AI Leaves Clues: New Watermark Could Expose Machine-Written Schoolwork in Trinidad and Tobago Classrooms

As the new school year unfolds across Trinidad and Tobago, educators are confronting a rapidly growing challenge: assignments produced with artificial intelligence. Students increasingly turn to AI tools for essays, prompting teachers to seek reliable ways to spot machine‑generated work.

Anthropic, the U.S. firm behind the Claude chatbot, announced that models launched after August 2 now embed an invisible statistical watermark in their output. The measure satisfies transparency obligations under the European Union’s AI Act and will be rolled out worldwide. Rather than a visible logo or metadata, the watermark is a subtle pattern created through low‑stakes word choices.

The technology draws on SynthID‑Text, originally developed by Google DeepMind. Google has already used SynthID to watermark text generated by its Gemini platform since 2024. In contrast, OpenAI currently applies provenance signals only to AI‑generated images and audio, leaving ordinary ChatGPT text unmarked.

Anthropic cautions that the watermark can only estimate the probability that Claude contributed to a passage; it cannot confirm that an entire assignment is AI‑written. Detection struggles with short answers, factual excerpts, code snippets, and lightly edited content, and extensive rewriting or translation can diminish confidence. No single detector can reliably identify text from every AI model.

For teachers, a watermark detection should serve as a clue rather than a verdict. Clear classroom policies, knowledge of each student’s typical writing style, and follow‑up discussions remain essential to address potential misuse while preserving academic integrity.