In Liu Cixin’s *The Three-Body Problem*, the Dark Forest theory describes a universe where every civilization is a hunter. Revealing your coordinates means annihilation. The safest strategy is silence, concealment, making sure no one knows where you are.
This same law is playing out in the digital world, except the hunters are AI companies and your coordinates are your patterns of thought.
A Hacker News post with 376 upvotes captured a growing unease that many feel but few articulate well: you use AI every day, you know these companies are collecting your data, you want to reduce your dependence, but you realize you can’t function without it anymore. The irony cuts deeper. When you want to figure out how to reduce your dependence, your first instinct is to ask the AI itself.
This isn’t a technical problem. It’s a cognitive trap.
Your Prompts Are Your Cognitive Fingerprint
Most people think privacy means not leaking their ID numbers or credit card details. But privacy in the AI era runs much deeper.
Consider the prompts you send to ChatGPT or Claude every day. You ask how to write a job rejection email, and it learns you’re job hunting. You ask how to handle a conflict with a coworker, and it maps your workplace relationships. You have it revise your resume, and it catalogs your skills and professional anxieties. You ask it to help you write something romantic, and it knows your relationship status.
Each prompt looks harmless on its own. But string together three months of someone’s prompt history and you get a map of their thinking more precise than any social media profile could offer.
Your social media presence is performative. You edit photos before posting, you craft your tweets carefully, you package your LinkedIn experience. But when you talk to an AI, you’re authentic. You have no audience, no need to perform. You just want to solve a problem.
This authenticity is exactly what makes it dangerous.
Meta spent over a decade using your likes, comments, and browsing time to infer your preferences. Google uses your search history to guess what you might buy. But AI companies don’t need to guess. You tell them directly. You type your needs, your confusion, your decision-making process into that dialogue box, word by word.
The Data Flywheel: Better Tools, Deeper Dependence
AI products follow a classic growth model called the data flywheel. More users generate more data, which improves the model, which makes the product more useful, which attracts more users. Once this loop starts spinning, it’s hard to stop.
But few people consider what this flywheel means from the user’s perspective.
You use ChatGPT to write some code. It’s not quite right, so you fix it and resubmit. That correction becomes high-quality training data. You’ve taught the model what the right approach looks like in that situation. Next time someone asks a similar question, the model gives a better answer.
You use Claude to analyze a business plan. In the conversation, you point out where its analysis went wrong and what it missed. That feedback makes the model smarter.
You use Copilot to write code. Every Tab acceptance and every Esc rejection tells the model what good code looks like and what to avoid.
You’re using these tools for free. But you’re also working for them for free.
This exchange looks fair early on. You get a powerful assistant, it gets some training data. Win-win. But over time, the balance shifts. The model learns more about you, gets better at predicting your needs, and you grow more dependent on it. What do you know about the model? Nothing.
You don’t know where your data goes. You don’t know how many people benefit from improvements based on your input. You don’t know if your contributions would be deleted if you stopped using the service tomorrow. Probably not.
This is the other side of the flywheel: the faster you help it spin, the harder it becomes to jump off.
The Cost of Outsourcing Thought
In 2011, Columbia University’s Betsy Sparrow published research showing that people remember information less well when they know they can Google it later. This became known as the Google effect or digital amnesia.
Fifteen years later, we can say search engines have changed how humans remember. We no longer memorize phone numbers, directions, or facts we can “just Google.” Our brains outsourced information storage to search engines and freed up space for other things.
This change has pros and cons, but at least search engines only outsource memory. You still do your own thinking, judging, deciding. Search engines give you information. You process it.
AI assistants outsource thinking itself.
When you have AI write your email, you’re not just outsourcing typing. You’re outsourcing word choice, tone calibration, information organization. When you have AI create a technical proposal, you’re not just outsourcing documentation lookup. You’re outsourcing solution comparison, tradeoff analysis, risk assessment. When you have AI write your code, you’re not just outsourcing keystrokes. You’re outsourcing algorithm selection, architecture design, edge case reasoning.
Each time you outsource, the related capability atrophies slightly. Just as wheelchair users lose leg muscle over time, people who rely on AI for thinking lose their capacity for independent thought.
This isn’t alarmist speculation. A 2024 Microsoft Research paper surveyed developers using AI coding tools and measured changes in their programming ability. Heavy users of AI coding tools showed declining ability to solve algorithmic problems without AI assistance compared to a year earlier. The decline wasn’t dramatic, but the trend was clear.
More subtly, this degradation happens gradually and imperceptibly. You won’t wake up one day unable to code. You just slowly become less willing to do it yourself. Because AI is faster and “good enough.” That “good enough” is the trap. It lets you feel competent while choosing a more efficient path. But when you actually need to solve a problem the AI can’t handle, you’ll find your instincts have dulled.
The Paradox of Escape
This is the cruelest part of the cognitive dark forest: you recognize the problem, but you can’t find the exit.
You want to reduce your dependence on ChatGPT. But you’ve grown accustomed to using it for organizing thoughts, drafting documents, analyzing decisions. Without it, your productivity drops sharply. Your boss won’t extend your deadlines because you’re “exercising independent thinking skills.”
You want to protect your data privacy. But you realize that not using AI tools puts you at a competitive disadvantage. Your colleague finishes in one day with Copilot what takes you three. Your competitors use AI for market analysis while you’re still reading reports manually.
You want to figure out how to maintain independence while using AI. What’s your first response? Open ChatGPT and type: “How do I reduce my dependence on AI?”
This is the paradox. The tool you want to escape from is exactly the tool you use to plan your escape route.
It reminds me of quitting smoking. Every smoker knows cigarettes harm health, but the quitting process itself creates stress, and stress makes you want to smoke. AI dependence doesn’t have nicotine’s physiological withdrawal, but the psychological mechanism is similar: the more anxious you feel, the more you want to use AI to ease that anxiety.
Exposed in the Dark Forest
Back to the dark forest analogy. In Liu Cixin’s universe, civilizations can choose silence. But in the AI cognitive dark forest, silence isn’t an option.
Because not using AI is itself a signal. When your peers all use AI to boost efficiency, your “silence” means falling behind. This isn’t a game you can unilaterally exit.
Your data exposure doesn’t only happen when you actively use AI. Your colleague uses an AI tool to process a document containing your information. Your company feeds customer data to AI for analysis. The email you send gets read and summarized by someone else’s AI assistant. The code you write gets included in Copilot’s training data.
You can choose not to use AI, but AI is using you.
This passive exposure is the hardest to defend against. You can choose not to use ChatGPT, but you can’t stop others from feeding information about you into ChatGPT. You can choose not to use Copilot, but you can’t stop your open source code from training Copilot.
Possible Exits
After all that pessimism, let’s talk about whether there are ways out.
The first path is local models. Open source models like Llama, Mistral, and DeepSeek can now run on consumer hardware. An M4 Max MacBook Pro can smoothly run 70B parameter models. Local models mean your prompts never leave your computer, won’t train anyone’s model, won’t get stored or analyzed by any company.
Local models aren’t as capable as GPT-4 or Claude, but the gap is closing fast. For daily code completion, document drafting, and information organization, local models are sufficient. You only need cloud models when you need the strongest reasoning ability.
The second path is conscious boundaries. Not avoiding AI completely, but drawing a line for yourself. For example: think through technical solutions yourself, only use AI to validate. Or: write the first draft yourself, only use AI to polish. Or: write core logic yourself, only use AI for boilerplate code.
Where you draw that line varies by person, but having the line matters. Usage without boundaries eventually becomes dependence without limits.
The third path is periodic disconnection. Just as some people do “digital detoxes” and turn off their phones for a weekend, you can do “AI detoxes.” Spend one day each week completely avoiding AI tools, forcing yourself to work independently. This isn’t to prove you don’t need AI. It’s to maintain your ability to work without it.
The fourth path is data sovereignty. Choose services that explicitly promise not to train models on your data. Claude’s API defaults to not using conversation data for training. ChatGPT’s API does the same. But free ChatGPT uses your conversations to improve models unless you manually disable it. Many people don’t know these details, but they determine where your cognitive data ultimately flows.
Living Awake in the Forest
The cognitive dark forest isn’t going away. AI companies won’t suddenly stop wanting your data, AI tools won’t suddenly become less useful, and your peers won’t suddenly stop using AI.
What you can do isn’t escape the forest. It’s stay awake while living in it.
Know what you’re giving up. Before opening that dialogue box, take one second to ask: what information about me does this prompt contain? Am I willing to let a commercial company permanently store this?
Know what you’re losing. Before having AI think for you, ask yourself: does this thinking process itself have value? If I keep outsourcing this kind of thinking, will I still be able to do it independently in three years?
Know what you’re feeding. Every interaction makes AI more powerful, better at understanding humans, harder to replace. This isn’t necessarily bad, but you should be informed and intentional, not unconscious and passive.
The most dangerous thing in a dark forest isn’t the hunters. It’s forgetting you can be a hunter too. When you clearly understand the rules of the game and know the costs and benefits of each move, you stop being prey. You become someone walking into the forest with a map. The map isn’t complete, but at least you know where you are and where you’re going.
Don’t outsource the map to AI either.
Stay updated with our latest AI insights
