How to detect text which has been written by ChatGPT (www.rte.ie)

Researchers and journalists outline how attempts to spot ChatGPT-written text fall into two camps: local detectors that judge individual documents and global analyses that hunt for linguistic trends across large corpora. Practical local signals include obvious giveaways (made-up citations or “as an AI language model”), statistical measures such as perplexity (AI text tends to be more predictable/low‑perplexity), watermarking schemes that embed subtle signals in generated output, and machine‑learning classifiers trained to pick up generation patterns. Global methods compare pre/post‑2022 language use or contrast known human vs. known AI text to flag spikes in particular words, phrases or syntactic habits. The story’s significance is the fragility of detection: none of these approaches are currently reliable enough for high‑stakes use because of false positives and an evolving adversary. Models and prompts change frequently (a recent Washington Post analysis of ~300k ChatGPT messages from June 2024–July 2025 found “delve” declining while “core,” “modern,” and emojis—notably 🧠 and ✅ in 70% of messages—are rising), so detectors can be evaded by model updates or user instructions. That creates an ongoing arms race with implications for academia, journalism and policy: detectors must account for model drift, cultural adoption of AI-influenced phrases, and the real risk of misattributing human writing as machine‑generated.