System-regulated
(alert tips only)
No agency over AI assistance provision
Research platform overview
AI-Enabled Chess Training
A custom chess-training platform used to study when AI assistance supports learning and when on-demand help weakens the effort that builds skill.
The field experiment
A chess-training environment for studying AI help, effort, and lasting skill.
The platform was built for a 12-week field experiment with more than 200 chess-club students. Participants trained on the same custom app, which combined adaptive engine games, AI-generated tips, chess quizzes, test games, and progress dashboards.
The central experimental question was simple but consequential: what happens when learners can request solution-revealing AI help on demand, in addition to the same automated alert tips everyone receives? The answer matters because assistance that improves immediate performance can also replace the productive struggle needed for long-term learning.
12 weeks
Intensive chess training
200+
Chess-club students
2 conditions
Same alerts; optional move-reveal button
30% vs. 64%
Learning gains after training
The platform
What participants actually used
The live research app required authenticated student and coach accounts. This public version preserves the platform context, intervention logic, and paper screenshots so the study can be shared without exposing the original backend or participant data.
Train / play
Adaptive practice with instrumented AI feedback
Students played against a chess engine calibrated to their level. The app identified critical positions where an optimal move existed, delivered automated alert tips in both conditions, and tracked whether students in the self-regulated condition requested additional move-reveal help.
Study interfaces
Progress, competition, quizzes, and feedback
Beyond training games, the platform gave students a profile page, leaderboard, chess quizzes, and post-quiz feedback. These paper figures show how learning activity, rankings, and knowledge checks were presented during the field deployment.
Experimental conditions
Same automated alert tips, one additional button
Alongside their regular coach-led training, students were randomly assigned to two conditions that used the same platform, the same engine calibration, the same incentives, and the same automated alert tips. Both conditions received alerts and post-move feedback at algorithmically identified positions. The self-regulated condition added one option: a button that could reveal the best move on demand.
Self-regulated
(alert tips + button to get on-demand move-reveal tips)
Agency over AI assistance provision (when and how often to click the button)
Alert tips and feedback
Guidance before and after the student's move
These automated alert tips were present in both experimental conditions. They signaled that the position contained an important opportunity. After the student moved, feedback either confirmed that the optimal move was found or revealed the missed optimal move with an explanation.
Move-reveal tips
The extra on-demand help in the self-regulated condition
This was the only intervention difference. Self-regulated students kept the same automated alert tips and feedback, but also had a button that could reveal the best move on demand. This added control made assistance easier to access, but also made it easier to skip productive search.
Key findings
More control over AI help did not produce more learning
Both groups improved, but students with on-demand access to AI help learned less than half as much as students whose assistance was system-regulated.
System-regulated learning gain
Students whose help was governed by the platform made substantially larger gains after the 12-week training period.
On-demand learning gain
Students who could request extra AI help whenever they wanted improved by less than half as much, despite having more access to assistance.
Productive struggle was displaced
The learning loss was concentrated when students requested help on tasks inside their Zone of Proximal Development: difficult enough to require effort, but still achievable with the right support.
Engagement fell as reliance rose
Students with on-demand help completed fewer training games, reported lower accomplishment, and increased their AI requests over time even while recognizing the risk of over-reliance.