A deeply concerning aspect of AI training is emerging: the models are not just being taught facts, but are also learning to replicate falsehoods, and the human trainers feel powerless to stop it. Due to flawed processes and perverse incentives, the very people hired to ensure accuracy are sometimes forced to approve or create content they know to be incorrect, effectively teaching the AI to lie.
This happens in several ways. Firstly, the pressure to meet deadlines means trainers often don’t have time to verify information. Faced with a choice between approving a plausible-sounding “hallucination” or being flagged for low productivity, many will choose the former. This injects false information directly into the AI’s training data.
Secondly, the system that forces non-experts to vet specialized content is a recipe for error. A trainer who knows nothing about a subject may inadvertently approve a subtle but significant inaccuracy, teaching the AI that the incorrect information is correct. Their lack of expertise becomes a permanent part of the AI’s “knowledge.”
Most disturbingly, some workers report being instructed to essentially make things up to fulfill a task’s requirements. This creates a situation where the AI is not just learning from human error, but from deliberate human fabrication, all in the name of completing a job. The result is an AI that is not only unreliable but is being systematically trained in the art of confident deception.