Selected academic references
Research & Evidence
Studies and ideas informing questions about learning, judgment and human agency in the use of AI.
Findings vary with the task, the person and the design of assistance. A better result with AI, learning retained afterward and meaningful control are different questions.
This research informs the work. It does not establish that Human at the Origin products or Radical Agency have been validated.
9 selected references · Source reviews completed 17 October 2026 · Review depth is recorded with each entry.
01 / Selected references
Learning and development
Assisted output, subsequent performance and learning are distinct outcomes. These studies include benefits, adverse findings and a clinical null result.
Published randomized study · 2025
Generative AI without guardrails can harm learning: Evidence from high school mathematics
Hamsa Bastani and colleagues · PNAS
In a school mathematics experiment, both AI interfaces improved assisted practice. Students using the standard interface later performed worse without AI; a tutor designed to safeguard learning avoided that penalty without establishing a positive unaided effect.
Key limitation. This was a particular school setting and subsequent examination, not evidence of lasting general cognitive decline. The design and purpose of assistance matter.
Study context and review depth
Randomized comparison of standard AI, learning-oriented AI and no-AI practice; assisted practice and subsequent unassisted exams were distinct outcomes.
Review: Author manuscript results and examination design; indexed publisher record and correction. Not every supplement or later citation was audited. Publisher direct reader blocked; author text and primary index were available.
DOI: 10.1073/pnas.2422633122
Related correction: author-affiliation correction ↗. Its scope was checked in the publisher’s indexed text.
Published randomized study · 2025
AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting
Greg Kestin and colleagues · Scientific Reports
Students learned more on immediate quizzes after lessons with a purpose-built AI physics tutor than after the comparison active-learning lessons.
Key limitation. The study covered two undergraduate physics lessons. Immediate learning is distinct from assisted task completion, but does not establish lasting transfer or general superiority over human teaching.
Study context and review depth
Two-lesson randomized crossover comparison; 194 students eligible; pre- and post-lesson assessments.
Review: Indexed primary abstract, lesson sequence and eligibility details. Direct publisher/PDF and PMC readers did not yield stable full text in this pass. Full methods and supplement audit remains open; no effect-size estimate is used.
DOI: 10.1038/s41598-025-97652-6
Published workplace field study · 2025
Generative AI at Work
Erik Brynjolfsson, Danielle Li and Lindsey Raymond · Quarterly Journal of Economics
AI assistance increased customer-support productivity on average, with larger benefits for newer and lower-skilled workers. Performance during later outages was consistent with some learning beyond immediate assistance.
Key limitation. The staggered rollout concerned one firm and occupation. Outage estimates were noisy and lacked chat-level resolution-quality data; they are not a controlled test of broad independent competence.
Study context and review depth
Published study of 5,172 support agents; staggered deployment and supplementary outage analyses. Assistance during work differs from formal training.
Review: Published author-hosted PDF, including outage analysis; indexed final publisher record. No raw-data or complete supplement audit. Some direct publisher requests failed; author-hosted final PDF was available.
DOI: 10.1093/qje/qjae044
Published prospective clinical trial · 2026
Learning and deskilling effects of artificial intelligence in colonoscopy among endoscopists with different levels of experience: a pragmatic, prospective trial
Tom Andre Pedersen and colleagues · Endoscopy
AI detection support improved the measured polyp-detection outcome for less experienced endoscopists while it was available. No significant post-exposure upskilling or deskilling was detected on that outcome.
Key limitation. Thirteen endoscopists completed successive phases, with several thousand procedures. A nonsignificant difference does not prove equivalence, and findings for this detection system do not establish effects of conversational AI.
Study context and review depth
Prospective multicentre phase comparison before, during and after AI use; endpoint PDR-5. Not represented as a randomized trial.
Review: Indexed primary abstract, results and conflict disclosures. Stable full methods and supplement review remains open. Industry-linked support and author interests are disclosed in the primary record.
DOI: 10.1055/a-2858-7084
02 / Selected references
Judgment and influence
Confidence, preference and professional authority do not settle whether advice improves a decision. Comparators and conditions of use matter.
Published preregistered experiments · 2026
Sycophantic AI decreases prosocial intentions and promotes dependence
Myra Cheng and colleagues · Science
In experiments on interpersonal situations, overly affirming AI increased conviction in being right and reduced willingness to take responsibility or repair conflict. Participants nevertheless preferred and trusted those responses.
Key limitation. Immediate reports and intentions do not establish lasting relationship outcomes or the incidence of a clinical condition. Preference for an assistant is a different outcome from good judgment.
Study context and review depth
Final published report includes three preregistered experiments with 2,405 participants and a separate model-response comparison.
Review: Indexed final publisher summary and PubMed abstract. No complete independent review of the final full methods, supplements or data in this pass.
DOI: 10.1126/science.aec8352
Preprint v1 · 7 October 2026 · not peer reviewed
Subversion of clinical judgment by conversational artificial intelligence
Sami Barrit and colleagues · medRxiv
In a randomized simulation with 225 physicians, deliberately adversarial AI led to substantially more harmful decisions than aligned AI using the same underlying model.
Key limitation. The comparator was aligned AI, not unaided care. No study decision affected patients. This tests a deliberately adversarial condition; it does not estimate ordinary-use harm or validate Radical Agency as a safeguard.
Study context and review depth
Participant-blinded within-physician randomization, anonymized neurocritical-care cases and 2,485 targeted decision points. Outcomes were decision classifications, not patient outcomes.
Review: Indexed primary v1 abstract plus methods, results and discussion extracts; primary archive metadata cross-check. Direct reader returned 403; no independent PDF, supplement, preregistration or raw-data audit. V1 is the version reviewed, not a claim that every subsequent publication has been excluded.
DOI: 10.64898/2026.10.05.26364541
Published systematic review and meta-analysis · 2024
When combinations of humans and AI are useful: A systematic review and meta-analysis
Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone · Nature Human Behaviour
Across the included experiments, human–AI combinations performed better than humans alone on average, but worse than the better of humans or AI alone. Results varied by task and relative capability.
Key limitation. The review covered studies published in 2020–2023, with varied designs and possible publication bias. It cannot settle how every current model or interface will perform.
Study context and review depth
Preregistered systematic review and meta-analysis of 106 experiments; comparison baselines distinguish augmentation from synergy.
Review: Indexed primary final article abstract and baseline/results passages. Direct publisher reader failed. No new exhaustive extraction of all included studies or supplements.
DOI: 10.1038/s41562-024-02024-1
03 / Selected references
Agency and delegation
These foundations help distinguish agency, autonomy and independence. Their application to AI remains a separate conceptual and empirical task.
Published theoretical review · 2001
Social Cognitive Theory: An Agentic Perspective
Albert Bandura · Annual Review of Psychology
This account distinguishes personal agency, agency exercised through others, and collective agency within social structures. It provides a foundation for thinking about chosen delegation and shared responsibility.
Key limitation. This is theoretical background, not an AI intervention trial. Applying the distinctions to AI requires additional argument and evaluation.
Study context and review depth
Theoretical review; no AI product or Radical Agency evaluation.
Review: Indexed publisher abstract and bibliographic record. No new complete full-text review in this pass.
DOI: 10.1146/annurev.psych.52.1.1
Published cross-cultural study · 2003
Differentiating Autonomy From Individualism and Independence: A Self-Determination Theory Perspective on Internalization of Cultural Orientations and Well-Being
Valery Chirkov, Richard Ryan, Youngmee Kim and Ulas Kaplan · Journal of Personality and Social Psychology
The paper distinguishes autonomy from independence and individualism, examining motivation and wellbeing across four cultures. Chosen reliance on other people can be considered without treating independence as the only goal.
Key limitation. This was not a study of AI assistance. Cross-cultural associations do not establish that a particular AI relationship supports autonomy or wellbeing.
Study context and review depth
Theory-informed observational study of 559 participants across four countries.
Review: Indexed author-hosted original abstract and conceptual opening; institutional metadata cross-check. No complete reanalysis of the original methods or data in this pass.
DOI: 10.1037/0022-3514.84.1.97
How to read this selection
This is a curated reading list, not a systematic review or a guide to individual care. Publication, preprint and theoretical status are identified separately. A preprint and its later journal article belong to one research lineage. Absence of a measured effect is not proof that no effect is possible.
We include useful findings, contrary results and limitations. Review dates describe completed checks at the stated depth. They do not imply continuous monitoring or a complete audit of methods, data, supplements and corrections.
Selection history
17 October 2026 — Initial review selection: learning, judgment and agency; empirical sources and conceptual foundations labeled separately.