Dive Brief:
- Artificial intelligence appears somewhat helpful at measuring certain sub-skills related to critical thinking when intentionally designed for educational use, but AI does less well in assessing other building blocks, according to new research on how the technology rated argumentative essays written by students in grades 6 to 12.
- Accounting for both of those variations in proficiency and the fact that AI often sounds convincing even when it’s wrong, further study is needed before it should be implemented in this regard, according to the research conducted by AugmentED, a program of the Advanced Education Research and Development Fund.
- Despite being trained on examples of how humans rated the various core components of critical thinking, with fine-tuning and additional prompting, no sub-skill of critical thinking was measured at better than 78% accuracy, according to research led by AugmentED.
Dive Insight:
“The big takeaway is that it’s not ready for prime time, to engage in all of the sub-skills of critical thinking,” said Sherry Lachman, founder and executive director of AugmentED. “Much work needs to be done … before we can be confident that it’s ready.”
Lachman draws two lessons from the research findings. “Rather than assuming AI is a genie in a bottle that is infinitely able to do anything, you have to validate what it can and can’t do in classrooms for educational purposes,” she said.
“The other thing it points to is, specifically within those critical thinking spaces, where there is additional work to be done. We’re not saying, ‘Close the door.’” Rather, the research “points to where there is additional work for the field to do, to measure AI’s ability to measure critical thinking,” Lachman added.
While not everyone agrees precisely on the definition of “critical thinking,” the AERDF research team used a rubric developed by ETS and the Carnegie Foundation that outlines six critical thinking sub-skills, Mills said. The results showed that AI performed highest when evaluating whether students synthesized information from multiple sources and when evaluating evidence strength. AI did less well assessing the other four sub-skills: using counterarguments, distinguishing facts and opinions, drawing conclusions, and seeing logical fallacies.
To ensure AI strengthens students’ critical thinking skills, rather than outsourcing those to AI, “You have to be careful to match your use of AI to your goal,” Lachman said. “In the age of AI, durable skills like critical thinking are not less essential, but even more essential for kids to thrive."
Lachman added, "There’s a real risk with AI that if you use it in particular ways, for example if a student uses it to write their essay without ever learning to write an essay on their own … it will short-circuit students’ ability to ever develop skills like critical thinking.”
But an AI tool that measures students’ prior knowledge and interest to deepen their reasoning would be pedagogically beneficial, she said.
Mills urged educators to ensure transparency in the way AI is being used by the student and educator.
“Don’t take the human out of the loop,” she said. “Failure is part of the process. Taking feedback is part of the process. Having to grapple with hard things is part of the process. We don’t want to remove that friction. That’s also what prepares [students] for the real world.”
While she would never say educators shouldn’t trust AI under any circumstances, “When you get to things that are subjective, like durable skills, the real problem is that we don’t have a validated model for pedagogical purposes,” Mills added. “It could not only be inaccurate but also harmful, not just not aligned with pedagogical concepts but also telling the student wrong information.”
The AugmentED study needs to be followed by further research not only to validate how well AI measures students' critical thinking skills but for whom it works well, Lachman said.
“Even if it works well 80% of the time, what about the other 20%?” she said. “What about students who the data are not as well trained on, like Black students or low-income students? We’re not saying you should never use AI. We want it to be used for valid pedagogical purposes.”
