−17%
The Crutch Effect
Bastani et al., Wharton / PNAS 2025
In a randomized controlled trial of ~1,000 high school math students, those who used ChatGPT freely scored 48% better during AI-assisted practice — then 17% worse on the actual exam when the AI was removed. Their brains had been conditioned to offload thinking rather than build it.
A second group used an AI tutor with guardrails — one that gave hints instead of answers. They scored 127% better during practice and retained those gains on the test. Same underlying model. Completely different design. Opposite outcomes.
Our response
DecodeLM never gives you the answer directly. When you ask a question, the tutor asks one back — guiding you to construct the understanding yourself. The knowledge that results is yours.
2×
Socratic AI Beats the Best Classrooms
Kestin et al., Harvard / Scientific Reports 2025
194 physics students were randomized into two groups: AI-tutored vs. active learning classrooms (the gold standard of human teaching). The AI tutor produced learning gains more than double the classroom group. Students also reported higher engagement and motivation.
The same researchers found that unguided ChatGPT use — giving students access without structure — made students worse. The technology was identical. The design was the variable.
Our response
The Harvard tutor was designed around one rule: never reveal the full solution; give one step at a time and make the student try. That is the rule DecodeLM's tutor follows by default.
Struggle first
Productive Failure
Kapur, 2008–2024 · Sinha & Kapur, Review of Educational Research 2021
Across 15+ years of research at ETH Zurich and NTU Singapore, Manu Kapur found a consistent pattern: students who attempted a problem before being taught the solution learned significantly more deeply than students who were taught first. Failure activates prior knowledge, surfaces gaps, and creates 'readiness to learn' that makes the correct answer land differently.
Traditional instruction and most AI tools eliminate this struggle. They optimize for getting to the answer quickly. Cognitive science says this is exactly backwards.
Our response
DecodeLM is designed to let you struggle productively. The tutor will not rescue you from difficulty — it will help you navigate it. The gap between 'I don't know' and 'I understand' is where learning actually happens.
30–50%
Retrieval Practice Beats Re-Reading
Adesope et al., 2017 meta-analysis · Rowland, 2014 (g = 0.50)
Testing yourself improves long-term retention by 30–50% over passive re-reading. The mechanism is called the 'testing effect': the act of retrieving information from memory strengthens the neural pathways that store it. Reading or watching passively does not.
Most AI study tools give you summaries and explanations — optimizing for how easy the session feels. Easy learning creates the illusion of understanding. Effortful learning creates durable knowledge.
Our response
DecodeLM builds spaced repetition into the tutor loop. It tracks what you've learned and revisits concepts at scientifically-timed intervals — making you retrieve, not re-read.
Neural degradation
What Unrestricted AI Does to the Brain
Kosmyna et al., MIT Media Lab 2025
MIT researchers put EEG monitors on 54 participants across 4 months, split into three groups: ChatGPT users, Google Search users, and no-tools. ChatGPT users showed systematically weaker brain connectivity. When the tool was removed, they could not recover — their brains had been conditioned to lower cognitive effort. They also couldn't accurately quote essays they had written with AI assistance.
The no-tools group showed the strongest, most distributed neural networks. The difference was not talent — it was the tool's design. ChatGPT optimizes for producing an answer; it does not require the user's brain to do anything.
Our response
Every interaction with DecodeLM is designed to require cognitive effort. The tutor asks before it tells. It checks whether you understood before moving on. It builds a model of what you know — and what you don't — so it can find the edge of your understanding and push there.
