In the era of AI, how can we protect our 'conversation privacy'?

We are becoming increasingly accustomed to conversing with AI. Writing a plan, analyzing a contract, organizing customer information, modifying a piece of code, or even encountering work problems, the first reaction may be to copy the content and ask 'help me take a look'. This usage method is very natural. AI is becoming more and more like a colleague who can be reached for help at any time, and we are also getting used to telling it our own problems directly.

But recently I have started to notice an easily overlooked question: What exactly are we telling AI when we ask for its help?

A customer list, a quotation, an internal company policy, an undisclosed product code, a contract under negotiation, and even a chat record containing personal information such as name and phone number. These things may seem like work materials to us, but from a different perspective, they are actually valuable information.

In the past, when we used AI, our biggest concern was often whether the answers given by AI were correct or not. Now that AI is truly entering the workforce, another question is equally worth paying attention to: is this content really suitable for being directly handed over to AI?

Many privacy issues are not because a person really doesn't care about privacy. On the contrary, often it's just because the whole process is too smooth. Seeing a lengthy contract, copy, paste, and have AI summarize it; View a customer data table, upload it, and have AI analyze it; Encountering a piece of code, paste it and let AI find the problem. The whole process may only take a few seconds, and we rarely stop in the middle to think about what information AI really needs in this section.

For example, if we ask AI to help us analyze a contract, it needs to understand the contract terms, obligations of both parties, payment terms, breach of contract liability, and other related content. But does it really need to know the customer's ID number? Do we really need to know the customer's personal phone number? Do we really need to know the real names of everyone in the contract? Many times, it is not necessary.

For example, let AI help sales organize customer communication records. What is really needed may be customer demand, purchase intention, communication progress and product preference, rather than customer ID number, private phone number and complete personal information. This made me realize a very simple principle: we used to be accustomed to handing over complete data to AI and letting AI process it on its own; Perhaps in the future, we should gradually shift to handing over only the information needed to complete tasks to AI. These two approaches may seem one step apart, but the underlying thinking is completely different.

Of course, after realizing this issue, we cannot go to the other extreme and think that since there is a privacy risk, we should not give anything to AI in the future. This is also unrealistic. The value of AI lies in its ability to understand the context and obtain certain information in order to truly help us accomplish tasks. You cannot let AI analyze a complex contract for you without knowing anything, nor can you let an enterprise agent truly participate in internal work without accessing enterprise knowledge.

So the real question is not whether to let AI know, but what it should know, to what extent it should know, what information can be given, what information should be hidden, what things can be done automatically, and what things must be confirmed by humans. This is no longer just a simple privacy issue, but more like a boundary design problem. Just like we don't give all the company's permissions to a new employee, nor do we let an ordinary employee casually view all customer information. AI is the same. If AI truly becomes the digital employee in enterprises in the future, the boundaries between permissions, data, tasks, and responsibilities may become increasingly important.

Because of this issue, I have recently started thinking about an interesting direction. Nowadays, our relationship with AI usually involves humans directly facing AI. We input a paragraph, upload a file, and then wait for the result. But what if there is an additional layer in the middle? Becoming a human goes through a security layer before reaching AI. This layer does not necessarily require users to learn complex security knowledge, it can help us make some judgments before we send them.

For example, it can indicate that personal information has been found in this content, or that it may contain sensitive enterprise data, or that this task does not require the complete raw data to be handed over to AI and can be desensitized first. Even further, it can determine which part of the information can be sent and which part should not be sent. For example, the original content is "Please help me analyze Zhang San's customer contract in the last year, his ID number number is XXX, and his mobile phone number is XXX". After processing, AI may only see "Please help me analyze customer A's contract in the last year". For AI, the information needed to complete tasks is still available, but sensitive information unrelated to the task has been blocked. At this point, AI can still help us complete tasks, but we don't need to hand over everything without reservation just to make AI work. This may be a direction worth exploring: not to prevent people from using AI, but to provide an additional secure buffer layer when using AI.

This incident reminds me of a change. When AI first became popular, the most discussed topic was which model was smarter. Later on, we started discussing which model was better for coding, which model had stronger reasoning ability, and whether the agent could complete the task for us. But when AI truly begins to work, the problem may gradually become what it should know, what it can do, what it cannot do, who can let it do it, and who will confirm when problems arise. That's also why I increasingly believe that the next stage of AI is not just about making models smarter, but also about solving another thing - how to make us dare to use it.

Because if an AI doesn't know anything, it's difficult for it to truly help you work. But if an AI knows and can do everything without boundaries, it will also generate new problems. So what I am more interested in now is not simply creating a tool that prohibits users from uploading sensitive information. What I am more interested in researching is whether AI can still do things well without knowing the information it shouldn't know. This is actually a set of tool directions that I have been thinking about recently. It may not be a simple privacy detector, nor just a simple code for mobile phone number and ID number. I hope it can eventually become a layer of infrastructure between humans and AI, identifying risks before sending, desensitizing when needed, confirming permissions when involved, and trying not to hand over data when it can be processed locally.

Of course, there are still many issues in this direction that need to be further verified. What information must be protected, which information is not necessarily overly protected, can AI still accurately complete tasks after desensitization, and why are companies and individuals truly willing to pay. I don't have a final answer to these questions yet. But I think there is at least one issue that is becoming increasingly clear: AI can become smarter, but we also need to become clearer about our own boundaries.

In the future, when we use AI, it may not just be about inputting, sending, and waiting for answers, but gradually become about judging, protecting, entrusting to AI, and completing tasks. This may be a usage method that we need to relearn in the AI era. It's not about letting AI know more, but about letting AI help us do more with more confidence. This is also something I want to continue researching next.