AI Trends

Trending story

University of Maryland and Google DeepMind propose IdeaLens

1 reports1 story updatesUpdated Nov 17 18:55

Understand this story

AI overview

Researchers from the University of Maryland and Google DeepMind proposed the IdeaLens algorithm in a preprint that has not undergone peer review. By analyzing the argumentation and content structure of a text, it determines whether an idea came from a human or a large language model, and it can be tried on their website. The study trained on approximately 800,000 texts annotated by Pangram; in testing, for texts conceived by humans with AI participating in the writing, Pangram judged up to 92.4% as AI-generated, while IdeaLens reduced this to as low as 7%; for texts in which humans wrote based on AI-generated structures, IdeaLens identified the AI origin in 68% of cases, while the Prosa detector identified less than 10%. Independent researchers noted that the system is not suitable for "co-authored" scenarios of human-AI joint ideation, and that the basis and weights of its judgments are not transparent.

AI-generated from public reports

Latest development · Report

University of Maryland and Google DeepMind propose IdeaLens, attempting to identify whether ideas in text come from ChatGPT

Developments

1 confirmed developments

  1. University of Maryland and Google DeepMind propose IdeaLens

    Source:heise online

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 17

  1. heise online

    University of Maryland and Google DeepMind propose IdeaLens, attempting to identify whether ideas in text come from ChatGPT

    Researchers from the University of Maryland and Google DeepMind propose IdeaLens in a preprint that has not undergone peer review—an algorithm that determines whether ideas come from humans or large language models by analyzing the argumentation and content structure of text, and it can be tried on its website. The study trained on approximately 800,000 texts annotated by Pangram; in tests, for texts conceived by humans with AI participation in writing, Pangram judged up to 92.4% as AI-generated, while IdeaLens dropped to as low as 7%; for texts written by humans based on AI-generated structures, IdeaLens identified the AI origin at a rate of 68%, while the Prosa detector was below 10%. Independent researchers point out that the system is not suitable for the "co-authored" scenario of human-AI joint conception, and that the basis and weights of its judgments are not transparent.

Related stories

  1. Related storiesGoogle DeepMind paper finds RL amplifies the gap between supervised and unsupervised model behavior
  2. Related storiesTomorrow.io Reports That Improved Initial Atmospheric Estimates Can Enhance AI Weather Forecasting
  3. Related storiesGoogle DeepMind uses AlphaProof Code to formally verify the Signal protocol
  4. Related storiesApple research team submits paper to arXiv disclosing the HALO system