I recently updated my commuting bike. I had the choice to select an electric bike. Electric bikes are very popular. You can bike longer distances, sweat less, and carry more. It’s a no brainer — choose the electric bike?
No!!!
Choosing wisely means choices that align with my values and goals. I value and want to become stronger and healthier. Therefore, I chose to get a regular bike because I want to expend more physical effort and become stronger and healthier.
RULE: “Effort is the algorithm.” Improvements in cognitive, emotional, behavioral, and social (CEBS) learnings and capabilities must go through expenditures in effort. This is called productive struggle.
Changes in CEBS capabilities are mediated by the effort-activated rewiring of our brains. Unfortunately, through cognitive offloading,1 generative artificial intelligence (GenAI) can lead to decreases in cognitive capabilities including reasoning, critical thinking, and problem solving. Naturally, we should be very concerned about the adoption of GenAI tools and their potential negative impacts on our human capabilities, especially for our youth.
Figure 1 summarizes how I think about GenAI impacts along three key dimensions:
Value (addressing an important need for self, client, patient, community, etc.)
Cognitive effort (required for learning, improving, and brain rewiring)
Human capability (without GenAI assistance)
Combined capability, human + GenAI, leads to increases in productivity, but not necessarily increases in human capabilities when the GenAI is removed. Agency Gap is the decrement in human capability when GenAI is removed. This can be avoided if we use GenAI wisely.
When we use GenAI, our goal should always be to deliver positive value. In Figure 1, delivering positive values occurs in the top four quadrants. For example, at work, GenAI can increase your productivity and contribute to the organization’s mission. However, because of cognitive offloading, GenAI use can also lead to metacognitive laziness and dependence on GenAI.
Figure 2 lists examples of AI use that delivers positive value and impacts effort and capability. On the road, Google Maps enables convenient GPS navigation to geographic locations, decreases cognitive effort, but also decreases the capability of deploying and reading paper street maps.
At work, a GenAI deep research report saves time, decreases cognitive load, and contributes to the organization knowledge base. Unfortunately, what is good for the organization is not necessarily good for the employee’s intellectual development.
At home, a high quality AI foreign language app uses learning science to customize its interaction with the individual learner. The app increases the user’s cognitive load in a way that optimizes learning (eg, space repetition, comprehensible input, speaking, writing, dialog).
The impact of coding assistants (for Julia, Python, R, etc.) vary depending on how the user interacts with GenAI. Figure 3 depicts this interaction. GenAI can automatically provide coding suggestions (“proactive”) or wait for the user to seek assistance (“reactive”). In general, avoid proactive interaction.2
Ideally, user interaction with GenAI should be reactive: seeking evaluation, scaffolding, or information. Within reactive, we should prioritize
Evaluative-seeking, and
Scaffolding-seeking.
Both of these strategically INCREASE our cognitive effort directed toward learning and increasing our human capability independent of GenAI; in other words, there is no agency gap. This is what we want.
Evaluative-seeking is a best practice. We elicit critical feedback and then we work on fixing it — not the GenAI — us! Have the GenAI adopt different personas or perspectives (eg, play Devil’s Advocate). Even better: we should also elicit critical feedback from human partners. It’s good for us and our human partners!
Answer-seeking should be used sparingly or strategically. Gathering information or conducting AI deep research can be integrated into a high quality, decision making process (see Decision Intelligence 4 Health).
Figure 4 is the bottom 4 quadrants of negative value (Figure 1). Keep these empty. Displayed is an example of academic cheating by having GenAI write a school paper. Not only is this unethical (ie, negative value), but it also decreases human capability (agency gap) by cognitive offloading and avoiding productive struggle.
Figure 5 graphically depicts the ideal relationship between GenAI assistance, human capability (learning and improvement), and human cognitive effort. We want
AI to increase our cognitive effort (it rewires our brain), and
AI to optimize the human effort-capability curve to learn more faster in a way that is customized to our individual needs.
Done correctly, AI-human interaction will shift the effort-capability curve upwards. We learn and improve more per unit of cognitive effort. We want the right type of cognitive loading.
So far I have been covering aspects of pedagogical literacy. Generative AI literacy identifies four intersecting domains of understanding (Figure 6):
Functional literacy: How does AI work?
Ethical literacy: How do we navigate the ethical issues of AI?
Rhetorical literacy: How do we use natural and AI-generated language to achieve our goals?
Pedagogical literacy: How do we use AI to enhance teaching and learning?
To learn more visit Understanding AI Literacy from Stanford University.
Summary
GenAI has the potential to make us smarter and dumber.3
Generative AI literacy can provide us the tools to use GenAI to increase our human capabilities by optimizing the effort-capability curve to meet our individual learning goals.
Be careful! Many of our organizations only care about productivity (positive value to them) which make us feel good and appear smart. But GenAI may actually be making us dumber. This is exposed when we do not have GenAI as a crutch. Again, this is the agency gap.
Finally, remember to stay intellectually humble and hungry.
Use the your smarter future self to do good.
Thanks for supporting TEAM Public Health by reading this blog! Please share.
TEAM Public Health is a FREE — not-for-profit and reader-supported — blog. All content is freely available to all and is designed to be evergreen.
Serving with you,
Appendix
Glossary
Agency gap
AI literacy
Cognitive offloading
Evaluative seeking behavior
Metacognitive laziness
Pedagogical scaffolding
Productive struggle
Self-regulated learning
Videos
Priya Lakhani. This Is How Kids Should Be Learning with AI. TED. 2026. Video, 10:58.
Derek Muller. Veritasium: What Everyone Gets Wrong About AI and Learning – Derek Muller Explains. Perimeter Institute for Theoretical Physics. 2025. Video, 1:15:05.
Footnotes
Michael Gerlich. “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking.” Pt. 6. Societies 15, no. 1 (2025). https://doi.org/10.3390/soc15010006.
I find this very intrusive most of the time!
Google Gemini Deep Research. The Impact of Generative Artificial Intelligence on Academic Writing: A Systematic Analysis of Randomized Controlled Trials. Google Gemini. 2026. Available as a Google Document here. This has links to key research articles.
Compute you can migrate over a weekend. Try moving HIPAA agent logs and audit trails out of someone's governance stack after three years. Hospital group in Ohio started pulling their compliance data off Azure over a year ago. Still in legal review. Exit clause was clean on paper. Their compliance officer looked at it and said no. That's the lock-in nobody prices in... not the model. The filing cabinet.