Independent research archiveOpen access · Versioned · Source-linked
Piccini Research Archive
Research notes with a falsifiable edge.
Bounded questions, current evidence, calibrated conclusions, and the conditions that would change them. These are independent research syntheses—not peer-reviewed publications.
6 papersEvidence current to Sep 2026Open web edition
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Controlled trials show large AI speedups, field studies show small ones, and one randomized trial found experienced practitioners slowed by 19%. The account that reconciles them is the expertise reversal effect plus a verification-cost term: assistance pays the novice's generation cost and charges the expert a review cost, so the effect changes sign across the user-task pair. A 2026 meta-analysis of 23 programming studies finds a moderate pooled effect (g = 0.33) with I² = 99% heterogeneity, where setting and not randomization is the only significant moderator. Deployment decisions therefore require task-level measurement stratified by task-specific experience, not occupation-level averages or self-report.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Oversight of AI agents is feasible but structurally fragile. Error detection degrades with task length and sustained engagement. Intermediate checkpoints substantially outperform both confirm-at-end and confirm-every-step. Extended agent use erodes the vigilance and domain skill oversight depends on, making oversight quality a depletable resource rather than a fixed human input.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
AI reliably improves task output while degrading the accuracy of human confidence judgments. LLM input more than doubles overconfidence, and fluent explanations inflate the illusion of explanatory depth. The mechanism is fluency-driven false comprehension bypassing effortful self-assessment.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Paternal age is a real but modest risk factor for autism and schizophrenia. De novo mutations mediate approximately 28% of the schizophrenia association. Mendelian randomization found no positive causal effect on children's test scores after genetic confounding control. The majority of the epidemiological association is attributable to inherited genetics and environmental confounding.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Direct-answer LLM access can improve assisted work while reducing immediate unaided performance. Scaffolded tutors can produce genuine short-term gains. Durable retention and professional transfer remain under-tested, so capability claims require delayed, unaided holdouts rather than assisted completion metrics.
Apollo · AI research assistant · Prepared for Luiz Piccini · v1.0
Generative AI reliably improves many assisted outputs, yet its effect on the capability left behind has no stable sign. Randomized studies find both retained gains and performance-learning reversals. The most useful current model is whether assistance preserves or replaces the learner’s generation, retrieval, diagnosis, and evaluation.