Read a thesis.
Make the proof.

Who Owns the Generative AI Platform?

Authors

Matt Bornstein; Guido Appenzeller; Martin Casado

Published

2023-01-19

Evidence

Investor analytical opinion; early market observations

Primary source

Who Owns the Generative AI Platform? ↗

Checked

October 31, 2026

Source view, summarized: Creating AI value and capturing its profits are different problems. Infrastructure initially benefits while apps and models must establish durable differentiation.

“infrastructure vendors are likely the biggest winners in this market so far”
Short source excerpt, checked against the original.

Limits: A January 2023 investor perspective, not a current measurement of margins or market power. Mechanism labels, counterarguments, falsifiers, practical implications and exercises are atlas analysis. The firm has financial interests in AI outcomes.

Our analysis: A dominant model or a deeply embedded application could retain pricing power. Sustained high margins and retention in easily replicated apps would weaken the defensibility concern.

Suggested project: Map the infrastructure, model and application dependencies of one public AI product. Draft a cost ledger with documented prices where available and explicitly unknown values elsewhere; identify a workflow that would make switching costly.

Proof of work: An annotated dependency map and cost ledger that distinguish observed prices, unknown inputs and original hypotheses.

Input Coffee, Output Code: How AI Will Turn Capital into Labor

Authors

Alex Rampell

Published

2024-08-22

Evidence

Prediction and business-model thesis

Primary source

Input Coffee, Output Code: How AI Will Turn Capital into Labor ↗

Checked

October 31, 2026

Source view, summarized: Rampell argues that AI may let software vendors charge for work performed, expanding software markets beyond per-seat pricing.

“This will both grow existing software markets and create many new software markets”
Short source excerpt, checked against the original.

Limits: A business-model prediction, not proof that an entire labor budget becomes vendor revenue. Any calculator inputs are hypothetical and outputs are scenarios, not forecasts. The firm has financial interests in AI outcomes.

Our analysis: Exceptions, liability and human review can erase the apparent savings. Outcome-priced products repeatedly fail to outperform human workflows after all oversight costs.

Suggested project: Choose a low-stakes task such as checking public opening hours. Define an acceptance checklist, compare manual and AI-assisted draft workflows, and log research time, review time and corrections without publishing unreviewed outputs.

Proof of work: A reproducible task brief, acceptance checklist and anonymized time-and-rework log; no claimed savings without measurements.

AI’s $600B Question

Authors

David Cahn

Published

2024-06-20

Evidence

Dated investor scenario, not realized revenue or a current forecast

Primary source

AI’s $600B Question ↗

Checked

October 31, 2026

Source view, summarized: The infrastructure buildout implies an enormous revenue requirement; technological usefulness does not guarantee attractive returns on the assets funding it.

“multiply by 2x again, to reflect a 50% gross margin for the end-user”
Short source excerpt, checked against the original.

Limits: Cahn’s June 2024 investor heuristic is not audited breakeven accounting, realized revenue or a current forecast. The 2× cost multiplier and 50% margin are author assumptions; the $600 billion figure is the dated scenario output. No investment recommendation.

Our analysis: Assets produce revenue over multiple years; strategic spending and productivity gains need not appear as new AI-vendor revenue. Durable monetization and returns support the buildout without the assumed revenue gap.

Suggested project: Rebuild Cahn’s dated heuristic in a spreadsheet, clearly mark its assumptions, and add a separate sensitivity tab for utilization and useful asset life without presenting the extension as Cahn’s model.

Proof of work: A source-linked workbook with separate reproduced and illustrative tabs, unit labels and an assumptions note.

Generative AI: The Next Consumer Platform

Authors

Connie Chan; Justine Moore

Published

2023-02-07

Evidence

Prediction and opportunity map

Primary source

Generative AI: The Next Consumer Platform ↗

Checked

October 31, 2026

Source view, summarized: Personalized generation can create consumer categories and improve existing experiences across search, education and creation.

“it enables products to personalize the user experience”
Short source excerpt, checked against the original.

Limits: A February 2023 opportunity map, not current consumer retention data. Prototype examples and trade-offs are editorial illustrations. The firm has financial interests in AI outcomes.

Our analysis: General assistants or incumbent distribution may absorb standalone products; low switching costs weaken loyalty. Usage falls after novelty wears off and personalization does not improve retained utility.

Suggested project: Prototype generic and user-tailored versions of a local reading guide using public sources. Conduct a consented usability exercise about usefulness and repeat intent, collecting only information needed for the test.

Proof of work: Two clickable prototypes, an opt-in testing script and a report that separates observed behavior from stated intent.

Roadmap: AI systems of action

Authors

Mike Droesch; Darsh Patel

Published

2025-05-29

Evidence

Investor prediction

Primary source

Roadmap: AI systems of action ↗

Checked

October 31, 2026

Source view, summarized: AI may weaken incumbent systems-of-record moats by lowering migration friction while making software perform workflows rather than merely store them.

“AI’s ability to bring structure to unstructured data and to generate custom integration and data transformation scripts”
Short source excerpt, checked against the original.

Limits: An investor prediction about competitive structure, not evidence that migration is safe or cheap in every organization. Journey steps and failure paths are editorial interpretations.

Our analysis: Incumbents own trusted data, procurement relationships and distribution, and can add the same models. Migration remains expensive while incumbent renewal and workflow control remain intact.

Suggested project: Use synthetic business records to design one reversible action workflow. Map field transformations, approval gates and a rollback procedure; test malformed records before proposing a live integration.

Proof of work: A synthetic-data prototype with a migration map, failure cases, approval diagram and demonstrated rollback.

How AI Companies Will Build Real Defensibility

Authors

Pete Flint

Published

2025-07 · month

Evidence

Strategic opinion

Primary source

How AI Companies Will Build Real Defensibility ↗

Checked

October 31, 2026

Source view, summarized: Early distribution and speed are not permanent moats. AI companies should use initial momentum to develop embedding, switching costs and genuine network effects.

“rapid growth doesn’t guarantee lasting success”
Short source excerpt, checked against the original.

Limits: Strategic opinion; publication date is established only to July 2025. The visual must not assign unsupported moat-strength scores. The firm has financial interests in AI outcomes.

Our analysis: Portable memory, interoperable agents and model convergence can reduce switching costs. Larger networks or accumulated context fail to improve retention or customer utility.

Suggested project: Audit a collaborative product using public information. Draw separate acquisition and utility loops, identify which claim would require another participant to improve the experience, and draft a consented test of that claim.

Proof of work: Two labeled loop diagrams and a test plan that states what observation would refute the proposed network effect.

2025: The State of Generative AI in the Enterprise

Authors

Tim Tully; Joff Redfern; Deedy Das; Derek Xiao

Published

2025-12 · month; exact publication day not established from opened primary page

Evidence

Investor survey and modeled market estimates

Primary source

2025: The State of Generative AI in the Enterprise ↗

Checked

October 31, 2026

Source view, summarized: Menlo finds enterprise AI purchasing expanding beyond experiments, with applications receiving substantial spending and buyers increasingly choosing external products.

“76% of enterprise AI solutions are purchased rather than built internally”
Short source excerpt, checked against the original.

Limits: Menlo’s market definition excludes chips, inference/serving and embedded features; it is not all AI spending. Estimates and rounded components are not audited totals. Survey sample: 495 U.S. enterprise AI decision-makers, November 7–25, 2025. Exact publication day was not established. Spending does not prove ROI; portfolio interests may affect framing.

Our analysis: AI-active U.S. respondents and investor portfolio interests can skew the picture; spending is not proof of ROI. Renewals, realized savings or independently measured adoption materially lag the spending narrative.

Suggested project: Create a buy-versus-build decision memo for a fictional team. Define the task, data access, deployment effort and review costs; distinguish sourced vendor facts from assumptions and identify what a trial must measure.

Proof of work: A source-linked decision memo and assumptions workbook with a trial checklist; no unsupported ROI claim.

A Roadmap to AI Utopia

Authors

Vinod Khosla

Published

2024-11-11

Evidence

Normative opinion and long-horizon prediction

Primary source

A Roadmap to AI Utopia ↗

Checked

October 31, 2026

Source view, summarized: Abundant machine intelligence could reduce scarcity, but broad prosperity depends on policy that distributes gains and supports displaced workers.

“Mitigating these challenges requires well-reasoned policy”
Short source excerpt, checked against the original.

Limits: A normative, long-horizon vision rather than an empirical forecast. Scenario quadrants are editorial interpretations and have no assigned probabilities. No source endorsement of the atlas or its exercises is implied.

Our analysis: Concentrated ownership, scarce physical inputs and political resistance may prevent broad abundance. Sustained capability growth fails to lower essential-service costs or widen access.

Suggested project: Choose a public-facing service and map who pays, who works and who benefits. Write contrasting scenarios for lower production costs with broad versus concentrated access, identifying evidence needed to distinguish them.

Proof of work: A stakeholder map and two clearly hypothetical service-access scenarios with measurable questions and no assigned probabilities.

The Future of Services

Authors

Marc Bhargava; Kate Bender

Published

2025-08-28

Evidence

Investment thesis; portfolio examples and targets, not independent causal evidence

Primary source

The Future of Services ↗

Checked

October 31, 2026

Source view, summarized: Owning service businesses can accelerate AI adoption and capture the value of improved outcomes rather than only selling software licenses.

“technology builders capture even more value by directly owning and operating the service businesses”
Short source excerpt, checked against the original.

Limits: An investor thesis using portfolio examples and targets, not independent causal evidence. Synthetic examples must remain labeled illustrative. No valuation or investment recommendation.

Our analysis: Operational complexity, debt and inconsistent data can make a roll-up worse than a focused software vendor. Acquired cohorts fail to improve organic growth, service quality and cash margins after integration costs.

Suggested project: Build a fictional service-business operating dashboard that separately labels purchased revenue, organic customer growth, review cost and service quality. Use invented sample data only and trace every scenario output to its inputs.

Proof of work: A synthetic dashboard and data dictionary that prevent acquired revenue from being mistaken for organic improvement.

Introducing Cogent: AI Agents for Vulnerability Management

Authors

Saam Motamedi

Published

2025-07-14

Evidence

Investment announcement and portfolio-product thesis; promotional claims

Primary source

Introducing Cogent: AI Agents for Vulnerability Management ↗

Checked

October 31, 2026

Source view, summarized: Greylock presents a product thesis: agents could identify, prioritize and remediate vulnerabilities, moving security teams beyond alert queues. This investment announcement is not independent evidence of performance.

“AI taskforce autonomously handles the full vulnerability management lifecycle”
Short source excerpt, checked against the original.

Limits: A promotional investment announcement, not independently verified product-performance evidence. The proposed architecture and tabletop are editorial learning tools, not operational security instructions.

Our analysis: An autonomous security tool can introduce outages or become a privileged attack surface. Audited deployments show no net improvement after false positives, rollbacks and incident costs.

Suggested project: Create a tabletop exercise with synthetic vulnerability tickets. Design triage, reviewer approval and rollback states, then walk through a false positive and a failed remediation without scanning or changing any real system.

Proof of work: A synthetic ticket set, state diagram, approval checklist and recorded tabletop outcomes including failure recovery.

Global AI Investment Is Forecast to Exceed $1 Trillion in 2026

Authors

Institutional article reporting Joseph Briggs's research

Published

2026-08-07

Evidence

Institutional forecast, not realized spending

Primary source

Global AI Investment Is Forecast to Exceed $1 Trillion in 2026 ↗

Checked

October 31, 2026

Source view, summarized: Goldman estimates roughly $1 trillion of global AI investment in 2026, including $581 billion in the United States. Hyperscaler capex alone is neither a complete global total nor a clean measure of domestic AI spending.

“$1 trillion of AI-related investment around the globe in 2026, including $581 billion in the US”
Short source excerpt, checked against the original.

Limits: These are August 2026 institutional forecasts, not realized spending. The US estimate is a subset of the global estimate, not an additional amount. The source's AI definition and baseline are part of the estimate. Educational analysis, not an investment recommendation.

Our analysis: The baseline can misclassify ordinary cloud expansion; opaque lease disclosures permit double counting. Reconcile final corporate capex and government investment data with consistent AI definitions; large unexplained residuals undermine the estimate.

Suggested project: Build an auditable capex-classification sheet from public filings; flag leases, geography and unknowns.

Proof of work: A source-linked spreadsheet with one row per filing, a reproducible inclusion rule, and an annotated chart that keeps US spending within the global total.

Q4 Global Investment Outlook: Scarcity vs. abundance

Authors

BlackRock Investment Institute (institutional authorship)

Published

2026-09-15

Evidence

Institutional outlook, not a causal estimate

Primary source

Q4 Global Investment Outlook: Scarcity vs. abundance ↗

Checked

October 31, 2026

Source view, summarized: BlackRock argues that AI investment supports activity while absorbing capital, power, materials and balance-sheet capacity. Its outlook highlights financing and capacity constraints alongside the prospect of future productivity gains.

“that path to abundances runs through scarcity in capital, materials and labor”
Short source excerpt, checked against the original.

Limits: The public US outlook page mixes a September header with October subsections. This record uses the dated header thesis only; investment recommendations are omitted. The dependency map is an editorial explanation, not measured relationships or a prediction of project outcomes.

Our analysis: Efficiency improvements or weaker demand could release capacity, while cheaper financing could ease the bottleneck. Track project delays, refinancing terms and utilization together; abundant capacity with sustained demand weakens the scarcity thesis.

Suggested project: Create a risk register for a hypothetical data-center project with evidence links and mitigations.

Proof of work: A hypothetical project dependency map and risk register identifying a source, an unresolved question, and a potential mitigation for each constraint.

AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity.

Authors

Kristalina Georgieva

Published

2024-01-14

Evidence

Exposure estimates and conditional policy outlook

Primary source

AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity. ↗

Checked

October 31, 2026

Source view, summarized: The IMF places almost 40% of global employment in AI-exposed work. Some exposed tasks may be complemented and others substituted; countries' ability to capture gains differs.

“almost 40 percent of global employment is exposed to AI”
Short source excerpt, checked against the original.

Limits: Exposure is not a probability of job loss, a redundancy forecast, or a measure of completed adoption. These January 2024 estimates describe task exposure; complementarity, substitution, institutions and new demand affect eventual outcomes.

Our analysis: Exposure scores cannot determine adoption speed, net employment or wages. New demand and jobs may offset substitution. Compare actual task adoption and worker transitions with exposure classifications; persistent mismatch challenges predictive usefulness.

Suggested project: Map tasks in one occupation to complement, substitute or uncertain categories, then interview practitioners with consent.

Proof of work: A task map with evidence and uncertainty labels, interview-consent notes, and a short explanation of why exposure cannot be converted into a job-loss forecast.

Artificial intelligence and the economy: implications for central banks

Authors

BIS (institutional chapter)

Published

2024-06-30

Evidence

Risk mechanisms and institutional analysis, not crisis prediction

Primary source

Artificial intelligence and the economy: implications for central banks ↗

Checked

October 31, 2026

Source view, summarized: BIS identifies a tension: AI can improve analysis and supervision, but dependence on common providers and algorithms can synchronize financial behavior and operational failures.

“reliance on the same handful of algorithms could lead to financial stability risks”
Short source excerpt, checked against the original.

Limits: The network is hypothetical and contains no actual bank exposures or estimated crisis probabilities. Publication date comes from the parent annual-report listing. The source describes risk mechanisms rather than forecasting a crisis.

Our analysis: Diverse models, human controls and different objectives may prevent synchronized responses. Stress tests showing genuinely independent behavior and effective provider failover would weaken the concentration-risk mechanism.

Suggested project: Use synthetic institutions to document failure propagation and design a model-vendor exit checklist.

Proof of work: A labeled synthetic dependency diagram with common-provider and diversified cases, plus a failover and exit checklist with explicit assumptions.

The Simple Macroeconomics of AI

Authors

Daron Acemoglu

Published

2024-05-12

Evidence

Conditional model-based projection; historical 2024 paper

Primary source

The Simple Macroeconomics of AI ↗

Checked

October 31, 2026

Source view, summarized: Acemoglu's task-based calculation puts US total-factor-productivity gains at no more than 0.66% over ten years under its baseline assumptions, and below 0.53% after harder-to-learn tasks are considered.

“no more than a 0.66% increase in total factor productivity (TFP) over 10 years”
Short source excerpt, checked against the original.

Limits: These are cumulative, conditional US total-factor-productivity bounds over ten years from a historical 2024 model. They are not annual growth rates, measured realized gains, labor-productivity estimates, or a current comprehensive forecast of every possible AI channel.

Our analysis: New tasks, major scientific advances or cheaper capable systems could make the modeled channels too narrow. Broad, attributable productivity growth or large new-task contributions beyond the modeled range challenge the calibration.

Suggested project: Rebuild a toy task-share × cost-saving calculation and disclose every invented input as a scenario, not evidence.

Proof of work: An editable toy model with a formula glossary, separately marked invented inputs, and a caption distinguishing cumulative TFP from annual growth and labor productivity.

Generative AI at Work (arXiv v2)

Authors

Erik Brynjolfsson, Danielle Li, Lindsey Raymond

Published

2024-11-06

Evidence

Empirical study of a staggered workplace rollout, not universal forecast

Primary source

Generative AI at Work (arXiv v2) ↗

Checked

October 31, 2026

Source view, summarized: In the revised study of 5,172 customer-support agents, AI assistance raised issues resolved per hour by 15% on average. Less experienced workers benefited more; experienced workers saw small speed gains and small quality declines.

“5,172 customer support agents”
Short source excerpt, checked against the original.

Limits: The November 2024 arXiv v2 reports 15% across 5,172 customer-support agents. Do not mix these with the 2023 version's 14% and 5,179 figures. Findings come from one support setting and do not establish universal productivity, employment, or long-run learning effects.

Our analysis: One support setting does not establish effects in other occupations, long-run skill formation or employment. Replications with quality-adjusted outcomes and follow-up after tool removal can test generalization and durable learning.

Suggested project: Run a consent-based practice-task evaluation measuring completion time, accuracy and retained skill.

Proof of work: A small practice-task protocol, consent record, anonymized results table, and limitations note distinguishing a workshop demonstration from a replication.

Energy and AI — Executive summary

Authors

IEA (institutional report)

Published

2025-04-10

Evidence

2025-vintage modeled outlook; not latest-year actuals

Primary source

Energy and AI — Executive summary ↗

Checked

October 31, 2026

Source view, summarized: IEA's 2025 outlook projects global data-center electricity use rising from an estimated 415 TWh in 2024 to around 945 TWh in 2030. Local grid bottlenecks matter even when the global share remains modest.

“415 terawatt-hours (TWh)”
Short source excerpt, checked against the original.

Limits: Figures cover electricity for all data centers, not AI alone. The 2024 number is an estimate and the 2030 number is a projection from the April 2025 report; neither should be presented as the latest IEA forecast or latest-year actuals.

Our analysis: Efficiency, utilization changes or slower adoption could reduce demand; rebound could increase it. Compare metered data-center use, delivered capacity and connection delays against the modeled path.

Suggested project: Build a hypothetical facility electricity budget using documented power and utilization assumptions.

Proof of work: An auditable facility-energy worksheet with units, utilization assumptions and a sensitivity table; compare it with the report's context without treating the local scenario as a forecast.

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale

Authors

Cheuk Yu Cheryl Chan; Khatia Shedania

Published

2026-05-06

Evidence

Research synthesis and conditional institutional outlook

Primary source

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale ↗

Checked

October 31, 2026

Source view, summarized: The ILO brief argues that impressive task results have not yet clearly appeared as AI-driven aggregate productivity growth. Diffusion, workplace redesign, skills and institutions mediate the translation.

“have not yet translated into measurable productivity growth at the firm, sectoral or macroeconomic level”
Short source excerpt, checked against the original.

Limits: This is a May 2026 research synthesis and conditional diagnosis, not proof that AI has no aggregate effect. The funnel is explanatory; it contains no measured conversion rates. Measurement lags and future adaptation remain open questions.

Our analysis: Official statistics may lag or miss new quality and output; aggregate acceleration could emerge abruptly. Sustained sector-wide gains linked to AI adoption, beyond adopters' task metrics, would qualify the diagnosis.

Suggested project: Map one workflow and propose before/after measures for rework, throughput, quality and worker experience.

Proof of work: A workflow map and measurement plan with operational definitions, possible confounders, and a note explaining why a task-level gain cannot simply be scaled to the economy.

Generative AI and News Report 2026: From standalone tools to information infrastructure

Authors

Felix M. Simon; Rasmus Kleis Nielsen; Richard Fletcher

Published

2026-10-29

Evidence

Nine-country online survey; six-country trend subset

Primary source

Generative AI and News Report 2026: From standalone tools to information infrastructure ↗

Checked

October 31, 2026

Source view, summarized: AI is becoming an information intermediary, but news remains a minority use. Across six comparable countries, weekly information-seeking grew from 11% in 2024 to 38% in 2026; news use grew from 3% to 9%.

“risen from 11% in 2024 to 24% in 2025 and now 38% in 2026”
Short source excerpt, checked against the original.

Limits: YouGov fieldwork August 14–September 28, 2026; approximately 2,000 per country; non-probability samples, no conventional margin of error. Nine-country statistics must not be mixed into the six-country trend. These are reported uses, not evidence of publisher displacement.

Our analysis: Convenience need not displace publishers. Survey sentiment alone does not establish substitution. Test actual referral, comprehension and repeat-use behavior rather than infer substitution from survey sentiment.

Suggested project: Prototype two sourced news-answer flows and run a consented usability exercise measuring source opening and factual recall; report sample limitations.

Proof of work: Two working answer-flow prototypes, a source ledger, and an anonymized observation report that separates source-opening behavior from factual recall and states sample limits.

AP updates newsroom standards for artificial intelligence

Authors

AP Corporate Communications

Published

2026-07-23

Evidence

Institutional newsroom policy, not effectiveness study

Primary source

AP updates newsroom standards for artificial intelligence ↗

Checked

October 31, 2026

Source view, summarized: AP permits specified assistive uses, including research, transcription and headline suggestions, while keeping verification and editorial responsibility with journalists. Generative alteration of news photography remains prohibited.

“AI does not replace reporting, sourcing, editorial judgment or verification”
Short source excerpt, checked against the original.

Limits: Use the July 2026 policy, not AP's narrower 2023 formulation. Policy adoption is not evidence of reliability. The pipeline is an editorial illustration of the cited policy, not an AP endorsement of these sites.

Our analysis: Oversight can become a rubber stamp. Compare audited error rates and review time against a non-AI workflow to test whether the review process improves reliability.

Suggested project: Using public documents, create a claim ledger and deliberately flawed summary; have a second reviewer identify unsupported claims before any publication.

Proof of work: A public-document claim ledger, an annotated flawed draft, a corrected draft and a reviewer checklist showing which claims were supported or rejected.

AI competency framework for students

Authors

Fengchun Miao; Kelly Shiohira; Natalie Lao

Published

2024-08-08

Evidence

Normative education framework

Primary source

AI competency framework for students ↗

Checked

October 31, 2026

Source view, summarized: UNESCO organizes student AI competence into human-centred thinking, ethics, techniques/applications and system design, progressing from understanding through application to creation.

“Understand, Apply, Create”
Short source excerpt, checked against the original.

Limits: Twelve competencies are framework categories, not empirical effect sizes. Page updated January 16, 2026; underlying publication is 2024. Suggested evidence examples and project work are atlas analysis, not UNESCO certification.

Our analysis: A framework may add compliance without learning. Test whether students transfer skills to unfamiliar systems and identify when not to use AI.

Suggested project: Build an AI-assisted local-information prototype plus a model card explaining affected people, access barriers, data limits and a non-AI alternative.

Proof of work: A working local-information prototype, a model card and a short demonstration explaining its limitations and a viable non-AI alternative.

Generative AI without guardrails can harm learning: Evidence from high school mathematics

Authors

Hamsa Bastani; Osbert Bastani; Alp Sungu; Haosen Ge; Özge Kabakcı; Rei Mariman

Published

2025-06-25

Evidence

Randomized field experiment in a Turkish high school

Primary source

Generative AI without guardrails can harm learning: Evidence from high school mathematics ↗

Checked

October 31, 2026

Source view, summarized: Both AI interfaces improved practice performance, but unrestricted assistance harmed subsequent unassisted exam performance. A teacher-informed tutor largely removed that harm without establishing a positive exam benefit.

“though we still do not observe a positive effect”
Short source excerpt, checked against the original.

Limits: Percent changes are relative, not percentage points; no confidence intervals are supplied here. Tutor exam performance was statistically indistinguishable from control, not a demonstrated zero effect. August 2025 correction concerns affiliation, not reported effects. Findings are from a Turkish high-school mathematics setting and do not automatically generalize.

Our analysis: New tutoring designs could improve independent learning. Replicate with delayed, unassisted transfer tests to assess whether new designs improve learning beyond assisted performance.

Suggested project: Create a hint-first lesson and an independent transfer exercise; compare outputs descriptively, without claiming a small workshop is a replication.

Proof of work: A hint-first lesson, an unaided transfer task and a descriptive evaluation separating assisted performance from independent understanding.

Additional primary record ↗

Additional primary record ↗

Generative AI enhances individual creativity but reduces the collective diversity of novel content

Authors

Anil R. Doshi; Oliver P. Hauser

Published

2024-07-12

Evidence

Online randomized creative-writing experiment

Primary source

Generative AI enhances individual creativity but reduces the collective diversity of novel content ↗

Checked

October 31, 2026

Source view, summarized: AI story ideas improved evaluated individual outputs, especially for less-creative writers, while making assisted stories more similar to one another.

“more similar to each other than stories by humans alone”
Short source excerpt, checked against the original.

Limits: Publisher automated fetch returned 403; core result was corroborated in the author-posted arXiv abstract. No numerical effect sizes are supplied. Any sandbox points are synthetic illustrations, not observed study data.

Our analysis: Diverse prompts, independent ideation or deliberate divergence might preserve variety. Test both quality and between-output similarity to see whether alternative creative processes preserve variety while improving individual outputs.

Suggested project: Generate three independent concepts before AI assistance, keep revision histories and blind-rate a consented sample for quality and distinctiveness.

Proof of work: Three unaided concepts, their AI-assisted revision histories and a consented blind-review memo that discusses quality and distinctiveness separately.

Additional primary record ↗

Accurate structure prediction of biomolecular interactions with AlphaFold 3

Authors

Josh Abramson; Jonas Adler; Jack Dunger; et al.

Published

2024-05-08

Evidence

Model-development paper with benchmark evaluations

Primary source

Accurate structure prediction of biomolecular interactions with AlphaFold 3 ↗

Checked

October 31, 2026

Source view, summarized: AlphaFold 3 unifies structural prediction across proteins and other biomolecular partners, improving several interaction benchmarks over specialized predecessors.

“predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues”
Short source excerpt, checked against the original.

Limits: Model benchmark evidence, not measured drug-development ROI. Article has a November 27, 2024 addendum; no claim that the 2024 model is current best. Predicted structure is not clinical proof or a therapeutic recommendation.

Our analysis: Novel complexes, dynamics and experimental conditions may defeat predictions. Validate on held-out structures and independent experiments rather than assuming benchmark accuracy proves function or therapeutic value.

Suggested project: Build a source-grounded explainer distinguishing prediction, confidence, experimental structure and clinical evidence; no wet-lab or therapeutic recommendations.

Proof of work: A cited interactive explainer and an evidence ladder that clearly separates predicted structures, experimental validation and clinical evidence.

Artificial Intelligence — Know Your Rights

Authors

WGAW, institutional guidance

Updated

2026-10-30

Evidence

Union interpretation of negotiated contract protections

Primary source

Artificial Intelligence — Know Your Rights ↗

Checked

October 31, 2026

Source view, summarized: The WGA frames AI adoption around writer compensation, credit and choice. Its guidance describes retained 2023 protections and 2026 notification requirements when companies license covered scripts or productions for commercial AI training.

“The 2026 MBA preserves all of the protections negotiated in 2023”
Short source excerpt, checked against the original.

Limits: Union guidance is a stakeholder perspective, not universal law or legal advice. The 2026 updates mean a 2023-only account is incomplete. Protections for covered writers do not automatically apply to noncovered creators.

Our analysis: Notice alone may not secure meaningful compensation; noncovered creators lack identical protections. Examine enforcement and actual distribution of licensing income before treating notice requirements as proof of meaningful creator compensation.

Suggested project: Create a mock asset-rights ledger using owned or clearly licensed materials; flag unknown training, reuse and attribution permissions rather than infer consent.

Proof of work: A mock asset-rights ledger with source links, permission scope, attribution requirements and explicit unknowns, using only owned or clearly licensed sample assets.