The Rise in AI Writing
By Sherry Shu
Can you tell the difference between text written by humans and by artificial intelligence (AI)? With the recent release of language generation models like ChatGPT (the Chat Generative Pre-trained Transformer), this question is becoming increasingly unclear.
https://www.markettailor.io/blog/how-businesses-can-use-chatgpt-for-content-marketing
ChatGPT is a state-of-the-art language generation model developed by OpenAI. It has revolutionized the way we interact with AI by providing human-like responses to natural language queries. The model has been trained on an immense corpus of text data and can generate high-quality text in a wide range of styles and genres.
One of the most impressive features of ChatGPT is its ability to mimic human writing. The text generated by the model is often difficult to distinguish from text written by a human. This has led to a growing concern among experts about the potential dangers of AI-generated text, particularly in fields such as journalism, where the line between human-written and AI-generated text is becoming increasingly blurred.
…Now, could you tell that the past two paragraphs were entirely written by ChatGPT?
Yeah, me neither. Don’t worry, I promise that I’ll be the one writing from now on!
Even though AI-generated text now seems eerily similar to human writing, a few AI-powered solutions have been developed. Princeton student Edward Tian created GPTZero over the winter, an application made to estimate the likelihood that a passage of text had been written by artificial intelligence. The system makes the prediction by calculating a text’s perplexity (language complexity) and burstiness (variation in sentence lengths). A few days ago, OpenAI released its own AI text classifier to accomplish the same task.
Unfortunately, these systems haven’t completely solved the problem. For instance, the OpenAI classifier only correctly identified AI-generated text 26% of the time. 9% of the time, it would incorrectly classify human-written text as being written by AI. Both systems have primarily been trained in English, meaning that they are not very useful for prediction in other languages. In addition, both systems require that inputted text meet a minimum length (250 characters for GPTZero, and 1,000 characters for the OpenAI text classifier), which allows a significant portion of text online to go undetected.
So, what does this all mean for the future of AI writing? Professor Mor Naaman of Cornell University warns that AI technological development is “racing ahead with no guardrails.” He argues that language generation models can continue to train off new databases of text and improve their ability to create text that evades classifier detection. This makes services like ChatGPT vulnerable to abuse from malicious actors, without any awareness or accountability from the public (e.g. cybercriminals have already generated malware using ChatGPT). He encourages schools, social media sites, and publishing platforms to create policies governing the use and distribution of AI-generated text. He stresses the importance of labelling and distinguishing between the two to convey honesty and awareness to readers.
The future of writing in the post-AI world is still deeply unclear. AI could be a tool that helps writers, journalists, and students complete their tasks with greater information or a tool that destroys the uniqueness of human writing. Regardless of the end result, the hope is that technological development and policymaking can help us preserve a world where human voices still can be heard among the buzz of machines.
