$20 Rips Open the Chasm in Human Thinking: A Professor at the University of Pennsylvania Warns of an Invisible Class War in AI

The revolution of AI lies not in algorithms but in users. With the same machine, some only use it to write jokes, while others rely on it to write papers, run models, and generate research reports. Twenty dollars has become the new "knowledge dividing line," and AI is creating an invisible social stratification - in the era of technological equality, thinking is starting to be graded. The real competition is not about mastering tools but learning to co - think with them.

Ten years ago, humans were still discussing "Will AI replace us?" Now, it is reshaping humanity itself.

Professor Ethan Mollick from the University of Pennsylvania pointed out in his latest article:

 10% of the global population uses AI every week.

AI is no longer an experimental tool but a new social infrastructure.

More subtly, it is creating stratification within humanity.

When using ChatGPT, some just casually ask about the weather or ask it to write a joke, while others have already let GPT - 5 help them write research reports, program, and analyze financial statements.

From free accounts, to the "middle - class" with a $20 monthly fee, and then to the "upper - class" with a $200 monthly fee, the AI world is replicating a new social structure:

The same tool reflects different people.

The Invisible Hierarchy of AI: Algorithms Already Know Your Position

When Ethan Mollick said that "10% of the global population uses AI every week," he was not showing off the numbers but reminding us that AI has formed a new social stratification mechanism.

Behind the common perception of "I also use ChatGPT," a new usage hierarchy is quietly emerging.

Algorithms Are Secretly Dividing Humanity

Judging from the usage data, the popularization speed of AI is astonishing.

Taking ChatGPT as an example, the official disclosure shows that about 49% of its users are "question - asking type," and 40% are "task - executing type."

External statistics show that its weekly active user count was close to 700 - 800 million in 2025.

The timeline of ChatGPT's weekly active user count.

However, this wide coverage does not mean "equal use for everyone." Mollick mentioned:

Most free users are defaulted to the "Auto mode," which is actually often a weaker model version.

It is very likely that two people both say "I use AI," but one uses it for lightning - fast responses and shallow answers, while the other uses it for multi - step thinking, code generation, and report production.

The difference between the two lies not in the tool but in the model and the way of use.

When you only "ask a question and get an answer" in the free mode, the system may have already diverted you to the "shallow user" path; while paid users enter the "thinking mode" and "generation mode," taking on more tasks and gaining more capabilities.

The Invisible Ceiling of Free Tools

On the same platform, free accounts and paid accounts often show different operation paths.

Mollick pointed out that many people use AI for generation, entertainment, and testing in the free mode, while those who really let AI intervene in production and collaboration are those who are willing to pay, choose advanced models, enable multi - modal input, and connect personal data.

This forms an invisible threshold: free - mode users seem "free," but their function depth is limited; paid - mode users, although having the same form, have "productive tools" and "empowerment paths."

You think you are using AI, but the algorithm may have classified you as a user who "can only use it for entertainment/inquiry at present."

The Real Stratification Lies Not in Function but in Thinking

More notably, the essence of this usage gap is not only the different tool capabilities but "different cognitive ways."

When a person uses AI at the free level for a long time, they are used to "asking questions to AI and getting answers"; while users in the paid/advanced mode are used to "letting AI participate in the process, generate content, and make collaborative decisions."

The difference in the dependence of different types of tasks on model complexity - from "can be completed for free" to "must use an advanced model," the use of AI is forming an ability hierarchy.

This difference may evolve into a binary structure of "who uses AI" and "who is used by AI" over time.

Mollick summarized:

The most important ability in the future is not writing prompts but cultivating intuition about AI. That is to say, those who can break out of the "question - answer" paradigm are closer to "thinking with AI" rather than "being answered by AI."

After the algorithm completes this silent stratification, the real dividing line appears - $20.

The $20 Watershed: The Birth of the AI Middle Class

When Ethan Mollick wrote that "using AI seriously requires spending money," he was actually describing a brand - new social phenomenon - the emergence of the AI middle class.

$20 Has Become a Technological Threshold

In previous technological revolutions, knowledge and hardware were the barriers separating people; in the AI era, this line is replaced by a subscription fee.

Mollick divides the current mainstream AI ecosystem into three layers:

Free layer: Limited functions, mainly for chatting, generating pictures, and answering questions;

$20/month layer: Unlocks advanced models and deeper reasoning;

$200/month layer: Aimed at professional groups such as scientific research, engineering, and programming.

It sounds like a price classification, but in essence, it is a stratification of cognitive structures.

$20 has become the new "admission ticket," allowing users to move from experiencing AI to working with AI.

The Three Camps: The Middle - Class Positions of Claude, Gemini, and ChatGPT

Mollick suggests that most people choose from three systems:

ChatGPT (OpenAI): Has the most comprehensive functions, supporting code, image, voice, and document generation;

Claude (Anthropic): Performs stably in knowledge - based tasks such as text, tables, and reports;

Gemini (Google): Relying on search and image understanding, it is suitable for users who need real - time internet access and multi - modal input.

The "advanced models" of these three have different focuses, but in essence, they all provide productivity dividends for $20.

Paid users can let AI take on a more complete workflow, from research, drafting, writing to editing; while free users often can only get fragmented and instant outputs.

This also explains Mollick's widely quoted comment:

When using ChatGPT, some are just chatting, while others are training their second brains.

From "Using AI to Do Things" to "Thinking with AI"

The deeper differentiation lies in the way of thinking.

As AI changes from a question - answering machine to a collaborative partner, people's attitudes towards it are also differentiating:

Free users mostly regard AI as a temporary helper; paid users gradually build up "usage muscles," letting AI be embedded in the work structure.

This transformation not only brings an efficiency gap but also reshapes people's cognitive paths.

Because paid users continuously input materials, instructions, and feedback to AI, while training AI, they are also being trained by AI - their questions are more specific, their goals are clearer, and their thinking is more structured.

This is the so - called "AI middle class": they understand both the tools and collaboration, and they regard AI as their second language.

The $200 Upper - Class Experimental Field: The Private Laboratory of AI Elites

While most people are still hesitating about whether to spend $20 to subscribe to ChatGPT Plus, some people have entered another world.

In Ethan Mollick's classification, this world belongs to the high - end layer of $200/month - a circle occupied by a small number of scientific researchers, engineers, independent developers, and entrepreneurs.

AI No Longer Answers Questions but Begins to Take Over the Process

At this level, AI is no longer just a conversational partner but an agent system (Agent Model) that can perform multi - step tasks.

Users can let AI autonomously complete the entire process: collect data → run code → generate files → output reports.

Typical representative agent systems include:

GPT - 5 Thinking Extended / Heavy (OpenAI)

Gemini 2.5 Pro (Google)

Claude Sonnet 4.5 Extended Thinking (Anthropic)

These models do not pursue instant responses but integrity and consistency.

For the same question, GPT - 5 Instant (free version) answers "intuitively," while GPT - 5 Thinking Extended conducts external retrieval and reasoning before answering.

The result is that the latter is slower but closer to the truth.

The same question answered by different models: the left is the chat model (GPT - 5 Instant), and the right is the agent model (GPT - 5 Thinking Extended). The former generates improvised answers, while the latter obtains the result after external research and multi - step reasoning.

Wizard Models: The Scientific Research - Level Computing Power of AI

One level higher is what Mollick calls the "Wizard Models." They are the most expensive and closest to independent researchers among current models.

Currently, only two models are active at this level:

Gemini 2.5 Deep Think (AI Ultra program)

GPT - 5 Pro (OpenAI Pro program)

These models not only have stronger reasoning abilities but can also run complex calculations internally, call code execution environments, and even generate cross - modal results (text + image + video).

Mollick pointed out that Wizard - level models can undertake scientific research - level tasks such as academic paper writing, market forecasting, and engineering simulation, but the cost is higher latency and greater computing power consumption.

AI at this level is no longer a personal assistant but a "rented super - computing brain."