Albawaba - As the race in generative AI accelerates, discerning between machine-written and human-written content emerges as a critical challenge for the technology industry. Novel AI services like ChatGPT, GPT-4, and Google Bard exhibit remarkable proficiency in crafting convincing written material, leading to a dichotomy of positive and negative implications. While these technologies expedite software code creation, they also have the potential to propagate factual inaccuracies and misinformation.
Recognizing the significance of this endeavor, OpenAI, the pioneering creator of ChatGPT and GPT-4, unveiled a "classifier to distinguish between text written by a human and text written by AIs from a variety of providers" in January. Despite acknowledging that it is challenging to achieve perfect accuracy in detecting all AI-written text, the implementation of robust classifiers is crucial in addressing numerous problematic scenarios. Such scenarios include combating false claims of AI-generated content being human-authored, tackling automated misinformation campaigns, and preventing the exploitation of AI tools for academic dishonesty.
Unfortunately, less than seven months later, OpenAI had to discontinue the project due to the classifier's subpar accuracy. In a recent blog post, the company expressed its commitment to incorporate feedback and explore more effective techniques for determining the provenance of text.
The implications of this development are far-reaching. If OpenAI, a leading entity with Microsoft's support, struggles to identify AI writing, it raises concerns about the wider online information landscape. Already, spammy websites employ AI models to churn out automated content, propagating deceptive narratives and generating ad revenue. The blending of AI-produced data with training new models is another worry, as researchers fear a phenomenon termed "AI Model Collapse." By utilizing model-generated content in training, subsequent models develop irreversible defects, as highlighted in recent research.
To avert the dangers of the AI Model Collapse and safeguard the benefits of large-scale data from the web, researchers stress the significance of preserving genuine human interactions in datasets. The value of such data becomes increasingly crucial in the face of proliferating content generated by large language models (LLMs) scraped from the internet.
The crux of the issue lies in the inability to distinguish between human and machine-written content. This poses an existential problem that must be addressed promptly to ensure the responsible and ethical use of AI-generated text online. Seeking insight from OpenAI regarding their discontinued AI text classifier and its implications, including Model Collapse, yielded limited response, leaving the industry at a crucial juncture in its pursuit of solutions.
