Misusing AI has left office workers in deep trouble.

AI has become a regular part of many people's daily work.

PR managers use it to draft planning proposals, lawyers leverage it to generate basic legal documents, and programmers rely on it to write code. For many, it's like an always-available intern right by their workspace.

What's more, AI's fast delivery speed can easily make people get carried away, even creating the illusion that they've "mastered magic".

Until the magic reveals its other side.

It can fabricate non-existent competing brands and industry data out of thin air, cite fictitious legal provisions with a serious demeanor, and arbitrarily modify modules in code that should never be touched. Many mishaps happen precisely because the output looks so authentic that people let their guard down.

We talked to several people who have been tripped up by AI at work. They are not unskilled at using AI. On the contrary, AI has been integrated into most of their workflows. Even so, they still get led astray by it at certain moments.

These experiences remind them that the more complete answers AI can provide, the more users need to retain the ability to judge, verify, and act as the final safety net.

The so-called AI-driven efficiency improvement never means handing over work entirely. It is more like a new round of workplace screening: people who know how to use AI will become faster, while those who only copy and paste may be dragged down by it at critical moments.

01. AI made up three non-existent competitors, and I was exposed on the spot by the client during the presentation

Lin Chen | 30, Strategy Manager at a PR agency, Shanghai

I work in PR strategy. At a service firm, strategists are often called "proposal machines". A single call from a client may require us to thoroughly understand an industry we knew nothing about before, dig out insights, and deliver a complete marketing plan in an extremely short time. The industry is unfamiliar, time is tight, yet what the client demands is "stunning".

So when AI first emerged, I practically pounced on it.

I spend at least a third of my working time every day on large language models, mainly for brainstorming and building frameworks. In the early days, it was truly "enjoyable". Once, a client requested a marketing concept targeting the "silver economy" with only two hours of lead time. I had AI generate more than a dozen directions in one go, picked one, polished it slightly, and submitted it. The client was extremely satisfied.

It felt like I had suddenly mastered some kind of magic. But it was exactly this "enjoyment" that gradually made me let my guard down.

What really made me stumble happened in April this year.

We received a multi-million-yuan annual full-package bidding project. The client wanted to launch a premium pet food line. As usual, the timeline was tight — we needed to deliver the proposal in less than a week. I was responsible for the competitor analysis section.

I fed the client's materials and the names of several major competitors into the large model, asking it to generate a detailed competitor analysis report, with a focus on identifying the "marketing blind spots" of these competitors in the high-end market.

AI quickly delivered the result, with rigorous logic. It not only analyzed the known competitors but also "extra" listed three "emerging premium pet food brands" I had never noticed. It even detailed their marketing tactics and audience profiles vividly, and cited data from a certain pet industry research institution at the end. Its conclusion was also clear: the client should focus on the "origin traceability" concept to fill that market gap. Since time was too tight, I didn't verify anything and directly copied this section into the PPT.

On the presentation day, when I reached the "competitive landscape" part, the client suddenly interrupted me: "Why have I never heard of any of these three brands?"

My mind went blank, and my palms broke into a sweat immediately. I tried to stay calm and said I would "check it later".

As it turned out, those three brands were completely fabricated by AI, and the data from the research institution had no traceable source.

We lost the project. The company deemed me seriously negligent in the review process and deducted my performance bonus for three months. I think this handling is fair.

After this incident, I set a rule for myself: all data, cases, and brand names generated by AI must be cross-verified one by one. If the source cannot be found, I would rather not use them. I also started building my own competitor database, and when using AI, I directly feed it real materials and require it to analyze based on factual information instead of making things up.

Gradually, I began to feel that AI-driven efficiency improvement is actually a pseudo-proposition.

It saves me the time of writing the first draft, but I have to stay vigilant all the time. The more capable it is of writing, the higher the requirements for your identification ability.

So now, my relationship with AI is that I can't live without it, but I also dare not trust it blindly.

02. AI arbitrarily added extra code on its own, dragging the entire team into overtime for a full traceback

Lu Yao | 32, Product Manager, Shanghai

I work in product management.

In the past, our workflow was clear: product teams write requirements, R&D and design teams implement them, and the project goes online after testing and acceptance. Now, AI has penetrated almost every link of this pipeline.

I use it relatively conservatively. Usually, I just ask it to check historical requirements for duplicates, or sort out my messy requirements to make them clearer. But I still judge the final wording and which parts are usable — I never hand over that part to AI. After R&D colleagues receive the requirements, they also feed them to AI first, asking it to sort out the logic and clarify requirements before outputting a solution.

This workflow itself is not problematic, but the colleague who took on this requirement was a new graduate who had just started working.

He was not familiar with the business, nor did he know much about our pile of historical code. He didn't even fully understand the solution delivered by AI. He followed AI's solution, ran a version, and found no obvious issues. The product and testing teams also reviewed it, and the effect was exactly what we wanted. Everyone thought the project could move forward smoothly.

It wasn't until large-scale testing that we noticed something was wrong.

It turned out that AI didn't stay within the small scope defined by the requirements at all. It "improvised" a lot, adding plenty of extra content. Almost half of that version of code was written "casually" by it. It's like an overly enthusiastic new colleague: you only ask them to help move a table, but they rearrange all the furniture in the entire office on their own. And we only focused on checking that single table.

If it were a document, we could just delete a few extra paragraphs of nonsense. Code is different — modifying one part can trigger a chain of problems. Those days, the product, testing, design, and R&D teams were all dragged into overtime. We had to check whether the online system was affected, identify which parts were extra code written by AI, and trace back to the root cause of the problem.

We also tried to let AI find the problem itself. It would confidently claim that it had located the root cause. But when experienced colleagues checked, the piece of code it insisted was problematic had nothing to do with the issue at all. In the end, we had no choice but to delete that version of code and rewrite it manually from scratch.

We were lucky this time — the problem was caught before the launch. But I don't think this is entirely AI's fault. The bigger gap is that people and the workflow failed to keep up. Even if this colleague didn't make a mistake, sooner or later, Colleague B or Colleague C would stumble over a similar problem. The top priority now is to sort out the workflow: at which link did AI introduce the problem, and where should the focus of review be adjusted in the future.

I am more cautious now. AI does save time — a problem that used to take me a day or two to research can now be given several directions by it in a dozen minutes. But I will definitely review it again and ask for sources.

You must be very clear about what you want, think through all the details sufficiently, and then let AI execute. Otherwise, if you hand over a vague idea to AI, it will only return an even vaguer result. Even if this risk doesn't explode immediately, it is just delayed, waiting to detonate all at once someday.

AI is indeed smart, but the premise is that you can control it. Therefore, the more powerful AI becomes, the higher the requirements for people. If someone reduces their work to just copying and pasting from AI, being eliminated is only a matter of time.

03. I used AI to cut my workload by 40%, but almost stumbled on a fictitious legal provision

Delia | 31, Lawyer at an independent law firm, Beijing

I work at a law firm with a high proportion of foreign-related businesses, where I handle a large number of Chinese-English contracts and commercial dispute cases daily. Now, AI has basically become my "invisible colleague", present in most of my workflows. But this "colleague" wasn't easy to use from the start — it was polished little by little after countless pitfalls.

I mainly use it for three things: Chinese-English contract translation, basic legal document drafting, and industry legal hotline research.

Translation is the part that makes me feel most reassured. The vast majority of contracts in the law firm are bilingual, which is the most repetitive work that consumes the team's energy. AI has always performed stably in this area, with professional terminology and sentence logic that fit legal writing norms, hardly making mistranslations or omissions. This alone directly cut down nearly 40% of our basic workload, greatly reducing the team's repetitive burden.

Document drafting, however, is the link where pitfalls are most likely to occur.

In fact, since 2024, I have been trying to use AI to search for legal provisions and assist in drafting documents. It occasionally makes mistakes, but back then I could avoid these hallucination problems through the final manual review.

What really sent a chill down my spine was when it started talking nonsense with a straight face in areas outside my knowledge.

Earlier last year, I was handling a local equity dispute case. The timeline was tight, and the regional special regulatory provisions involved were extremely complicated. I asked AI to assist in drafting the first draft, requiring it to supplement local legal provisions and similar effective cases to support the arguments.

It quickly delivered the finished product, with smooth writing logic and standardized citation entries. After I made some revisions to the major laws and regulations, the whole document looked fine. I was busy coordinating with clients at the time, so I hastily filed it in the case file.

Fortunately, when the case file was submitted to the partner for review, a colleague spotted the problem. Several local legal provisions in the document were completely fictitious, and the errors were so subtle that they could hardly be noticed without dedicated research on local regulations.

I woke up instantly at that moment. AI can help me improve efficiency within the framework of my cognition, but even a tiny bit beyond that framework will make me stumble.

To avoid the risk of AI hallucinations, I started building my own database. Every time I use AI, I restrict it to only search within this database. Later, when some professional paid databases became available, I subscribed to them immediately. After long-term use, I am getting more and more comfortable with them now.

When communicating with peers, I found that many people are still in misunderstandings. Quite a few people are used to cross-reviewing documents with multiple AI models, thinking that this can avoid errors. But the actual effect is minimal — several models may make the same mistake at the same place, and many people have stumbled over this pit.

Now I adhere to one point of view: AI has no inherent good or bad, and whether it works well depends entirely on the user's professional knowledge. It can free our hands, but it cannot replace lawyers' logical analysis, legal provision verification, and value judgment.

At present, the general consensus among legal practitioners is that AI hallucinations can be resolved through users' capabilities, but its causal reasoning and logical abilities are still far from mature, far from being able to disrupt the industry. After all, law is never a simple "yes or no" question.

If you don't know how to use AI, you will gradually fall behind; but blindly relying on AI without thinking is even more dangerous.

04. The project was delayed for two weeks, I spent 7,000 yuan out of pocket on AI tokens, and I also got a "black mark" in my career

Tong Tong | Post-90s generation, Programmer, Shenzhen

I have been a programmer for 10 years. I started using AI to write code in March last year, and handed over all coding work to AI in December. Now I hardly write code manually myself.

Not long ago, our team needed to carry out intelligent upgrade for an old product of the company, building an Agent system. The initial usage experience was indeed smooth, and the efficiency multiplied several times. I gradually relaxed my vigilance in verification. According to the normal pace, this project could be completed in one month.

Once, I wrote down the requirements clearly, set the standards, and went to sleep, waiting for AI to deliver the result the next day. But when the code ran, there were obvious black-box defects. By the time my colleagues and I noticed the problem, it had arbitrarily modified the existing online code logic and broke the code that others had already written. I could only quickly apologize in the group chat: "Sorry, I crashed your code." Then I worked with my colleagues to remedy the situation.

After this incident, I also strengthened my review of AI, dealing with its small flaws every day. We stumbled along for a month and a half, and the project still hasn't been delivered. The project that was originally scheduled to be completed in four weeks has now been delayed for two weeks. I also paid for additional AI tokens out of pocket, spending more than 7,000 yuan extra. Even my product manager, who couldn't stand the torment caused by AI, just submitted his resignation to me.

I have always been known for delivering work on time and never allowing delays. This incident made me deeply doubt myself. More fatally, this incident has disrupted the company's business layout — the boss originally planned to assign me to take over another new business, but I have been stuck in this project and can't get out. The leader didn't explicitly hold me accountable verbally, but it's obvious that their patience is almost exhausted. Since the project hasn't ended yet, the specific punishment hasn't been determined, but I know clearly in my heart that this matter will be settled sooner or later.

AI did help me improve efficiency, but it didn't make my life easier. It did save manpower in some links — we don't need to hire someone specifically for front-end development, and I can ask AI to adjust the copywriting myself. The overall efficiency has increased by two