Summary of Key Points
Data annotation is the “human foundation” behind AI. ChatGPT’s ability to write poetry, autonomous vehicles’ recognition of traffic lights, and voice assistants’ understanding of commands all rely on human annotators to process data. However, the industry has become highly segmented: salaries range from 2,000 to 65,000 yuan per month, a difference of 30 times. The nature of the work has evolved from mechanical tasks (such as “drawing frames around objects”) to requiring specialized knowledge for AI training. Companies follow a pattern where large corporations set the rules and outsourcing firms handle the execution. Additionally, automated annotation technologies are replacing low-level jobs, leading to a growing shortage of expert annotators with domain-specific knowledge.
Detailed Analysis
1. Why such a huge income gap? Why do some earn 2,000 yuan while others earn 65,000 yuan for the same job?
Income in data annotation follows a “pyramid” structure:
- Lower level (2,000–5,000 yuan): This includes part-time and crowdsourced work involving repetitive tasks like outlining traffic light frames in images or transcribing speech. Education and experience requirements are minimal, and payment is based on the number of days worked (around 100–200 yuan per day); anyone can do this.
- Middle level (around 10,000 yuan): Full-time annotators perform more complex tasks but still follow established rules; they are somewhat replaceable.
- Upper level (20,000–65,000 yuan): These annotators don’t just label data; they set the rules. For example, they develop annotation standards for large models, evaluate AI code for bugs, or manage the quality of medical image annotations. They typically have a master’s degree and a background in computer science, medicine, finance, etc. Such skills make them hard to replace, resulting in higher salaries.
For instance, Baidu offers interns in autonomous vehicle annotation 500–600 yuan per day (with a master’s degree required), while crowdsourced annotators earn only 185 yuan per day, a difference of over three times.
2. Annotators are no longer just “drawing frames”? Now they need to understand code, dialects, and even medicine!
Previous work involved simple tasks like drawing frames around objects, but now annotators must be more professional:
- Text annotation: In the era of large models, knowledge of coding is essential. Tencent, for example, looks for candidates who can identify bugs in AI programs.
- Speech annotation: Annotators need to understand dialects and detect emotions. Musk’s xAI project, for instance, requires Chinese speakers familiar with Sichuan dialects to identify pronunciation errors. JD.com requires French-speaking annotators with a high level of language proficiency.
- Image/video annotation: In the autonomous vehicle field, knowledge of sensors is necessary. A car company is hiring “intelligent driving annotation engineers” to work with 3D point cloud data (equivalent to analyzing three-dimensional street scenes); these positions pay 40,000–70,000 yuan per month.
- Multimodal annotation: These tasks involve handling text, images, audio, and video simultaneously. For example, annotators training AI to “describe images” need to understand aesthetic principles and language logic. This type of work accounts for 36% of the market demand.
3. The division of labor between large corporations and outsourcing firms: Do large corporations set the rules, and do outsourcing firms do the heavy lifting?
The industry chain is clear:
- Large corporations (JD.com, Tencent, Alibaba): They define annotation guidelines and recruit high-end talent (e.g., for code annotation or large model evaluation), ensuring the quality of core data.
- Outsourcing companies (Haitian Ruisong, Yunce Data): They break down large tasks into smaller pieces and assign them to annotators in third-tier cities. These annotators often don’t know the specific use of their work; they see only the individual tasks on a assembly line.
Why do outsourcing firms prefer third-tier cities? Because labor costs are lower there. In Beijing, it may be difficult to find 4,000 workers for a certain job, but in a small town, 2,000 yuan can attract many candidates.
4. The evolution of the industry over 30 years
Data annotation didn’t always exist; its development has gone through four stages:
- Pre-annotation era (2006–2014): Academics were involved in data collection. For example, Li Feifei used the ImageNet dataset, hiring undergraduates to annotate images. This was initially seen as a low-status task.
- Annotation factory era (2014–2017): It became a commercial activity. Companies set up “factories” in third-tier cities, hiring rural youth and mothers for basic annotation work; salaries were decent, but workers often didn’t understand the purpose of their tasks.
- Segmentation and upgrading era (2017–2020: The Ministry of Human Resources and Social Security included “artificial intelligence trainers” in the official job catalog. Annotators were categorized into three levels: low-level workers, quality controllers, and rule setters (with annual salaries over 300,000 yuan).
- AI’s impact era (2020–present): Automated annotation has emerged. GPT-3 reduced the need for massive amounts of manual data annotation, increasing automation from 30% to over 60%. However, new tasks (e.g., evaluating AI accuracy) still require human judgment and cannot be fully automated.
5. What’s the future? Will AI take away annotator jobs?
The answer is that low-level positions will be replaced, while those requiring specialized knowledge will become more valuable:
- Jobs to disappear: Purely mechanical tasks like drawing frames or transcribing text will be automated, with humans only needed for quality checks.
- Jobs to remain: Those with domain expertise (e.g., medical or financial annotators) and those who can manage projects and set rules.
HR at AI companies say, “We prefer to hire computer science graduates as trainers rather than promoting from low-level annotators” because they lack the necessary professional knowledge and logical skills.
In conclusion, the data annotation industry is undergoing a transformation. Mechanical tasks will be automated, but professionals with specialized knowledge will become increasingly valuable. This is both a challenge and an opportunity for ordinary people to enter the AI field.