收听本期播客
阅读正文
In an unexpected turn of events, it appears that the individuals employed to refine and improve artificial intelligence systems are becoming their most significant critics. A worrying development shows that many of these workers now distrust the technology they help build and are urging their own friends and family to exercise caution when using such tools.
The case of Krista Pawloski, an AI content moderator, is a compelling example. A defining moment came when she was tasked with labelling social media posts for offensive content. She narrowly avoided approving a post that included the term ‘mooncricket’, only discovering it was a serious racial slur after she decided to investigate further. This near-miss led her to a profound realisation about the countless similar errors she and thousands of her colleagues might have made unwittingly. Consequently, she has ceased her personal use of generative AI and advises her family to do the same.
Pawloski’s concerns are widely shared among other ‘AI raters’, who work for major tech corporations. After witnessing the inner workings of these systems, many have developed a deep-seated distrust. A fundamental issue they identify is that workers are frequently assigned to evaluate complex subjects, such as medical advice or sensitive historical events, without possessing any expert knowledge in these fields. This practice leads directly to the principle of ‘garbage in, garbage out’, which suggests that if the human feedback guiding the AI is flawed, the AI’s output will inevitably be unreliable.
According to these insiders, the root of the problem lies with the priorities of the tech giants. They report a relentless emphasis on developing and releasing products as quickly as possible, often sacrificing accuracy and quality for speed and profit. Raters are allegedly given vague instructions and unrealistic deadlines, meaning that crucial, detailed feedback may be overlooked. The result is the creation of AI models that can confidently dispense false information.
Ultimately, those working closest to the technology perceive its fragility rather than its futuristic promise. They understand that an AI is only as good as the data it is trained on, a process which is often far from perfect. By sharing their insights, they hope to make the public more aware of the technology’s current limitations, raising a critical question for us all: if the experts are losing faith, how much trust should we be placing in these tools?