Accuracy & misinformation
Fabricated citations and synthetic media can mislead. Verify claims and provenance before relying on an output.
Champions and challenges
Risks, evidence, and misconceptions
Evaluate a specific use, its evidence, and who bears its costs. Choosing to limit or decline AI use belongs in this conversation. Explore Kenji’s education questions →
Privacy & surveillance
Consider what data enters a system, who can access it, and whether people can meaningfully consent.
Bias & unequal impacts
Test who benefits, who is excluded, and how people can challenge an automated decision.
Learning, work & dependency
Ask whether a tool develops skills or substitutes for practice, and how it changes working conditions.
Energy, water & infrastructure
Examine site-specific impacts, utility demands, and community participation in data-center decisions.
Copyright & security
Review permissions and ownership. Untrusted instructions can exploit tool-enabled systems; limit access and check actions.
Six risks and misconceptions
Expand a topic for evidence, limits, and a question to take into a decision. The misconception statements are examples of claims to examine, not survey findings about what everyone believes.
Unreliable outputs
Misconception to examine: A fluent answer with citations must be correct.
What the source says: NIST describes confabulation: confidently presented false content. A reference in an answer still needs verification. NIST GenAI Profile, 2024 · sections 2.2 and 3 ↗
Limits and uncertainty: Error rates depend on the task, model, and evaluation conditions. A good demonstration cannot establish reliability in every setting.
A practical check: Open the cited source and check that it supports the exact claim. Recalculate important numbers. Identify who reviews errors before anyone acts.
Privacy and data use
Misconception to examine: A training opt-out settles every privacy concern.
What the source says: NIST identifies leakage, unauthorized disclosure, and re-identification as privacy risks. Training is only one part of data handling. NIST GenAI Profile, 2024 · section 2.4 ↗
Limits and uncertainty: Retention, sharing, access, and deletion need separate checks for the specific service and account.
A practical check: Read the relevant settings and terms. Ask where prompts, uploads, logs, and outputs go. Use public or fictional examples when possible.
Bias and representation
Misconception to examine: Using software removes human bias.
What the source says: NIST identifies historical bias and performance differences across groups or languages. Aggregate scores can conceal those differences. NIST GenAI Profile, 2024 · sections 2.6 and 3 ↗
Limits and uncertainty: Results may change with population, context, and the fairness measure used.
A practical check: Look for evaluation with affected people. Ask which groups were tested, what failed, and how someone can challenge a harmful result.
Labor and working conditions
Misconception to examine: Exposure to AI means those jobs will disappear.
What the source says: The ILO’s 2025 study estimates which occupational tasks could be affected. It expects job transformation to be more likely than full replacement because many tasks still require human input. ILO Working Paper 140, May 2025 · research estimate ↗
Limits and uncertainty: Exposure estimates are not counts of actual layoffs. Adoption, job design, and policy influence outcomes; displacement and changes in work quality remain concerns.
A practical check: Ask which tasks change, who receives the time savings, and whether staff have a voice, training, and workable expectations.
Environmental impacts
Misconception to examine: One per-prompt figure tells us AI’s environmental cost.
What the source says: The IEA’s 2025 analysis models electricity demand from data centers and shows that local impacts can differ from global averages. NIST also identifies energy and water used in training and operation as environmental concerns. IEA Energy and AI, April 2025 · estimates and scenarios ↗
NIST GenAI Profile, 2024 · section 2.5
Limits and uncertainty: A prompt estimate depends on workload and system boundaries. Electricity projections depend on adoption and efficiency. They do not establish every site’s water use, emissions, or community effects.
A practical check: Ask what was measured, where, and when. Compare a specific proposal’s power, water, noise, and land impacts with alternatives. Follow DMV infrastructure records →
Copyright and authorship
Misconception to examine: Writing a prompt automatically gives me copyright in its output.
What the source says: The U.S. Copyright Office’s January 2025 report says protection depends on sufficient human expressive authorship. AI assistance does not automatically disqualify human-created work; merely providing prompts is insufficient. U.S. Copyright Office, January 2025 · U.S. scope ↗
Limits and uncertainty: Copyrightability, permission to use material, attribution, and academic integrity are different questions. Application depends on the work and jurisdiction; the report does not decide your particular case.
A practical check: Record your own contribution and disclose assistance where required. Check source permissions and licenses separately from a tool’s account terms.
Sources checked October 11, 2026. Publication dates identify the evidence used; these are dated explanations rather than a live research feed. NIST offers voluntary guidance, the ILO and IEA provide analysis, and the Copyright Office describes its U.S. interpretation. The practical checks are this hub’s editorial synthesis.
Connect these topics to Kenji’s education concerns → · Use the evaluation guide →
Starting points: NIST risk framework · AI Now’s critical infrastructure toolkit · EFF’s digital-rights analysis. These are distinct institutional perspectives, not a single consensus.
Ask who benefits, who bears the risks, whether the evidence supports the claim, and what happens when a system fails. A decision to limit AI use can be as informed as a decision to adopt it.