A new AI detector looks beyond words to check whether the ideas came from AI
[post_content]
Disclaimer: This article has been automatically aggregated from
Most AI detectors look for telltale signs of AI-generated text by examining word and sentence patterns. A new tool called IdeaLens takes a different route to solving this problem. It spots AI-generated research by looking at the ideas behind it instead of the wording itself (via TechXplore).
Researchers from the University of Maryland, Google DeepMind, and Simon Fraser University built IdeaLens to check whether a document’s core arguments came from a human or a machine, regardless of who wrote the final draft.
How does IdeaLens detect AI-generated ideas?
According to the research paper, the team trained IdeaLens on around a million documents already labeled as AI-written or human-written. They used language models to strip away the original phrasing and reduce each document down to a simple outline of its core arguments. This forced the system to learn from ideas instead of writing style.
IdeaLens correctly flagged 91% of AI-generated work when tested on documents written entirely by one source or the other. It misidentified human writing only 0.6% of the time, though that still falls short of established detectors like Pangram.
When researchers tested writing that received minor AI edits, IdeaLens classified 95% of human-generated documents as human, compared with Pangram’s average of 44%.
Can IdeaLens reliably identify who came up with an idea?
The results were interesting in cases where human and AI ideas were mixed together. When AI wrote the text based on a detailed human outline, traditional detectors like Pangram caught 92% of that AI involvement, but IdeaLens flagged just 7%, since the ideas themselves were human. Now flip that scenario, where a human wrote the words from an AI-generated outline, and IdeaLens caught 68% of AI influence while Pangram caught only 8%.
Researchers found that human generated ideas tend to have more concrete details, verifiable sources, and acknowledged limitations. Whereas AI-generated ideas usually have generic details, unverifiable numbers, sweeping claims, and suspiciously perfect answers.
However, this AI detector is not yet ready to settle authorship disputes entirely. IdeaLens still struggles with some newer AI models, so the team says it needs more testing before it’s ready for wider use in journals and classrooms.
RELATED COVERAGE:
for informational purposes only. We do not claim ownership, accuracy, or liability for the content provided. All rights belong to the original publisher.