Evidence-backed Analysis

-Research suggests retrieval-augmented generation can substantially reduce hallucination relative to a model answering from memory alone, but it does not eliminate hallucination.

-In high-stakes domains, purpose-built RAG tools can still produce unsupported answers a substantial share of the time, with one legal study reporting roughly 17 to 33 percent across products.

-On tasks inside AI's reliable competence, novices and below-average workers often show the largest measured performance gains.

-The novice-favoring compression pattern has appeared across several independent field settings, not just one experiment.

-On greenfield, self-contained coding tasks, AI assistance can produce large measured speedups.

-Less experienced developers tend to show larger measured gains from AI assistance than more experienced ones.

-Research suggests two firms adopting similar information technology can get very different productivity returns, and the gap is associated more with management quality than with the technology itself.

-Structured management practices have been associated with a substantial share of productivity variation across plants, on the order of 20 percent in US Census data and about a quarter of TFP gaps in cross-country survey work, comparable to the share attributed to R&D or IT.

-The famous parity between interviews and cognitive tests rested partly on a statistical artifact that was revised downward in 2022.

-Adding structure raises an interview's predictive validity compared with an unstructured conversation.

-Pay transparency is not one policy but at least three with different and sometimes opposite effects: within-firm horizontal disclosure, cross-firm salary posting, and gender-gap reporting.

-Within-firm horizontal transparency, where coworkers can see each other's pay, is associated with lower average wages of roughly 2 percent.

-The employment effect of a higher minimum wage is conditional, not universal: its sign depends on local market structure and bite rather than a single law.

-The size and direction of employment effects scale with the bite, the floor relative to the local median wage.

-Research suggests banning non-competes for lower-wage hourly workers is associated with higher wages.

-Higher non-compete enforceability is associated with lower wages and a weaker return to tenure on average, consistent with reduced worker bargaining power.

-Occupations that require state-specific licensing exams show substantially lower interstate migration, around a third lower in one study, while occupations using national exams show no such gap.

-The labor-market mobility cost of licensing appears to come largely from license non-portability rather than from licensing itself.

-AI tutoring systems improve learning compared with business-as-usual instruction when they provide structured, adaptive practice in well-bounded domains.

-The key boundary between helpful AI tutoring and harmful AI assistance is whether the tool preserves productive student effort while providing timely support.

-When sports-event accounting includes full Scope 3 activity, the largest climate gains often come from transportation and event-system design outside the facility envelope rather than from building systems alone.

-On-site facility improvements remain meaningful because they reduce the operator's direct and purchased-energy footprint, lower utility and water impacts, improve waste outcomes, and create auditable annual performance data.

-Whether homeownership is financially advantaged over renting in the US is a conditional question.

-The user cost of housing framework is the accepted analytical frame in tenure-choice economics.

-The Tax Cuts and Jobs Act of 2017 reduced the number of itemizing tax filers from approximately 46 million (TY2017) to approximately 16 million (TY2018), a drop of roughly 30 million filers driven by the near-doubling of the standard deduction and the $10,000 SALT cap.

-The share of US tax filers claiming the mortgage interest deduction fell from approximately 21% in TY2017 to approximately 8% in TY2018, a 13-percentage-point drop in a single year driven by the TCJA standard deduction change, not by any change in mortgage holding or mortgage rates.

-Systematic phonics instruction in kindergarten and first grade has been associated with small-to-moderate gains in decoding and word reading compared with less-systematic or non-phonics instruction.

-Phonemic awareness has been consistently associated with later word reading, and short explicit phonemic awareness instruction linked to letters has been associated with gains in decoding for beginning readers.

-Oral reading fluency measures have been associated with moderate to strong prediction of early grade reading comprehension, with reported correlations often in the 0.

-Multi-stage or gated screening models, in which children flagged by an initial brief screener are reassessed through dynamic assessment or short-term progress monitoring, have been associated with lower false positive rates than single-gate screening.

-Academic gains from preschool programs have frequently faded by first or third grade, even when other effects persist.

-US kindergarten classrooms have become substantially more academic and teacher-directed since the late 1990s, with less time for play, art, and child-initiated activity.

-Juvenile rats deprived of play with peers show altered prefrontal cortex development and impaired adult social behavior, providing the strongest causal evidence for play's developmental functions in any species.

-Preschoolers' exploratory play is sensitive to evidence quality and can support causal inference, with children targeting ambiguous or confounded contexts for fuller exploration.

-Regular preschool exposure to curriculum-based educational programming such as Sesame Street has been associated with small-to-moderate gains in vocabulary, letter recognition, and other school-readiness skills.

-Children under roughly two and a half years generally learn less vocabulary and less transferable content from pre-recorded video than from equivalent live interaction.

-Educational apps for young children vary widely in quality and most top-downloaded apps score low on learning-science criteria.

-Manipulative design features are common in popular children's apps and may shift time away from learning content.

-High-quality center-based preschool programs have been associated with short-term gains on language and early literacy measures in the range of 0.

-Early Head Start has been associated with small positive effects on toddler language and cognitive outcomes at age 3 (around d = 0.

-High-dosage tutoring programs with small groups, multiple weekly sessions, and in-school delivery have been associated with moderate average effect sizes on reading outcomes in randomized trials.

-Phonics-based structured literacy interventions have been associated with statistically meaningful gains in reading and spelling for students with dyslexia or severe reading difficulties.

-Some studies observe workers approaching benefit eligibility thresholds constraining their hours or declining overtime to keep earnings below cliff cutoffs.

-When a worker simultaneously risks losing two or more benefits (e.

-During the 2023-2024 Medicaid unwinding, approximately 69% of the 25+ million disenrollments were procedural rather than eligibility-based, suggesting that the majority of coverage losses were driven by administrative failure to complete renewal processes rather than genuine changes in eligibility.

-Experimental evidence from LA County found that one-third of SNAP applications were denied due to missed interviews, five times the number denied for ineligibility.

-Research consistently finds that income thresholds in benefit programs have larger effects on the extensive margin (whether to work) than the intensive margin (how much to work).

-Observable income bunching at benefit thresholds is limited and concentrated among specific populations.

-Effective marginal tax rates for low- and moderate-income workers range from near-zero to over 80%, with some household configurations exceeding 100%.

-When multiple benefit programs phase out across the same income range, the combined effective marginal tax rate exceeds what any individual program designer intended.

-Research consistently finds that the extensive margin of labor supply (whether to work at all) is more responsive to tax-benefit policy than the intensive margin (how many hours to work).

-Optimization frictions, including fixed work schedules, adjustment costs, and imperfect information, help explain why micro-econometric studies find small labor supply elasticities while macro-level evidence suggests larger responses.

-Industry and analyst research consistently shows that inefficient manual processes are not marginal.

-Empirical case studies show dramatic performance differences between manual and automated execution in structured, repetitive tasks.

-Industry surveys consistently show high AI adoption rates across organizations, but deep workflow integration remains uncommon.

-A Federal Reserve working paper tracking U.

-Academic benchmarks and case studies show substantial variance in extraction accuracy depending on the approach and document complexity.

-Multiple case studies from named organizations report order-of-magnitude reductions in document processing time after implementing AI-driven extraction.

-Multiple large-scale studies converge on a consistent finding: email occupies a disproportionate share of work time.

-Research documents that email has evolved beyond communication into an informal task management platform.

-Multiple studies converge on a consistent finding: workers spend a substantial portion of their day locating information rather than using it.

-The economic costs of data fragmentation are documented at multiple scales.