Artificial intelligence (AI) is rapidly changing education, creating both new opportunities and new challenges for teaching and learning. As AI capabilities continue to evolve, teachers, researchers, and families are asking what role AI should play in schools and where human expertise should remain central.
The AugmentED program at AERDF is built on a specific high-stakes bet: that AI’s most valuable function in education is not to replace human judgment, but to enhance it. AI should work alongside teachers and students to build exactly the kinds of durable skills an AI-saturated world will require more of, not less. Skills like critical thinking.
This is the theory of change driving all AugmentED’s work. AugmentED brings together expert educators, researchers, and technologists to reimagine education for the AI era, developing new teaching approaches and AI-powered tools to help students learn durable skills.
As part of this work, AugmentED is investigating which roles AI can appropriately play in the classroom. One of the roles AugmentED is testing is assessment, with AI helping educators see how students’ durable skills are developing. In a new research publication, they examine how far AI can be trusted with that job, beginning with detecting critical thinking in student writing.
Why Study Critical Thinking?
Critical thinking is an essential skill in the AI era. As generative AI makes it easier to produce misleading information, students need to learn how to evaluate evidence, identify flawed reasoning, blend multiple perspectives, and draw well-supported conclusions.
Despite its importance, critical thinking remains difficult to measure in classroom settings. Unlike many academic skills with clear right or wrong answers, critical thinking is nuanced and multidimensional. Developing students’ critical thinking skills requires timely assessment and meaningful feedback, but providing that consistent level of support can be difficult.
AugmentED’s latest study explores whether advances in large language models (LLMs) can help address this challenge while also identifying where current AI systems still fall short.
Exploring AI’s Ability to Measure Critical Thinking
The study brought together leading AI and education researchers to answer one question: Can AI help measure subskills that underlie critical thinking?
The team developed and evaluated AI-powered methods for assessing critical thinking demonstrated in argumentative essays written by students in grades 6 through 12 using the PERSUADE dataset. Building on the Skills for the Future critical thinking skills progression, the team created a rubric with definitions of different levels of proficiency for several critical thinking subskills, including:
- Evaluating evidence
- Synthesizing information
- Using counterarguments
- Drawing conclusions
- Logical reasoning
The team then compared how expert human raters evaluated students’ essays with how multiple LLMs assessed the same work. Both the humans and LLMs scored the essays against the rubric. Rather than asking whether AI could replace human judgement, the study examined where AI can best support assessment and where additional research and safeguards are still needed.
Key Findings From the Research
AugmentED’s research offers encouraging evidence that AI has the potential to support assessment of certain critical thinking skills. It also highlights why rigorous validation is needed before these tools are used in educational settings.
Understanding both the capabilities and limitations of today’s AI models is essential for ensuring these technologies are used responsibly in education so that students receive accurate, meaningful feedback.
Key findings include:
- AI shows promise in assessing certain critical thinking skills:
The research demonstrates that LLMs can get better at measuring certain critical thinking subskills when they are intentionally designed for educational use. By scaffolding LLMs with critical thinking definitions and rubrics, and applying methods such as few-shot prompting and fine-tuning, AugmentED was able to improve model performance beyond what these systems could achieve “out of the box.” - More research is needed to understand AI’s limitations:
The research showed that LLMs currently perform poorly on measuring some aspects of critical thinking.The models struggled with more nuanced forms of reasoning and with distinguishing between similar levels of student proficiency. Even with scaffolding, there were critical thinking subskills where the models consistently fell short.
Understanding these limitations is as important as identifying AI’s strengths. Identifying the gaps cautions edtech builders against mindlessly putting AI into education tools and assuming it can do certain things it currently can’t do.
- Human oversight remains essential:
LLMs can produce outputs that sound convincing even when they are incorrect.
Without evidence of what AI can actually do well, there is a risk of integrating AI into educational tools in ways that provide inaccurate assessments or feedback to students.Rather than assuming AI is ready to evaluate complex skills, the field should continue building the research base that defines where AI adds value, where human expertise remains central, and how these technologies can be responsibly integrated into teaching and learning.
Building the Evidence Base for Responsible AI in Education
This publication represents one piece of AugmentED’s broader mission to build the evidence, infrastructure, and safeguards needed for responsible AI in education.
As AI becomes more deeply integrated into teaching and learning, research must continue to identify not only what these technologies can do well, but also where human judgment remains indispensable. By rigorously evaluating both the capabilities and limitations of AI, we can help ensure these tools strengthen education, rather than simply automate it.
Download AugmentED’s new research publication here to explore the complete methodology, findings, and implications for researchers, educators, and developers working at the intersection of AI and education.
