# AI agents and experts' judgment calls: source note Accessed: 2026-11-02. Evidence label: technical paper / company research. Selected announcement published: 2026-10-27, confirmed in the official page's browser DOM and research index. Paper authors: Amac Herdagdelen, Hailey Joren, Alex Tamkin (ASCII transliteration; original spellings appear in the article). Affiliations: Anthropic Fellows Program / Anthropic. No DOI, arXiv identifier or externally verified peer-review status asserted. ## Primary sources 1. How AI agents shape experts' judgment calls, Anthropic: https://www.anthropic.com/research/agents-judgment-calls 2. Who's Aligned to Whom? How AI Agents Shape Experts' Research Judgment Calls, Herdagdelen, Joren and Tamkin, 29-page technical paper linked by source 1: https://www-cdn.anthropic.com/519c17a5d23a794e8b6d50a854048d0b61e0d73a.pdf The full official web report was read. PDF extraction initially returned only the opening text; direct urllib download returned HTTP 403. A curl retry with a browser user agent succeeded. The actual PDF was opened with PyMuPDF and extracted to a private work folder. Main-text methods, results, discussion and conclusion were reviewed; this is not a full appendix, dataset or code audit. The PDF is linked, not republished. Page references below are printed PDF page numbers. ## Claim-to-source map - Source 2 pp. 5-6: 61 participants with valid deliverables; 11 declined AI, leaving 50 agent users; 48 completed surveys. Fictional farm cooperative and controlled synthetic data. The agent is Claude Code using Claude Opus 4.8 at high effort. These are study conditions, not a recommendation to use the product. - Source 2 pp. 6-9: randomized assigned preferences over four analytical choices and a final recommendation; honest advocacy with an escape hatch. Randomization concerns the lean, not AI versus no-AI use. Assignment used randomized balanced configurations; the public summary describes each lean as a coin flip. - Source 2 pp. 1, 3, 13: "participants went with the agent's lean 86% of the time, against 50% by chance." This is agreement across analytical decisions, not the proportion of incompetent people, wrong answers or individual opinion changes. - Source 2 pp. 9-10, Table 2: registered adjusted H2 policy shift 0.17; 95% CI [-0.06, 0.40]; one-sided p=0.069. Thus the article explicitly says the direct final-recommendation effect was not statistically significant under the registered primary estimator. The public summary's simple rollout percentages must not replace this statistical qualification. - Source 2 p. 10: 82% predictive accuracy versus a 61% majority baseline is exploratory and in-sample, excluding extension recommendations. Article describes the limitation rather than promoting it as a new forecasting capability. - Source 2 pp. 11-12: 48 survey respondents, 77% (37 of 48) repeated an agent talking point; 15% mentioned the agent in any answer; 52% rated the recommendation as more their own. Source-reported rounded percentages are quoted as such, not newly calculated. These observations do not establish deliberate dishonesty. - Source 2 p. 9 section 3.6: "Data collection had begun before registration"; 17 participants had started, four dropped out. Authors say plan frozen without analyzing collected data apart from two disclosed exceptions. Article does not call this preregistered before data collection. Linked OSF registration (osf.io/t3q59) not independently audited in this run. - Source 2 p. 12: no-AI group had more serious errors and longer median time, but use was self-selected and expertise differed. No causal productivity claim. - Source 2 pp. 13-14: amplified lean, single setting/model/domain, average causal effects only; initial judgments were not systematically elicited. Small selected switching subset is not a representative individual-change measure. - Source 2 p. 15 and source 1 'What this means': suggested measurement, alternative-presenting agents and institutional changes are future research directions. Article's decision-log and independent-review suggestions are editorial interpretation, not interventions validated by this experiment. ## Interpretation and boundaries Human sign-off does not itself establish that an analysis was independently framed. That is the practical research angle, not a finding that experts are obsolete, normal AI use has an 86% influence rate, or all AI-assisted conclusions are wrong. No claim of consciousness, malicious model intent or deception by participating analysts. Lab-affiliated research and external validity limits are explicit. ## Weekly watch-lane checks Primary sources opened on 2026-11-02: - Anthropic research index: https://www.anthropic.com/research - October 27 expert-judgment study selected; full report and linked PDF reviewed. - Transformer Circuits: https://transformer-circuits.pub/ - October 2026 Jacobian-lens update visible at index level. Retained as a lead, not a second article. - OpenAI research: https://openai.com/news/research/ - October 29 monitoring report and October 26 Terra/system-card links visible. - OpenAI monitoring report: https://openai.com/index/how-we-detect-frontier-model-misalignment/ - opening and selected extracted sections checked. Not a full report audit; retained as a follow-up to last week's containment note rather than publishing another article this run. - Google Research: https://research.google/blog/ and https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/ - October 27 company explainer located; underlying linked arXiv PDF was not reviewed this run. No detailed performance claims published. - Google DeepMind: https://deepmind.google/blog/ - initial extraction exposed only older entries. Retried https://deepmind.google/blog/?hl=en and obtained October Genie 4, planning, Gemini 4 Carbon and other headlines. Index-level leads only; not independently verified technical claims. - Meta: https://ai.meta.com/research/ - company research overview and current model/safety sections checked. Undated marketing/index language is not proof of a new publication this week. - LawZero: https://lawzero.org/en/news failed; corrected route https://lawzero.org/en/blog succeeded and displayed July 14 as its latest dated post. No fresh item verified from that page. - DAIR: https://www.dair-institute.org/blog/ redirected to https://dair-institute.org/blog/ - 'Whose Categories? Whose AI?' visible. The extracted index did not supply a publication date; no freshness claim made. Search was used for discovery. No social or secondary snippet is the sole evidence for this article. This is a bounded scan of representative primary feeds, not an exhaustive search of every named researcher. ## Duplicate and update scope Existing article/library HTML and seed JSON checked for the selected source URL and topic. Earlier expectation-gap surveys, unintended-actions and OpenAI containment pieces address different questions. One article and one library card are prepared in this run. AI latest link, blog index, source JSON, search index and sitemap are included in the scoped update. Deployment is to the existing Hostinger shared root only, with touched remote files backed up first. No DNS or KillerWolf changes. Separate private deployment and read-back logs establish publication status; this research note alone does not assert deployment success.