That is not a slogan. It is what the studies now show: sustained use of generative AI that optimizes for convenience is associated with weaker memory, reduced critical thinking, and a diminished sense of authorship. NOVARC's research program exists to build the opposite — models, agents, and instruments engineered so that using them makes people more capable.
Independent groups, different methods, one direction: the more thinking is delegated, the less capacity remains. Two studies anchor the pattern.
MIT Media Lab · 2025
Kosmyna et al.
Participants who used LLMs to assist with writing showed reduced neural connectivity, weaker memory of their own work, and a diminished sense of authorship over the output they produced.
“Cognitive debt — capability lost in exchange for short-term efficiency.”Read the study ↗
Microsoft × Carnegie Mellon · 2025
Lee et al.
Higher confidence in AI was associated with reduced critical thinking; higher self-confidence in one's own ability was associated with more of it.
“The more people trust the system, the less they exercise the faculties it is meant to support.”Read the study ↗
Any system that interfaces with human cognition should be able to demonstrate that it strengthens the faculties it engages — the standard we already apply to medicine, education, and physical training. NOVARC holds its own work to that standard.
The brain strengthens through challenge, not through service. NOVARC's approach rests on established results from cognitive science.
The learning signal is the gap between what we expected and what actually happened. When reality matches the prediction, nothing updates. When it doesn't, the brain changes.
Systems that answer instantly close that gap by design. No surprise, no learning.
Wolfram Schultz · University of Cambridge · The Brain Prize, 2017
Learners who struggle with a problem before being shown the solution develop deeper conceptual understanding than those given the solution upfront.
Tools that hand over finished answers remove the struggle understanding is built from.
Manu Kapur · ETH Zürich · Learning Sciences
Humans are systematically overconfident, and overconfidence correlates with worse learning. Targeted calibration training reliably improves metacognitive accuracy.
AI that always agrees produces confidence without competence.
Schraw · Dunlosky · Bjork · Metacognitive monitoring research
The program spans the stack — from how models reason to how people change through using them. Every area answers to the same question: does this leave the person more capable?
Select an area to read it
How models think, fail, and can be made trustworthy.
Agents that act with people, keeping judgment human.
Systems that strengthen human reasoning through use.
Interaction designs that add capability instead of dependence.
Evaluating whether AI leaves people more capable, not just faster.
The first instrument ships in 2026.
