A new study published in the MIT Technology Review has found that a significant portion of people paid to train AI models are themselves outsourcing their work to AI. Irony aside, this practice makes it likely to entrench AI mistakes and hallucinations, because mistrained AI will then be compounding that misinformation by training other AI models.
To take a step back, training artificial intelligence requires huge amounts of human intelligence, and labour. Workers are paid to complete mundane tasks – identify objects in images, tag and annotate text, label data – which is then fed into large models to train these AI.
Workers can sign up to complete these tasks on platforms like Amazon’s Mechanical Turk. So the Swiss Federal Institute of Technology hired 44 people to complete work and then analysed their work for telltale signs of using ChatGPT. They estimate that somewhere between 33% and 46% of the work had been done by AI.
Using AI-generated data to train AI could introduce further errors into already error-prone models. Large language models regularly present false information as fact. If they generate incorrect output that is itself used to train other AI models, the errors can be absorbed by those models and amplified over time, making it more and more difficult to work out their origins |
This is just your latest reminder to be skeptical of AI’s answers. As we all play around with AI and find new ways to incorporate it into our working lives it is critical to check your work. There are way too many of examples of AI hallucinating and getting it wrong, and studies like this suggest that isn’t going to be changing any time soon.
This is an excerpt from our Thought Starters email. Once a week we send you 5 interesting articles that have caught our attention, to get you thinking. No spam, we guarantee.