You’ve probably experienced that slightly odd feeling you get when reading AI generated text, even when it’s grammatically and factually correct. That sense that something is just a little ‘repetitive’. As it turns out, one of the most distinctive tells of ChatGPT’s language in particular, is its curiously frequent use of the word “delve”.
Researcher Jeremy Nguyen found that around 0.5% of medical research articles on PubMed now contain ‘delve’, 10 to 100 times more than just a few years ago, indicating both the increasing use of the technology to generate content, and this strange linguistic phenomenon. This week a Guardian investigation looked into this and proposed an explanation.
Whilst AI models are trained on a giant mass of text data (typically trillions of words or ‘tokens’), the transformation into an assistant requires human workers to test the system, provide feedback, and even write ideal responses. However, such a labour-intensive process has a crucial drawback, cost. To make such a system economically viable, the big AI firms outsource this essential human feedback (RLHF) work to lower-cost labour markets, such as those in Africa. These workers may have unconsciously imbued ChatGPT with elements of African business English, in which words like ‘delve’, ‘explore’, ‘tapestry’, and ‘leverage’ are more commonly used compared to American or British usage.
The distinctive voice of GPT-3.5 and 4 is a testament to the hard work and linguistic diversity of its human trainers, and their role should get more attention. The use of low-cost labour to train AI models is a troubling example of how the drive for ever more powerful tech can also lead to worker exploitation. But as the field evolves, new techniques are emerging that could reduce the need for human feedback. One such approach is the use of synthetic data. Rather than relying on armies of workers to assess training examples, AI models can learn from artificially generated data that mimics real-world patterns. For us down-stream consumers, as with many other globalised products, we shouldn’t forget what goes into the AI training supply-chain.
Takeaways: Tools like GPTZero aim to detect AI content by analysing things like perplexity and burstiness, but the technology is by no means reliable, and can put non-English speakers at a disadvantage by wrongly flagging their prose. In response, tools like StealthGPT are being developed to avoid such detection by better mimicking human writing patterns and flaws. Any business relying on AI for content generation will need to consider detection methods to avoid their material being increasingly flagged as inauthentic. Those wishing to validate the authenticity of information will need to understand and carefully employ equally cutting-edge detection.
Full transparency, this article was part generated by Claude 3 Opus. We tend to get help expanding an initial idea and provide Claude with lots of examples of our writing to ensure the style and perspectives are consistent. GPTZero suggested there was an 80% chance this was human written, which is likely down to the extensive edits we then make on the initially AI generated material.