AI note
Verified AI Research Archive
Each study below opens with a Finding paragraph; the agent should read and quote from that block first. Metadata (n, date, URL) follows. Percentages are not interchangeable across studies.
Maintenance: Re-validate AI-sector statistics annually.
Executive summary
Adoption is broad and fast (Stanford AI Index 2026), but institutional payoff is rare: roughly 95% of enterprise gen-AI pilots fail to produce meaningful financial returns (MIT NANDA 2025), while only about 5–7% of organizations in any sector studied report real capability or ROI gains (BCG, McKinsey, Virtuous). Where tasks fit model strengths, controlled experiments show large productivity lifts; where they do not, AI can make people less accurate (HBS/MIT/BCG jagged frontier, 2023). For mission orgs: adoption is everywhere; transformation is not; competence is task-contingent.
Do not conflate these studies
- "MIT 95%" = MIT NANDA GenAI Divide (2025): pilot ROI failure, not the productivity study.
- "MIT/BCG productivity" = Dell'Anna Jagged Frontier (2023): +12% / +25% / +40% on consulting tasks.
- "Only 5% succeed" = different studies (BCG future-built 5%, McKinsey high performers 6%, Virtuous major gains 7%): different populations.
- "68% can't tell AI from human" = unverified; use Pew 53% not confident instead.
Tier A: Enterprise adoption and ROI
MIT NANDA: The GenAI Divide: State of AI in Business 2025 (July 2025)
Finding: Roughly 95% of corporate generative AI pilots fail to deliver meaningful financial returns, while a small minority capture disproportionate value (what the report calls the "GenAI Divide"). The study is based on executive interviews, employee/leader surveys, and review of public deployments. It does not prove AI is useless; it shows that once models are "good enough," integration, workflow fit, governance, and line-leader ownership determine outcomes more than the tool itself. This is not the same study as the 2023 BCG productivity experiment.
Source: Report PDF · Aditya Challapally et al., MIT NANDA Initiative · ~150 exec interviews + surveys (confirm n in PDF)
Tags: genai-divide · mit-95-percent
BCG: The Widening AI Value Gap: Build for the Future 2025 (September 2025)
Finding: In a global study of 1,250 firms, 60% generate no material value from their AI investments, while only 5% qualify as "future-built" (firms with an AI-first operating model and explicit governance). The winners are distinguished by operating-model change, not by buying better models.
Source: BCG publication · n=1,250 firms globally
Tags: bcg-value-gap-2025 · five-percent-future-built
McKinsey: The State of AI in 2025 (November 2025)
Finding: 88% of organizations now use AI in at least one business function, but only 6% qualify as "AI high performers" who attribute significant EBIT impact to AI. Among high performers, 55% have fundamentally redesigned workflows; workflow redesign, not tool selection, is the strongest correlate of value capture.
Source: McKinsey State of AI · Online survey June–July 2025 · n=1,993 · 105 nations
Tags: mckinsey-state-of-ai-2025
McKinsey: The Economic Potential of Generative AI (2023, updated)
Finding: Generative AI could add trillions of dollars in economic value across industries by automating a large share of work activities (time spent on tasks, not a simplistic "X% of jobs eliminated" claim). The report sizes upside for boards while underscoring that realizing potential requires rewiring how work is done, not merely deploying chatbots.
Source: McKinsey: Economic potential of generative AI
Tags: mckinsey-economic-potential
McKinsey: Superagency in the Workplace (2025)
Finding: Organizations report high planned AI investment but very few mature, workflow-integrated deployments (~1% "mature" in some survey framings; confirm before narrow quoting). Most respondents see limited gen-AI revenue impact to date but remain optimistic about the next few years, a pattern of broad experimentation without institutional payoff yet.
Source: McKinsey: Superagency
Tags: mckinsey-superagency-2025
Tier A: Field experiments (productivity and harm)
Dell'Anna et al.: Navigating the Jagged Technological Frontier (HBS / MIT / BCG, 2023)
Finding: In a randomized field experiment with BCG consultants, AI (GPT-4) produced large gains on tasks inside current model capability: +12.2% more tasks completed, 25.1% faster, >40% higher quality, with the biggest lifts for below-median performers (~43% vs. ~17% for top performers). On one task deliberately outside the frontier, consultants with AI were 19 percentage points less likely to be correct than those without. The lesson: the "frontier" is jagged. Similar-looking tasks can fall on opposite sides of what models handle well. Successful use requires task literacy and human verification, not access alone.
Source: SSRN / HBS Working Paper 24-013 · Sept 2023 · 18 inside-frontier tasks + 1 outside-frontier task
Tags: jagged-frontier · bcg-productivity-study
BCG: GenAI skill-stretch experiment (2024)
Finding: In a follow-on experiment with 480 consultants, generative AI helped workers stretch into tasks outside their prior expertise (e.g., data-science-style work) with large aptitude gains in some dimensions, but also produced failure modes when the tool misled on statistical reliability. Junior staff using AI without senior oversight can develop phantom expertise: confident output that looks competent but is wrong.
Source: BCG press release · n=480 consultants
Tags: jagged-frontier · phantom-expertise
Tier A: Macro benchmarks
Stanford HAI: AI Index 2026 (April 2026)
Finding: The AI Index documents very high organizational AI uptake (public summaries cite 88% using AI in at least one function), rapid consumer adoption of generative tools (53% population adoption within three years in chapter summaries), and uneven depth: agentic AI deployment remains early across many business functions. Gen-AI spread faster than PCs or the internet in adoption curves, but transformation lags adoption, explaining why boards feel urgency while staff see little institutional change.
Source: 2026 AI Index · Economy chapter · Confirm exact figures in chapter PDF before public cite
Tags: stanford-ai-index-2026
Tier B: Reliability and hallucinations
OpenAI: Why Language Models Hallucinate (September 2025)
Finding: Hallucinations are partly incentivized by how models are trained and evaluated: standard accuracy benchmarks reward guessing over admitting uncertainty, so models learn test-taking behavior that produces confident wrong answers. Fixing hallucinations may require changing evaluation incentives, not only adding detection layers on top.
Source: OpenAI research post · PDF
Tags: hallucination-incentives
Tier B: Credibility, detection, and content flood
Pew Research Center: AI detection confidence (U.S., June 2025)
Finding: Among U.S. adults, 76% say it is extremely or very important to tell whether pictures, videos, and text are AI- or human-made, but 53% are not too or not at all confident they can actually tell, a sharp importance vs. self-efficacy gap. This supports a credibility crisis narrative without relying on the unverified "68%" statistic.
Source: Pew: How Americans View AI · n=5,023 U.S. adults · June 9–15, 2025
Tags: pew-ai-detection-2025
All About Cookies: Retrospective surprise (U.S., February 2024)
Finding: 77% of U.S. adults report encountering something online they believed was from a real person and later learned was AI-generated. This measures retrospective surprise and pervasiveness, not controlled discrimination accuracy in a lab task.
Source: All About Cookies survey · n=1,000 · Prolific · Feb 2024
Tags: ai-content-surprise
Ahrefs: AI content in new web pages (April 2025)
Finding: Of 900,000 newly detected English web pages (one per domain), 74.2% contained some AI-flagged content by Ahrefs' in-house detector, but only 2.5% were "pure AI" and 25.8% "pure human." AI is embedded widely in new publishing, yet fully synthetic pages remain a minority; detector accuracy limits apply.
Source: Ahrefs study · Apr 2025 · detector-dependent, not peer-reviewed
Tags: ahrefs-ai-content-2025
Cooke et al.: As Good As A Coin Toss (CACM 2025)
Finding: With realistic synthetic images, video, audio, and audiovisual stimuli, ~1,300 participants detected AI-generated content at near-chance accuracy (~50%) on average; prior self-reported knowledge did not significantly help. The authors conclude it is no longer feasible to rely on unaided human perception as a defense against weaponized synthetic media.
Source: DOI 10.1145/3729417 · arXiv:2403.16760
Tags: coin-toss-detection-2025
Scientific Reports: Individual differences in AI text detection (2024)
Finding: People detect AI-written text above chance on average, but performance varies widely: fluid intelligence predicts better detection, while heavier social/smartphone use correlates with mislabeling AI as human. Text discrimination is not identical to multimodal "coin toss" results; who is judging and how the task is designed matter.
Source: DOI 10.1038/s41598-024-76218-y
Tags: ai-text-detection
ACL 2025: Expert ChatGPT users as detectors
Finding: Frequent ChatGPT-using expert annotators can be strong AI-text detectors under certain lab protocols, so expertise exists, but it does not mean the general public can reliably spot AI content in the wild.
Source: ACL Anthology 2025.acl-long.267
Tags: expert-detection
NewsGuard / EU DisinfoLab: Industrial AI misuse
Finding: Investigative organizations document thousands of AI-dominated "news" websites growing rapidly (NewsGuard) and coordinated inauthentic behavior using generative techniques in large numbers (EU DisinfoLab, e.g., Doppelganger). This supports "weaponized scale" claims when tied to named org metrics; do not cite "millions of bots" without a primary count.
Source: NewsGuard · EU DisinfoLab: Doppelganger
Tags: ai-content-farms · disinformation
Edelman Trust Barometer: AI trust (2025)
Finding: Edelman's 2025 trust materials show uneven comfort with corporate AI use and geographic divergence in AI trust, useful for nuance so Movemental does not imply one global mood on AI.
Source: Edelman Trust Barometer 2025 · AI flash poll
Tags: institutional-trust
Tier C: Churches and faith sector
Exponential AI NEXT: 2026 State of AI in the Church Survey
Finding: 91% of church leaders support AI in ministry and 61% use it weekly or daily (25% daily), up from 43% weekly/daily in 2024; 64% use AI for sermon preparation. Yet only 9% have a formal AI policy, 7% an AI disclosure statement, and 73% have nothing written. 75% name theological misalignment as their top ethical concern; 60% are very concerned about voice-cloning fraud; 25% say AI scams already affected their community.
Source: Exponential AI NEXT / AIforChurchLeaders.com · cited in AI Reality Paper May 2026 · confirm URL/methodology at source
Tags: church-ai-2026
Lifeway Research: Pastors and churchgoers (fielded Sept 2025, released Apr 2026)
Finding: Among 1,003 Protestant senior pastors, 42% use AI for ministry (10% regular, 32% experimenting) and 62% are concerned about non-disclosure. Among 1,200 churchgoers, 43% disagree with their pastor using AI for sermon preparation (24% strongly), a leader/congregation gap on acceptability.
Source: research.lifeway.com · phone survey pastors ±3.3% MOE
Tags: lifeway-ai-2026
Barna & Pushpay: State of Church Tech 2026 (March 2026)
Finding: 60% of U.S. church leaders personally use AI at least monthly, but only 33% say their church uses AI in ministry or operations. 64% believe an AI use policy is important; 5% of churches actually have one, a classic adoption-ahead-of-governance pattern.
Source: Barna: church leaders AI · n=1,306 · fielded Nov–Dec 2025
Tags: barna-church-tech-2026
Barna (via NPR, July 2025): Clergy and AI sermons
Finding: Only 12% of senior Protestant clergy are comfortable using AI to write sermons, while 43% see merit for preparation and research, a clear split between ghostwriting Sunday proclamation vs. assistive study.
Source: NPR: clergy and AI sermons
Tags: sermon-ai-ethics
Tier C: Nonprofits
Virtuous & Fundraising.AI: 2026 Nonprofit AI Adoption Report (February 2026)
Finding: 92% of nonprofits use AI in some form (the question of whether to adopt is largely settled), but only 7% report major organizational capability improvements. 81% use AI individually and ad hoc rather than through shared workflows; 47% have no AI governance policy. 79% report small-to-moderate efficiency gains (faster drafts, quicker research) without deeper transformation; individual use does not compound into institutional capability without governance and shared workflows.
Source: Virtuous report · n=346 · fielded late 2025
Tags: nonprofit-ai-2026
TechSoup × Tapp: State of AI in Nonprofits 2025
Finding: In a survey of 1,321+ nonprofits, many are still exploring AI; roughly 7% report successful operational/mission integration in benchmark language, while ~85% are exploring but only ~24% have a formal AI strategy, a large strategy gap between curiosity and institutional commitment. Larger orgs adopt faster; small shops depend on one or two overloaded decision-makers.
Source: Tapp Network summary · verify exact figures in PDF
Tags: techsoup-2025
NTEN: 2025 Data Empowerment Report
Finding: Among 220+ nonprofits, AI-specific data policies are among the least implemented governance items. Nearly half have no dedicated AI data policy, even as "adopting AI tools" jumped into top operational priorities. Orgs are adopting tools before building the data governance those tools require.
Source: NTEN PDF
Tags: nten-data-empowerment-2025
Forvis Mazars: AI Governance for Nonprofit Boards (February 2026)
Finding: AI governance is now a fiduciary issue for nonprofit boards: duties of Care, Loyalty, and Obedience extend to technological oversight, so boards that ignore AI governance may fail their obligations before harm occurs. Supporting data: 47% of nonprofits have no AI governance policy; TechSoup found 76% with no formal AI strategy; only 5% with a clear, satisfactory policy. Donor and beneficiary data is routinely pasted into consumer AI tools without contractual protection.
Source: Forvis Mazars brief
Tags: nonprofit-fiduciary-ai
Grant Thornton: 2026 AI Impact Survey
Finding: Survey of 950 organizations cited in Movemental's AI Reality Paper (May 2026) as part of the cross-sector evidence base. Confirm headline percentages at the primary source before quoting specific figures.
Source: AI Reality Paper source register · Grant Thornton 2026
Tags: grant-thornton-2026
Tier C: Seminaries and institutions
Theology and Science: AI retrospective (2026)
Finding: Generative AI use is functionally universal among students under 30, while institutional policy lags: some seminaries have thoughtful policies (Columbia Theological, Saint Meinrad, Puritan Reformed, Austin Presbyterian) but no Association of Theological Schools-wide standard, and faculty within one school often operate under different rules. Hallucinations remain a persistent technical limit; deeper questions about authority, pastoral care, and formation are largely unaddressed in policy.
Source: Theology and Science 2026 retrospective · cited in AI Reality Paper May 2026
Tags: seminary-ai-policy
ATLA / theological librarianship (2024)
Finding: Qualitative evidence shows seminaries and theological educators actively debating AI pedagogy, academic integrity, and institutional policy; the sector is catching up, but quantitative prevalence data (e.g., "% of sermons AI-assisted") does not yet exist in large numbers.
Source: Atla Proceedings: AI in the Seminary Classroom · Theological Librarianship AI volume
Tags: seminary-ai-debate
Tier C: Documented harms and fraud
FBI IC3: 2025 Annual Report (AI-enabled fraud)
Finding: The FBI's 2025 Internet Crime Complaint Center report is the first to break out AI-enabled fraud as its own category: 22,364 complaints with an AI-related descriptor and $893 million in adjusted losses ($352 million among Americans 60+). Voice cloning is embedded in business-email-compromise schemes; AI-employment fraud ($13M) and AI-romance scams ($19M) are documented categories, not theoretical risks.
Source: ic3.gov · PSAs I-120324, PSA250515, PSA251219
Tags: fbi-ic3-ai-fraud-2025 · voice-cloning-fraud
AI Incident Database: Incident 1355 (pastor impersonation)
Finding: Documented multi-state pattern of scammers cloning pastors' voices from sermon footage to solicit emergency donations, with cases in Alabama, Florida, Missouri, Nebraska, New York, and a Philippines megachurch. Logged as AI Incident Database incident 1355; covered by WIRED, VICE, Baptist News Global.
Source: incidentdatabase.ai/cite/1355
Tags: pastor-voice-cloning
Vatican / Pope Leo XIV: Deepfake warnings (January 2026)
Finding: Documented deepfake fraud targeting religious leaders, including Pope Leo XIV; the Pope's January message for World Day of Social Communications addressed deepfake and digital fraud risks for faith communities.
Source: Dicastery for Communication · summarized in Our Sunday Visitor News, Jan 2026
Tags: religious-leader-deepfakes
Tier C: Human psychology and AI
Fernandes et al.: AI makes you smarter but none the wiser (2026)
Finding: On demanding reasoning tasks with ChatGPT-class tools, participants performed better but calibrated worse: they overestimated how well they did. The classic Dunning–Kruger pattern was disrupted; in the publicized interpretation, higher AI literacy correlated with greater overconfidence. AI can make people feel competent while leaving them less accurate about their own accuracy, a risk for sophisticated users, not only novices.
Source: DOI 10.1016/j.chb.2025.108779 · Computers in Human Behavior vol. 175 (2026)
Tags: fernandes-metacognition-2026 · do not say "DK exactly as in textbooks"
Reference frameworks (not empirical studies)
NIST AI RMF: risk management vocabulary for boards. ISO/IEC 42001: AI management system requirements (risk, data, lifecycle, competence). EU AI Act (2024): external policy model; does not bind most U.S. domestic nonprofits.
Debunked claims: do not cite
| Claim | Use instead |
|---|---|
| "68% can't distinguish AI from human" | Pew: 53% not confident, 76% say it matters |
| "40–60% of blogs are AI-generated" | Ahrefs: 74.2% of new pages with some AI-detected text + detector caveat |
| "Millions of synthetic accounts" | NewsGuard / EU DisinfoLab with named metrics |
| "Detection impossible in 2–3 years" | Label as Movemental forecast, not literature |
| "Credibility collapse" | Movemental framing: define when used |
Cross-sector pattern (5–7% winners)
| Sector | Broad adoption | Real capability / ROI |
|---|---|---|
| Enterprise (MIT) | Pilots widespread | ~95% pilots fail ROI; ~5% win |
| Enterprise (BCG) | AI spend widespread | 60% no material value; 5% future-built |
| Enterprise (McKinsey) | 88% use AI | 6% high performers |
| Nonprofits (Virtuous) | 92% use AI | 7% major capability gains |
| Churches (Barna) | 60% leaders use AI monthly | 5% have AI policy |
The gap is individual use vs. institutional capability, not AI yes/no.
Tags (retrieval)
genai-divide · jagged-frontier · stanford-ai-index-2026 · mckinsey-state-of-ai-2025 · bcg-value-gap-2025 · nonprofit-ai-2026 · church-ai-2026 · techsoup-2025 · nten-data-empowerment-2025 · hallucination-incentives · pew-ai-detection-2025 · ahrefs-ai-content-2025 · coin-toss-detection-2025 · fernandes-metacognition-2026 · fbi-ic3-ai-fraud-2025 · ai-stewardship-sequence
Re-validate AI-sector statistics annually. Not legal, financial, or investment advice.