Cognitive Surrender in AI-Assisted Math: What the ALEKS Study Shows for Builders and Educators

A new study analyzing millions of student interactions on the ALEKS math platform reveals a troubling pattern: students using AI to solve word problems complete them faster but learn less. The phenomenon, called cognitive surrender, describes how learners outsource cognitive work to AI, reducing deep understanding despite faster task completion. For AI builders and educators, the findings raise critical questions about how to design tools that preserve learning value while enabling efficient problem-solving.

What happened

Researchers from UC Irvine and McGraw Hill analyzed millions of student interactions with ALEKS, an online math platform used by over four million students annually. They compared two types of problems: word problems, which are easy to copy and paste into AI chatbots, and graphing problems, which require more effort to outsource.

Beginning in early 2023, time spent on word problems fell by 31% for high school students and 27% for college students by late 2025. Time on graphing problems remained largely unchanged. In unsupervised practice, students answered more word problems correctly. But on proctored placement tests, accuracy dropped from about 80% to 60%. Performance on graphing problems did not decline.

The study, released as a June 2026 working paper titled "Faster Completion, Less Learning," cannot definitively prove AI usage. But the changes appeared only in problems easy to outsource to AI, disappeared under supervision, and grew steadily over nearly three years.

Why AI builders should care

For teams building AI tutoring tools, coding assistants, or any educational product, the study signals a design risk. When AI provides answers too easily, learners may skip the reasoning steps that build durable skills. The ALEKS study shows this is not a universal penalty across all tasks. Graphing problems, which are harder to outsource, showed no learning erosion. The damage is concentrated where AI can fully replace the cognitive step.

Related findings from Anthropic and RAND suggest this pattern extends beyond math. Students report worrying that AI is weakening their critical-thinking skills while more of them admit using it for schoolwork.

Practical implications

Designing AI that preserves learning requires intentional friction. Carefully designed AI tutors that ask questions, personalize instruction, and withhold answers until students reason through a problem have improved achievement in controlled experiments. The key is that using AI this way should increase time on task, not decrease it.

For builders, this means:

• Prompt reasoning before revealing answers.
• Require users to explain or justify AI outputs.
• Track time-on-task as a signal of cognitive engagement.
• Design for tasks that are hard to fully outsource, like graphing problems.

Caveats

The study is a June 2026 working paper that has not yet been peer reviewed. Causality cannot be definitively established; researchers could not see what else was happening on students' screens outside of ALEKS. The findings show strong associations but may not generalize to all classrooms or subjects.

FAQs

What is cognitive surrender in the context of AI-assisted math problems?

Cognitive surrender refers to the tendency to accept AI-generated outputs with minimal scrutiny, effectively outsourcing reasoning and effort to AI systems. In the ALEKS study, this appeared as students using AI to answer word problems quickly without engaging the underlying math concepts, leading to reduced learning despite faster completion.

How does AI use affect students' ability to solve word problems vs graphing problems?

Word problems are easy to copy and paste into AI chatbots, so students spent 31% less time on them by late 2025. Graphing problems require uploading screenshots and using platform tools, making them harder to outsource. Performance on graphing problems remained stable, while word problem accuracy on proctored tests dropped from 80% to 60%.

Do AI tools improve or diminish learning outcomes according to recent studies?

It depends on design. Carefully designed AI tutors that ask questions and withhold answers until students reason through problems have improved achievement in controlled experiments. But the ALEKS study shows that when AI provides answers too easily, students learn less, especially on tasks that are easy to outsource.

What is the difference between proctored and nonproctored AI-assisted assessments?

In nonproctored settings, students using AI answered more word problems correctly but spent less time on them. In proctored settings, where AI use is blocked, accuracy on word problems dropped from 80% to 60%, suggesting that the learning did not transfer. Graphing problem performance was similar in both settings.