Today I read Ethan Mollick's *Choosing to Stay Human*. After finishing it, I sat down and thought about it for a long time. The more I thought about it, the more flaws I found; almost every argument I made could be challenged by myself. The first half of this piece reads like a book review, while the second half is more like a debate I had with myself. If my future self looks back, I hope these points of contention haven't been rationalized.
What is this article about?
Simply focus on the key points:
- AI is writing more and more things that look fluent but lack substance, like an "attention-stealing vampire."
- The cost of using AI to write code is the loss of the opportunity to develop your own style.
- A comparison of two experiments in Türkiye and Taipei: Treating AI as an "answer machine" leads to learning regression; treating AI as a "personalized tutor" results in learning improvement over 6-9 months.
- The phenomenon of "cognitive abandonment": People stop thinking and hand over all their thinking to AI, even if the AI is wrong.
- BCG consultant research: When faced with problems that AI cannot solve, humans using AI actually perform worse because they cannot identify the errors.
- Anthropic research: Programmers who fully outsourced their work to AI couldn't answer what they did; those who asked AI to explain, or only outsourced parts of the work, didn't degenerate.
- Conclusion: The key is to "consciously choose" what to delegate to AI and what to keep for yourself.
Why did this article hit me?
The opening paragraph, which directly addresses the topic of "social media is now flooded with AI content," immediately caught my eye. I think the English-speaking world may already be used to this, but the Chinese-speaking world is just beginning. Initially, those more sensitive to language probably felt slightly disgusted by AI output, but now I think even the average person feels uncomfortable.
One point in the article further confirmed why I felt uncomfortable: the content produced by AI lacked any human meaning; it was as if it was constantly consuming our attention without giving us any corresponding reward.
I think it might be necessary to distinguish the purpose of reading:
1. Good articles based on real experienceWe can learn new things from the perspectives of people who have actually experienced and gone through these events.
2. AI-generated long articleMost of them are probably fabricated stories, without any meaning given by the people behind them, and may even be generated by a fully automated process, making us feel like we are watching an empty story.
Some people originally wanted to learn on the community, but when they saw rows and rows of these AI-generated long articles, they felt they weren't getting anything in return, and even felt a bit disgusted.
My past writing experience has made me more vigilant.
Ethan mentioned that he had been very fortunate over the past decade or so of writing. I am also silently grateful, because I started to develop the habit of writing regularly about five or six years ago.
My writing habits, cultivated over time, prevent me from believing that AI can simply accelerate my writing skills. It's a bit like when we didn't have cell phones, we might stare at the words on a shower gel bottle when we were bored; even though we've grown up and have cell phones that easily eliminate boredom, we still vaguely remember those ways of passing the time.
However, I'm also afraid of falling into a framing: that people with past experience will naturally be alert, while those without won't. That's not actually the case. Many people ten years older than me are just as dependent on AI as young people; they also experienced the era without cell phones, but they haven't developed the same sensitivity as me. Conversely, there are also aware people in the younger generation; they don't have a control group, but they can sense "something's not right" in the moment.
So what's truly alarming is probably whether we've realized the value of comparison. My sensitivity to AI content might not simply stem from living in an era without smartphones, but more importantly, from the fact that I lived through it and am aware of what that era lost.
This distinction is quite important. Because "experience" cannot be copied to others, but "consciousness" can be guided. If I want to do community design, what I need to do is probably "create moments of awareness in the present moment," rather than just thinking about "reproducing the past." I'll write this down for now and come back to it later.
Using AI to assist in writing is fine, but it depends on how you use it.
I also use AI to assist in writing. After I dictate some content, I let it help me see if there are other examples or better materials to illustrate the points I want to make, and check if anything is missing.
The key is to ask yourself what your purpose is before you use it.
At first, I thought this "purpose" was simply a choice between "what I want vs. what AI wants." Later, I realized that "what I want vs. what AI wants" is merely a means to an end. The real purpose is a choice between three options:
- In order to learn → AI in the background (run the test yourself, then let AI find vulnerabilities)
- In order to complete → AI throughout (highest efficiency)
- In order to express myself → AI cannot draft; at most, it can polish it.
I used to constantly ask myself, "Am I abandoning cognition by doing this?" and "Am I using AI too much?" This is the consequence of mistaking the means for the end. The real question should be, "What is my goal today, and am I using the right methods?" Worrying about whether you've used too much or too little will only lead you astray.
If the goal is right, using AI, however much or however little, is correct. If the goal is wrong, using it, however little, is wrong.
AI should be the trigger that "pushes you back to thinking".
There's a passage in the article that really resonated with me: If you want to benefit from AI, you must change your approach. AI should become a catalyst for problem-solving, prompting you to think for yourself, rather than simply handing you the answer.
In other words,AI should be designed with a trigger point that "pushes you back to thinking"..
This made me reflect on how I used a lot of AI to help me solve problems in the past, whether it was for events or courses. In the process of dealing with these problems, I felt that I was gradually losing this "muscle" for problem-solving.
So I'm thinking that perhaps in the future, problem-solving skills will become more important, and AI should play a more crucial role in the problem-solving process.
I looked up the steps for Problem Solving in the middle of the process:
1. Define the problem
2. Disassembly Structure
3. Brainstorming solutions
4. Decision-making and actions
5. Review and Reflection
I've been thinking about what role AI will play in problem solving.
If our goal is learning, we must introduce obstacles and limitations at appropriate times; otherwise, we cannot achieve the desired learning effect. Therefore, I'm wondering if, in the process of problem solving, AI needs to implement more restrictions and verifications in the first step of "defining the problem"? Will "hypothesis and verification" become even more important in the future?
I try to go through the entire process myself as much as possible. Some parts of the role that AI plays can be "outsourced".
For example, I think that in diagnosing the root cause, we can further utilize AI to cover some directions we haven't thought of; we can use the "5 Whys" or MECE framework to find more possibilities.
However, in reality, the focus of "verification" and "solution confirmation" may lie in assessing which potential solutions can collaborate with AI. As for how to solve the fundamental problem, that is ultimately up to us to determine.
The likelihood of success for these things still rests on the premise that "you have actually done it yourself." Only because you have practiced it will you know whether the method can work, which echoes the earlier case about AI and consultants.
Here are a few points that can be incorporated into problem solving:
The blind spot of "problems that AI cannot solve".
The article mentions a point that puzzled me: Ethan asked the consultants to use AI to handle some problems that AI couldn't solve, but the group that used AI actually performed worse because they couldn't see where the AI went wrong.
But how do you know which problems AI can't solve? Are there any evaluation criteria?
If we can know in advance, it means we already have the judgment in this field; if we can't, you'll never know when you'll be misled by AI.
Therefore, I believe that "professional competence" will become even more important in the future. But if this isn't cultivated on the job, how exactly can it be done? And if AI is heavily used for implementation and collaboration, will it actually worsen the situation?
I don't have the answer to this question yet. I'll write it down first.
System design can protect the user from harm.
The results of the experiment in Taipei provide a direction, namely that restrictions should be imposed at the system level, rather than relying on the willpower of users.
However, we rarely see this approach in consumer products; business pressures often drive things in the opposite direction. All AI companies are vying to "reduce friction" because friction equals attrition. But friction is necessary in learning scenarios; without friction, there is no learning.
This means:Products that are "deliberately designed to create friction" may be the next opportunity..
Here are some ideas for using AI to help with brainstorming:
- Forced breakpointThe process includes a "stop and decide your next step" segment (this is part of the Claude Plan).
- Progress bar visualizationDraw out the thought process so that users can see which section of the thought process the AI is taking over from.
- Write first, then read.You must enter your thoughts of at least 200 characters to unlock the AI response.
- Explanation buttonNext to the AI's answer, there are always two buttons: "Explain it to me" and "Force me to think for myself".
I'm considering whether the system design can encompass and organize the following key issues together:
1. Define the problem
2. Objectives
3. Restrictions
4. Verification
Ultimately, these four items are less like features of a product and more like a checklist I want to run before I use AI every time. It's just that maintaining this level of self-discipline is really difficult for one person. If it's so difficult, wouldn't it be easier for a group of people?
What can I, as a member of the tribe, do?
Let's go back to the question we left earlier: "Consciousness" can be guided, but how?
As someone who's currently building a community, this question is particularly crucial to me. What I've always wanted is a spontaneous community, but writing this has made it clear that spontaneity itself cannot be designed.
Truly spontaneous people don't need my design; they'll find me naturally. My community is already doing this, but we just didn't see it clearly before. For those who haven't yet experienced comparison and therefore haven't developed spontaneous awareness, there are actually two things I can do: First, "catalyze comparative experience," for example, by organizing a week-long challenge for writing without AI, forcing those who haven't experienced "what it feels like to write for themselves" to have a taste of it; second, "replace personal will with community pressure," as the presence of the masses is more reliable than personal will, which connects to the conclusion of the Taipei experiment: system constraints are more reliable than personal will.
So the core paradox is this: I want a "spontaneous community," but all I can do is "design a community that can also be driven by non-spontaneous people."
Writing this also makes me realize: Could I be treating "designing a self-spontaneous system" as a new ritual and keep postponing my actions?
Conclusion: Four principles for your future self
I feel my thoughts are getting a bit jumbled as I write this, so I'll summarize in four sentences for my future self:
1. The purpose determines the method of use.The three purposes of completing, learning, and expressing correspond to three different uses. Instead of asking, "Have I been using AI too much?", ask, "What is my purpose today?"
2. Leave the process to humans, give the draft to AI.The problem-solving process cannot be outsourced, but the outputs can.
3. We don't outsource things we haven't done before.You can't even see your own mistakes.
4. Determine if alignment is requiredPersonal judgment can be off-target; you need to chat with people and discuss in online communities to stay on track.
Supplementary information
The original text of this article:
- Ethan Mollick,Choosing to Stay Human(One Useful Thing, 2026-05-27)
The study mentioned in the article:
- A Turkish high school math experiment (using ChatGPT for homework actually resulted in lower test scores):PNAS paper
- Python course experiments at ten universities in Taipei (AI-customized questions, progress equivalent to 6-9 months of additional study):Full research paper (PDF)
- BCG's experiment with 758 consultants (when faced with problems that AI couldn't solve, the humans using AI actually performed worse):Organization Science
- Anthropic Programmer Study (Full Outsourcing vs. Requiring AI Interpretation):Anthropic Research
- "Cognitive surrender":Wharton / SSRN
