A study by the Swiss Federal Institute of Technology suggests that many human AI model trainers could be outsourcing their work to AI, MIT Technology Review reports.

The researchers hired 44 gig workers on Amazon’s Mechanical Turk platform and asked them to summarise 16 extracts from medical research.

Companies often use Mechanical Turk to hire people to help train AI.

They are appointed to perform tasks that are difficult to automate, such as solving CAPTCHAs, annotating text, and labelling data.

The researchers analysed the responses from their contracted workers for signs of ChatGPT-generated output, such as low word variety, and keystrokes that could indicate the copy-and-paste workflows needed to take text from ChatGPT.

They estimated that between 33%–46% of the workers had used AI models such as ChatGPT to come up with their responses.

The practice could accelerate the process of AI model collapse — wherein AI models get gradually worse by being trained on AI-generated content.

Researchers from the US and Canada recently published a paper highlighting the issue, explaining how they were surprised by how quickly the process could occur.

Feeding AI-generated content into AI models led to “irreversible defects” in the models wherein they “forgot” their true underlying data distribution.

The increased amount of AI-generated content being posted online is expected to accelerate model collapse.

In addition to highlighting the problem with using gig workers to train AI models, the latest study shows that it is necessary to develop greater capabilities to detect AI-generated content.

The creator of ChatGPT itself — OpenAI — is one of several AI developers that is outsourcing some of its AI training to part-time, often low-wage workers.

source:Recipe for disaster — Human AI trainers are letting AI do their work (mybroadband.co.za)