AInception
Study shows that workers hired to train AI often use AI to do their job.
An AI system without data is as helpful as a head without a brain. To make predictions and perform tasks, the AI needs to be fed with good data. A LOT of data. And this data doesn’t grow in trees. Companies often use platforms like Amazon Mechanical Turk, Upwork, or Prolific to hire workers, usually poorly paid, to perform tasks, for example, labeling images, solving CAPTCHA, and summarizing texts. In a recent preprint, researchers found that more than 30% of the summarized texts submitted by these workers were actually being generated by AI tools, such as ChatGPT.
The ‘invisible’ AI workforce earns an average of 2 dollars per hour and often deals with toxic content generated by the AI that must be filtered. The conditions are poor, and the pressure to perform the work efficiently pushes people to seek ways to increase productivity.
But why don’t we outsource all the work to AI? Artificial data can be problematic. Any problems in the system will pass through the artificially generated data, and when used to feed the AI system, all errors and issues will be reinforced. It is like your dog is constantly barking at the postal worker, and instead of correcting the behavior, you give him a treat after the barking. The dog will keep associating the barking with something good.
The researchers also highlight the importance of creating new methods to distinguish AI from human-generated data and reevaluate the data generation process. They added, “We do not believe that this will signify the end of crowd work, but it may lead to a radical shift in the value provided by crowd workers.”
