sustainable business model.
Part III: The "Cost Cutting" Paradox (Firing Your Own Customers)
Here's where it gets darker. Every previous tech disruption redistributed money while keeping it in human hands:
- YouTube/TikTok → Created the influencer economy, video editors, thumbnail designers
- Amazon → Killed malls but spawned dropshippers, FBA sellers, logistics workers
- Uber → Created gig economy (exploitative, but still jobs)
- Etsy/Shopify → Empowered artisans, makers, small businesses
AI cost-cutting is different. It extracts money from the economy:
- Fire the voice actor → Money goes to AI company shareholders
- Fire the graphic designer → Money goes to Midjourney subscription
- Fire the customer service reps → Money goes to AI vendor
- Fire the copywriters → Money goes to OpenAI API calls
The money flows from "distributed among workers" to "concentrated in AI company equity." And here's the kicker: If you fire everyone to cut costs, you're also firing your own customer base.
The YouTube viewer who lost their job can't afford your product. The laid-off copywriter isn't subscribing to your service. Every corporation is racing to be first through the "efficiency" door without realizing that if everyone does it, there are no customers left with money to spend.
It's capitalism eating itself, optimized by spreadsheets that can't see beyond the next quarter.
Part IV: The Supply Side Delusion (Building Faster Toward Nothing)
When barriers to entry drop to zero, you don't get a gold rush—you get a noise competition.
- Pre-AI: Hard to build apps → Only serious ideas got made → Manageable signal-to-noise
- Post-AI: Anyone can ship in a weekend → 1,000 identical SaaS apps per day → Who can even find the good ones?
Developers can now "wing a SaaS app in a single weekend." Cool. So can 1,000 other people. You're innovation #457 for the week, globally. Nobody cares. Nobody's downloading. The App Store is a landfill.
The whole "move fast and ship" mentality made sense when you were discovering what people wanted. Now it's "move fast and build shit nobody asked for, using AI nobody trusts, to hit metrics nobody benefits from."
What's the point of churning out more code and delivering apps faster if the apps themselves are just crickets? Mostly business apps with an "AI inside" sticker slapped on.
The Calculator Salesman Problem
Imagine bringing a calculator into a fresh market and asking people, "Hey, want to calculate something? 2 cents per math problem!"
Not gonna happen.
The problem has to exist first and be painful enough that someone will pay to solve it. But AI apps are solution-first:
- "I have this cool AI hammer"
- "Let me find nails to hit"
- "Actually let me just hit everything and see what sticks"
- crickets
And here's the real issue: If someone needs a calculator badly enough, they already built one themselves. The gap between "problem exists" and "vendor shows up" is filled by people just figuring it out. Especially now when any in-house vibe coder can use AI to whip something up in 20 minutes.
You're not competing with other vendors. You're competing with "eh, we'll just handle it ourselves."
The market isn't small—the market is already solved. Every toilet-sitting idea has been crystallized into apps even before AI-assisted coding. The pie is gone. AI just lets more people fight over crumbs, faster.
It's like Chinese factories manufacturing products with robots at incredible efficiency, then shipping the cargo into the Sahara desert. The production optimization is perfect. There's just nobody there to buy it.
Part V: The Straight-A Problem (Why AI Needs Taste, Not Just Truth)
We've solved General Intelligence (being able to do anything okay). We haven't solved Specific Taste (being able to do one thing beautifully).
As developers, we look at LLM outputs and feel a specific kind of fatigue. The code works, the summaries are accurate, and the tone is polite. But it feels... empty. It's the textual equivalent of stock photography.
The industry thinks the solution is "more parameters" or "higher temperature." They are wrong. The problem isn't the compute; it's the classroom.
The "Bad Teacher" Problem (RLHF)
Current models are trained using Reinforcement Learning from Human Feedback (RLHF). To scale this, companies hire thousands of raters to grade model outputs.
This setup accidentally mimics a bad high school teacher:
- The Metric: Correctness, Harmlessness, Helpfulness
- The Goal: Minimize errors
- The Vibe: The "Straight-A Student"
The Straight-A Student is terrified of being wrong. They follow the rubric exactly. They hit the word count. They repeat the textbook definition because they know it's safe.
The Result: We get a model that is the average of what 1,000 random people think is "good." We get regression to the mean.
Real "taste"—the kind that makes you actually want to read an essay—comes from the B+ student. The one who risks a weird metaphor, ignores the rubric to make a point, and maybe gets a fact slightly wrong but nails the feeling.
Current AI tries to please 100% of people, so it thrills 0% of them.
The "Temperature" Fallacy
Here is where most devs (and the industry) get it twisted. We assume that if we want "creativity" or "style," we just crank up the temperature parameter.
This is a category error.
- Temperature = Entropy (Randomness)
- Taste = Intent (Structure)
Let's look at a prompt: "Explain Newton's Laws using a daily occurrence."
The "Straight-A" Model (Low Temp)
"Newton's First Law states that an object in motion stays in motion. For example, a car driving down the street continues until the brakes are applied."
Verdict: Accurate. Boring. Zero soul.
The "High Temp" Model (Randomness)
"Sir Isaac's decrees are omnipresent. Consider the feline on the sill; it remains a statue of fur until the external force of a vacuum cleaner propels it into a frantic trajectory."
Verdict: This isn't style. This is just a thesaurus having a seizure.
The "High Taste" Model (Intent)
This is the model we don't have yet. This model understands the assignment on a structural level.
"Newton is the reason why, when the bus driver slams on the brakes, all the freshmen standing in the aisle fall onto the seniors sitting in the back. The bus stopped; our bodies didn't."
Do you see the difference?
The "High Taste" example isn't random. It made specific, high-intent choices:
- Setting: The school bus
- Social Dynamic: Freshmen vs. Seniors
- Physics: Applied correctly to the social dynamic
That is not entropy. That is a highly ordered set of decisions. It is the opposite of randomness.
The 30 Different Essays Problem
When you give 30 high school students the same prompt—"Write about Newton's Laws in daily life"—you get 30 different angles:
- One kid talks about feeling heavy after a big meal (inertia)
- Another writes about dropping their phone (gravity + regret)
- Someone makes it about basketball shots (projectile motion)
- Another does it as a conversation with Newton at Starbucks
That's not randomness. That's 30 different lenses on the same material.
Each student is:
- Drawing from their own experiences
- Connecting Newton to what THEY care about
- Choosing an angle that makes sense to THEM
- Optimizing for "does this answer convey understanding in an interesting way?"
The loss function isn't "be random" or "be correct." It's "synthesize the requirement with your perspective in a way that satisfies both."
Current AI temperature just makes it pick weirder words. It doesn't give it a perspective, taste, or judgment about what angle would be interesting.
Industry thinks taste/style = temperature = randomness. But that's fundamentally wrong.
It's the difference between:
- Randomness: Throwing paint at a canvas
- Style: Choosing which colors go where based on taste
Part VI: The Convergence (Why This All Matters)
Here's how it all connects:
- VCs pump trillions into AI companies based on "future potential"
- Companies fire workers to cut costs and show "AI adoption"
- Free tiers are subsidized at 95% loss to juice "user growth"
- AI tools flood the market with near-zero barriers to entry
- Content becomes slop because models optimize for correctness, not taste
- Consumers reject it and engagement craters
- But the metrics still look good to people 2-3 layers removed from actual users
- The cycle continues until... what?
We're in a trillion-dollar game of musical chairs where:
- Money flows upward (to AI companies and shareholders)
- Jobs flow outward (people get fired)
- Quality flows downward (content becomes slop)
- And everyone pretends this is "progress"
The gaslighting is that this is presented as inevitable. As if we have no choice. As if this is what "the future" demands.
But the future isn't predetermined. It's being chosen by people optimizing for quarterly metrics instead of human flourishing.
Part VII: What Comes Next?
The Optimistic Case: Models improve dramatically. Within 5-10 years (look at the AlexNet → GPT-4 trajectory in just 13 years), AI closes the taste gap. We train models on "would a human with taste actually want to read this?" instead of "is this correct and inoffensive?" The output becomes genuinely good. People stop resisting. The economics work out because the product is actually valuable.
The Realistic Case: We hit the wall. Consumer rejection continues. The B2B2C chain breaks when businesses realize their customer acquisition costs are skyrocketing because nobody wants AI slop. The free tier rotation continues and conversion rates stay abysmal. The VC subsidy runs out. Thousands of AI SaaS companies collapse. We're left with a few actual winners (the infrastructure layer: OpenAI, Anthropic, Nvidia) and a graveyard of wrappers.
The Cynical Case: The gaslighting works. People adapt. Just like we adapted to algorithmic feeds, just like we adapted to ads everywhere, we adapt to AI slop. Quality becomes a luxury good. The "authentic human-made" becomes a premium tier. Everyone else gets the slop. Wealth concentrates further. Jobs don't come back. The economy increasingly serves capital instead of people.
Conclusion: We're Optimizing for the Wrong Loss Function
The core issue isn't that AI isn't powerful enough. It's that we're pointing all that power at the wrong target.
We're optimizing for:
- Cost reduction instead of value creation
- Correctness instead of taste
- Scale instead of quality
- Metrics instead of human satisfaction
We've built incredibly sophisticated technology and used it to make everything... worse. Cheaper, yes. Faster, yes. But worse.
The question isn't "can AI do the thing?" The question is "should we be doing the thing this way?"
And right now, the answer from actual humans—the people who have to consume this content, use these products, live in this economy—is increasingly: No.
But the VCs, the CEOs, the startups? They're seeing bright. The metrics look good. The line goes up.
That's the trillion-dollar gaslighting.
What do you think? Are we training the soul out of AI in the name of efficiency? Is the resistance to AI content temporary or fundamental? Drop your thoughts below.
Top comments (0)