Key points
- Perplexity and watermarking help identify AI writing
- False positives remain a significant challenge
- Spikes in word usage suggest AI influence
- ChatGPT’s favourite words and emojis are changing
ISLAMABAD: It has been three years since ChatGPT was released, transforming industries reliant on writing and reading. Shortly after its launch, efforts to detect AI-generated text began, with methods split into local approaches, focusing on individual articles, and global approaches, which track broader linguistic trends in AI-written texts.
Some clear indicators of AI-generated text include fabricated references or phrases like “As an AI language model”. However, more advanced methods are often required. These methods include perplexity, a statistical measure that assesses how predictable a sequence of words is. AI-generated text tends to be more predictable, showing lower perplexity compared to human-written text.
Other techniques involve watermarking, where a hidden signal is embedded in the text, or machine learning algorithms that identify patterns distinguishing AI and human-written texts. While improving, none of these methods are sufficiently reliable for practical use, often due to false positives, such as a student wrongfully accused of using AI for an assignment.
Finding words
The global approach focuses on finding words, phrases, or syntactic patterns linked to AI writing. One method compares texts written before and after 2022, looking for unexplained spikes in word usage.
For example, the phrase “I rise to speak” has become more popular in British political speeches, though the reason remains unclear, suggesting AI influence. Additionally, the word ‘delve’ has been notably favoured in scientific writing by ChatGPT.
However, tracking these changes is challenging as AI models continuously evolve. For instance, a recent Washington Post study found that ChatGPT’s use of ‘delve’ is declining, while words like ‘core’ and ‘modern’, as well as emojis such as 🧠 and ✅, are rising in popularity.
Ultimately, the challenge remains: how to reliably identify whether a piece is AI-generated or human-written. We must continue exploring this complex issue.
