AI detector identifies machine-generated research ideas with 91% accuracy
IdeaLens analyzes the structure and flow of an argument rather than the wording of a paper, but researchers warn that the tool can produce errors.
Credit: EvgeniyShkolenko/Getty
Researchers have developed a detector designed to distinguish ideas created by humans from those generated by artificial intelligence. The tool, called IdeaLens, could help measure AI use in conference submissions and scientific journal papers, according to its developers.
Rishanth Rajendran, a computer scientist at the University of Maryland, College Park, and a co-author of the preprint, says IdeaLens could be useful if applied carefully to conferences and journals that are “overwhelmed” with artificially generated content.1
Using AI to write scientific papers is becoming increasingly accepted, Rajendran says, provided that “the core idea is the researcher’s own.”
How IdeaLens detects AI-generated ideas
Unlike many AI detectors, IdeaLens analyzes the summary or outline of a work rather than its exact wording. To use it, researchers first extract and outline the text, such as a scientific paper, and remove its formatting and style.
The system then searches for patterns associated with human- or AI-generated ideas, including the structure of the discussion and the flow of the argument.
AI-generated ideas detected 91% of the time
In testing, the detector identified AI-derived ideas 91% of the time. Nikhil Garg, a computer scientist at Cornell Tech in New York City, described the result as impressive.
“I was surprised that AI could detect an argument written by an AI just from the outline,” Garg says.
Researchers warn that the detector can be unreliable
Garg cautions that IdeaLens is error-prone and should not be used to decide whether an individual piece of work was created by a human or by AI. Applying the tool to a different type of data from the material used to train it could also produce inaccurate results.
“What AI ‘tells’ us in one domain may simply mean humans are better at writing in another,” he says.
Human and AI writing differ in structure
The research group’s earlier work found that human-written stories have a different structure from stories written by AI. The researchers also found that different chatbots tend to use the same arguments when writing essays.
Together, these findings suggest that it may be possible to develop detectors that identify whether an idea was generated by AI, Rajendran says.
Source: www.nature.com