Reading room.

Choose a reading connected to a question from the discussion, or use the seven patterns below to examine an example from your work.

Readings by question

Generative AI without guardrails can harm learning

Bastani and colleagues · PNAS · 2025 · High-school mathematics experiment

Students improved during AI-supported practice, but those using the less constrained tutor later performed worse on their own. Teacher-designed hints largely mitigated that harm. Read it for the difference between assisted performance and what a learner retains.

Discuss claim 1: evidence of learning · Claim 3: developmental effort · Study and limits

The Triage Trap

Eli Talbert · War on the Rocks · 2026 · Military decision-making essay

Talbert asks what happens when AI filters the options before a commander sees them. His argument gives us a way to examine which choices reach a school leader’s desk. That application to education is our analogy.

Discuss claim 2: judgment before approval · Context and limits

Own the Outer Loop

Addy Osmani · 2026 · Software-engineering commentary

Osmani describes the responsibility of the person who sets an agent’s constraints, checks its work, and approves what ships. Consider what someone in your institution would need to explain when an agent acts on their behalf.

Discuss claim 2: approval and responsibility · Source context

The Irreducible Officer

Jack Shaw · 2026 · Essay on professional education

The essay asks how educators can see who owns the reasoning when AI helps produce the work. Section II describes invisible delegation; sections VI and VIII propose ways to preserve developmental effort and examine what a learner can defend.

Discuss claim 4: the meaning of better · How we adapted the essay

Examining and Addressing Barriers to Diversity in LLM-Generated Ideas

Deng, Brucks, and Toubia · 2026 · Research preprint

The authors investigate why independent AI outputs can resemble one another and test prompts that increase their diversity. Their results give us both a reason to question shared defaults and ways to broaden the ideas AI supplies.

Discuss claim 5: shared frames · Findings and limits

From Trust to Appropriate Reliance

Raees and Papangelis · 2026 · Review of 22 studies

Feeling confident in AI, following its advice, and following advice that deserves acceptance are different things. The review examines how researchers measure those differences. Use it when the conversation turns to what would justify accepting a recommendation.

Return to claim 2 · Review and limits

Navigating the Jagged Technological Frontier

Dell’Acqua and colleagues · 2023 · Field experiment with consultants using GPT-4

AI helped on some tasks and reduced correctness on another. The task-specific results make a useful companion to the reliance review: success with one kind of work gives us a reason to test the next kind.

Return to claim 2 · Experiment and limits

Seven patterns to look for

A finished answer can hide who chose the problem or the standard for success. Jack Shaw describes these seven patterns in The Irreducible Officer. Use them to ask which choices a person examined and which they accepted from AI.

The frame arrives before the question

The first plausible account of the problem becomes the boundary of the work. Later improvements stay inside it. The frame may be reasonable and still leave out what matters.

Frame capture. “Improve attendance” becomes “send stronger reminders,” before anyone asks why students are absent.

The language does the convincing

Balanced prose, confident structure, and well-placed caveats can make a recommendation sound considered before anyone has explained the decision.

Fluency substitution. A recommendation lists every risk but never says which one should govern the choice.

The synthesis comes too soon

Connections arrive before the reader has worked through the disagreements. A coherent summary can conceal tensions worth preserving.

Premature synthesis. “Stakeholders favor the proposal” compresses incompatible reasons and a consequential minority objection.

Success in one task becomes permission for another

People may trust a recommendation because the agent performed an earlier task well. Rejecting sound advice simply because it came from AI can fail too.

Uncalibrated reliance. An agent accurately summarizes records, so its recommended intervention gets the same confidence.

An ordinary request hands over the criteria

A request for help can leave the agent to choose criteria the person never examined.

Invisible delegation. “Make this more strategic” lets the system decide what strategic means. “Choose the best option” lets it supply the definition of best.

Different answers share the same assumptions

A group can receive varied wording and still work within a narrow set of frames. Similar tools and defaults may amplify that risk.

Institutional monoculture. Every proposal assumes efficiency is the objective; none asks whether access, fairness, or learning should take priority.

The recommendation loses its author

“The system recommended it” can make a human choice sound inevitable or objective. An approval may record the decision without showing who examined the reasoning.

Responsibility laundering. A plan is accepted because it passed through several reviewers, but no one can explain the decisive assumption.

Practice and tools

Explore the essay companion, try a spoken argument, or consider whose perspective is missing.