Lisa Guan's Innovation World

OpenAI vs. China’s Consumer AI Models

OpenAI crossed the $10B ARR mark by August 2025. Meanwhile in China, leading consumer LLM apps—Doubao, DeepSeek, Yuanbao, Tongyi—are still free. Not ¥20 a month. Not even ¥2.

There are plenty of reasons for this gap, but let’s zoom in on one that doesn’t get enough airtime: years of feed-first product design have rewired user behavior. When you build for passive consumption at massive scale, you capture attention—but you also weaken two muscles that paid AI relies on: the ability to search and the ability to ask good questions.

How “Don’t Make the User Think” Backfired

Since 2016, Chinese consumer apps optimized for “don’t make the user think.” They pushed their algorithms to the extreme in order to win more users and keep them hooked.

It worked—spectacularly. But now the boomerang has come back. AI is a tool that rewards thinking. When your core audience has been trained to avoid effort, the perceived value of AI drops. Monetization suffers.

Consequence #1: The Search Muscle Has Atrophied

If your daily internet routine is “open app → scroll → tap whatever the system hands you,” you stop practicing search literacy. You don’t refine keywords, compare results, or evaluate sources. Over time, people lose the instinct to ask precise questions.

That matters because large models shine when you do ask specific, layered questions. If you can’t, outputs feel generic. And if results feel generic, you won’t see value in the tool—certainly not enough to pull out a credit card. Try to charge and users churn.

Plain truth: weak question → shallow answer → “AI is meh” → no willingness to pay.

Consequence #2: The Rise of the “Feed-Brain”

Recommendation engines plus short video created a kind of passive-stimulus mode. Call it the “feed-brain”: high dopamine, low friction, zero contemplation. The feed-brain resists anything that takes more than a minute of focused thought. It bails at the first hint of effort.

Paid AI usually asks for the opposite. To get real value, you need to think through the problem, break it down, and express it in a structured prompt. That’s how you transfer intent to a model. But to a hardened feed-brain, even writing a clear prompt feels like homework. The defense mechanism kicks in: skip the thinking.

So will a feed-brain pay for an app that makes it think more? Not likely.

Why Subscriptions Struggle in China’s Consumer AI Market

Subscriptions thrive when users plan to use a tool to accomplish something specific—draft a contract, analyze a dataset, sketch a marketing plan. That’s intentional, goal-driven behavior. It requires a moment of setup: clarify the goal, supply context, choose a workflow. In return, you get leverage.

But feed habits bias toward impulse, not intent; toward reaction, not reflection. Until that shifts, consumer AI in China will keep bumping into the same wall: people won’t pay to think.

OpenAI can charge because many of its users arrive with intent and a habit of asking questions. Much of China’s consumer internet, by contrast, optimized for passive acceleration. That won the feed—but it starved the very skills that make paid AI feel magical.

What Will Work in Consumer AI in China (for Now)

Given this landscape, there are two consumer AI features that actually match today’s feed-era incentives. Neither demands deep thinking; both amplify what users already do.

1) The Mouthpiece: One-Tap Replies for Comment Wars

When users need to spar in the comments, an AI “mouthpiece” makes sense. It would:

• Read the thread and the topic at hand.
• Infer the user’s stance and preferred tone from their history.
• Generate a one-tap reply tailored to the user’s style (snarky, deadpan, data-driven, whatever).

This aligns with existing behavior (argue, dunk, react) and removes friction (no need to compose). It feels like power without requiring thought. Engagement would spike. Ethically, it’s a minefield—amplifying flame wars isn’t exactly civic uplift—but in pure product terms, it fits the feed.

2) The Comforter: Algorithmic Affirmation Engine

Another fit is the “psychological massager.” The idea is simple: mirror the user’s values back to them with context-sensitive affirmations. If someone frequently watches content that validates a certain worldview or personal struggle, the AI offers sympathetic, on-brand encouragement.

• Parent-content watcher? “You’re doing your best; it’s hard, and you care deeply.”
• Relationship-drama content? “It’s complicated, and responsibility isn’t one-sided.”
• Career-grind clips? “Your discipline today compounds into tomorrow’s options.”

Never ask user “Do you need to talk about the issue?”, whoever ask user to think or type, will loose.

3) Beyond the Feed-Brain: Possible Ways Out

There are ways forward, but they run upstream of the current habit stack:

• Teach prompting without saying “prompt.” Build lightweight wizards that feel like chat, but nudge users to add constraints, examples, and success criteria. Reward the extra detail with visibly better outputs.
• Bundle AI to real jobs-to-be-done. Tie the model to workflows where outcomes are tangible—tax prep, visa forms, lesson planning, small-business ops. When the win is concrete, payment follows.
• Shift from content consumption to content creation. Tools that help people produce—sell, teach, design, automate—create clear ROI. That’s the quickest bridge from “free” to “worth paying for.”

But unfortunately, I suspect the first two—the mouthpiece and the comforter—will be far more attractive to companies in the short run.

Let’s see.