After graduating, this group of college students may have to work as nannies for AI

Indeed, AI has caused countless companies to lay off employees, but on the other hand, AI is also generating new job opportunities.

According to the latest data of the cloud recruitment month activity of Internet enterprises launched by the Ministry of Human Resources and Social Security, more than 5000 Internet enterprises released more than 200000 jobs this summer. Jingdong, Tencent, ByteDance, Meituan and other leading enterprises have contributed more than 46000 jobs in total, covering cutting-edge directions such as AI algorithm, large model application and high-performance computing.

Not only have job positions increased, but job types have also increased. For example, algorithm engineers, prompt word engineers, and so on. And the recently emerging botsitting is likely to become a new reservoir of labor.

What is botsitting? Babysitting refers to the nanny who takes care of babies, while botsitting refers to the nanny who takes care of robots (AI).

Microsoft’s Copilot department has positions such as “AI Trainer,” “Digital Adoption Specialist,” and “AI Advocate. Their work, to some extent, is a form of botsitting.

Specifically, it means teaching colleagues how to use AI, checking the quality of AI output, and embedding AI into specific business processes.

In practical work, from the answers output by AI to their actual application in real-world scenarios, the two are often not the same thing.

AI may output illusions, or it may completely misunderstand your meaning. In short, you need to spend some time modifying the answers that AI outputs to you.

How long is this specific time?

According to the “Work AI Index 2026” report jointly released by Glean’s Work AI Institute and universities such as Stanford and UC Berkeley, a white-collar worker needs to spend nearly a day’s work every week on botsitting.

Moreover, botsitting is not a difficult task, and with a little experience in using AI, one can easily get started. Therefore, it is just suitable for the group of college students who have just graduated and want to learn about the industry.

Just like the data annotators of the past, there were very few people working at the beginning, but with the development of AI, people have become increasingly aware of the importance of this position, so the position has also grown.

What is botsitting all about?
So what exactly does botsitting do? For example, you will understand.

For example, you can use AI to help you write a market analysis report. At the beginning, AI doesn’t know about your company’s product line, so you need to feed in the background information first. After the results come out, you need to verify the competitor data one by one.

Finally, you need to format the content output by the AI. This process takes about one or two hours to complete.

According to the Work AI Index 2026, 87% of white-collar workers are already using AI in their work, claiming to save an average of 13 hours per week.

However, they spend an average of 6.4 hours per week on botsitting. That is to say, half of the time saved by AI will be returned by you personally.

Another thing is that botsitting has a tendency towards self deterioration. 69% of people submit AI generated content without any review.

The report further points out that 37% of the time employees spend on AI each week is spent on botsitting, 36% on using AI to produce things, and the remaining 27% on learning tools and building agents.

This data is indeed somewhat outrageous, but reality is even more outrageous.

The reason why botsitting is more time-consuming than using AI in practice is that 36% of AI sessions directly “fail” and need to be completely overturned and restarted. Another reason is that there are too many tools, with 77% of AI users having to switch back and forth between multiple AI tools every week, and 33% using four or more at the same time.

Among Claude’s users, only 0.5% only use Claude, and on average, each person is running four other AI tools at the same time. Every time you change the tool, the company background and project context that were previously fed in are interrupted, and you have to feed them again later.

The report calls this’ context tax ‘.

The report shows that for every 10% extra time spent feeding context to AI, the probability of employees feeling exhausted increases by 25%.

Moreover, heavy AI users have more than twice the frequency of botsitting compared to light users. It means that the more you use AI, the higher the proportion of time you spend as a nanny.

The more unstable the output quality of AI, the more botsitting people need; The more botsitting, the more tired a person becomes; The more tired a person is, the more likely they are to skip the review and submit directly; The more people skip the review, the less organizations can see the true returns of AI.

The report shows that 75% of individual users believe that AI has improved productivity, but in reality, only 13% of people say that businesses have achieved significant improvements because of AI. 62% of this is due to a lack of botsitting.

Why botsitting is naturally suitable for recent college graduates
The Job Outlook 2026 report from the National Association of Colleges and Employers (NACE) states that 45% of employers rate the job market as “fair” in 2026, which is the worst rating since 2021.

In the autumn of 2025, employers predict a recruitment growth of only 1.6% for fresh graduates. Although this number rebounded to 5.6% in the spring update, large technology companies have reduced their recruitment of fresh graduates by 25% in 2025.

However, the “Work AI Index 2026” believes that botsitting will create employment opportunities for a large number of college students.

The first reason is that the entry threshold for botsitting is extremely low, but the industry exposure is extremely high.

When many people hear about jobs related to AI, what comes to mind is writing code, tuning parameters, and training models. Bottiting is completely different from that.

It doesn’t require you to have any algorithmic knowledge, it just requires you to have the most basic human judgment.

Can you tell if this passage written by AI is nonsense? Can you determine if the conclusion of this data analysis report matches the previous data? Can you recognize the pile of seemingly professional but actually hollow terms that AI stuffed in to make up for the word count?

As long as you are a normal human with basic university education, you possess such abilities.

The second reason is that this group of college students are AI natives.

A fact is that most of the papers for college graduates in 2026 were completed by ChatGPT, Claude, and DeepSeek. But many senior employees still use AI as a faster search engine. Ask a question and get an answer.

But this graduating class is different. In the process of being repeatedly tricked by AI, they have established an intuition about when to trust AI’s output? When should I take another look? When does AI seem to be answering, but actually go in circles?

More importantly, this generation naturally understands what the “prompt word project” is, which is the lesson they learned through failing exams.

They also don’t think that repeatedly modifying the output of AI is an “extra job”, because this is already their daily workflow: let AI generate the initial draft, make major revisions themselves, then let AI polish it, and then check it themselves.

In other words, botsitting may be an “extra burden” for veteran employees, but for this graduating class, it is the “normal way of working” they have learned.

The third reason is moderate labor intensity.

As mentioned earlier, it takes 6.4 hours per week. And for those college students who are good at AI, the time will be even shorter.

NACE’s survey also shows that 70% of employers are already using skill first recruitment methods, an increase of 5% from last year.

What is skill first? It doesn’t depend on your school or major, it depends on whether you can do this job.

Another easily overlooked point is that botsitting is not a job that ends at a glance, it has a clear upward path.

Companies like Scale AI and Surge AI directly recruit full-time quality analysts and project managers from the top performing trainers. The starting salary is around 10-20 US dollars per hour, and after becoming a regular employee, the annual salary can reach 40000 to 60000 pounds.

For a newly graduated college student, your promotion criteria can be quantified. If you correct more mistakes and are more accurate than your peers, then you can be promoted.

Will botsitting become normalized?
There is a more fundamental question that has not been answered: is botsitting a temporary phenomenon in the transitional stage? Is it still a regular position?

To understand this, we need to first pull out a reference object – data annotation.

Ten years ago, the term ‘data annotation’ was almost unheard of.

Take the issue of “how to make the model recognize a cat” as an example. The solution at that time was to hire someone to label the pictures one by one, this one is a cat, this one is not a cat, and the face of this cat is partially covered but still a cat.

Fortune Business Insights once released a set of data that around 2015, there were only a few tens of thousands of data annotation practitioners in China, with a total market size of approximately 500 million yuan. By 2020, the size of China’s data annotation market had reached 3.1 billion yuan. By 2025, it will exceed 10.5 billion yuan.

On a global scale, the market for data annotation tools is estimated to be between $1.7 billion and $3.6 billion by 2025, and is expected to grow to $14 billion to $38 billion by 2034, with a compound annual growth rate of over 26%.

Since models are becoming stronger, why do more and more people need to annotate them?

The answer is simple, the stronger the model, the more complex the task to be processed; The more complex the task, the more refined the training data required; The finer the data, the more irreplaceable the manual judgment required.

As AI takes one step forward, the requirements for data annotation become higher.

Botsitting is repeating the path of data annotation and will go further.

Firstly, the core bottleneck of botsitting is not technology, but organization.

Harvard Business Review published a heavyweight article in March 2026 titled “The ‘Last Mile’ Problem of Slowing Down AI Transformation”.

The authors of the article include Karim Lakhani from Harvard Business School, Jared Spataro, head of Microsoft AI at Work, and Jen Stave from Harvard D ³ Institute.

The core conclusion they presented in the article is that the model can become infinitely stronger, but as long as it does not understand the internal logic of your company, you must send someone to deal with its aftermath.

This is not a problem of the model not being good enough, it is an information gap between the model and our real world. Every enterprise has its own unique context, such as unwritten rules, corporate culture, and so on.

These things are not in any public dataset, and AI can never learn them on its own. But as long as AI is to be effective within an organization, someone must translate, supply, and correct these things.

And this is precisely the core value of botsitting.

It is not about replacing the insufficient capabilities of AI, but about bridging the information gap between AI and organizations. As long as the enterprise is unique, this gap will always exist, and there will always be people in need of botsitting.

Secondly, the way AI works determines that it naturally requires humans to provide support.

Forbes’ article in 2026 titled ‘Is AI Replacing Jobs?’? The new data suggests that it may be increasing the workload “mentions a viewpoint that AI will not make jobs disappear, it will only make jobs shift.

The article argues that as long as AI enters the real workflow, there must be someone to supervise, edit, verify, and backstop. ”

In 2025, a scandal occurred in the Mississippi federal court in the United States. Several lawyers from Butler Snow law firm wrote AI fabricated precedents directly into court documents and didn’t even glance at them when submitting them.

Judge Britton Manasco exploded after reading the verdict, stating that ‘fabricating legal grounds constitutes serious misconduct’.

The conclusion drawn from a survey of 85 legal professionals conducted by the National Law Review in early 2026 is that the differentiation competition point for lawyers in the future may not lie in which AI is used, but in their ability to output verification.

Human machine collaborative workflow, quality control, and defensible review processes will become the core competitiveness of the legal industry, rather than optional safeguards

Morgan Stanley launched an AI tool called Morgan Stanley Debrief in 2024, and by the end of 2025, 98% of wealth management advisors will be using it.

However, Morgan Stanley has a rule that advisors must “review and adjust the AI generated output before finalizing” when using AI generated meeting summaries and investment recommendations.

At the same time, the Financial Industry Regulatory Authority (FINRA) in the United States has specifically added a regulatory framework for “AI systems capable of autonomously executing tasks” in its 2026 annual regulatory report released in December 2025, which clearly requires that once AI systems are able to take action in brokerage workflows rather than just generate content, the company’s supervision, bookkeeping, and governance obligations must be substantially upgraded.

In plain language, AI can help you write documents, conduct analysis, and even make decision recommendations, but the one who signs in the end must be a living person.

Not only that, the person who signs must also explain and be responsible for what they sign. It can be delegated, but responsibility cannot be outsourced, so botsitting is destined to exist forever.

Thirdly, empirical evidence has emerged.

The World Economic Forum’s 2025 Future Jobs Report states that AI and big data experts are among the fastest-growing jobs by 2030, but also points out that non-technical positions such as AI governance and AI strategy are also growing rapidly at the same time.

According to a report by Mercor, an AI talent platform, the demand for human assessors and trainers in the global market is growing at a rate of 25% to 35% annually. And the report also mentioned that most positions are completely remote and do not require technical background, but place great emphasis on domain expertise and judgment.

When AI transforms from an efficiency tool for individuals to an infrastructure for an organization, botsitting shifts from ‘whoever has time to do it’ to something that must be handled by a dedicated person.