This week: AI is widely used, and its mistakes still carry real costs.
SIGNAL 25 · AI / TRUST · HIT
Saturday result: HIT under the published three-check rubric v2.1. Assignment, continuation-shaped output, and fallout are supported by reviewed sources; contrary findings are retained. The evidence review was completed before the scheduled 3:30 PM EST public readout.
Signal: 25 · ID: GRV-2026-W46 · Receiver: AI / Trust · Source publication window: November 7–13, 2026 (Eastern dates) · Review: November 14, 2026. Some studies' underlying fieldwork predates their publication.
01 / WHAT THE WORLD SHOWED
AI is increasingly positioned as a source of fast, accurate business insight and resolved customer-service needs. Yet a GWI survey released on November 10 found frequent inaccurate results and self-reported operational costs among heavy business AI users. A separate November 9 consumer-service survey showed growing trust when resolution is guaranteed. A November 10 experimental preprint found that user reliance could remain high even when AI advice was consistently wrong, unless participants received accuracy feedback.
Scope: These sources cover distinct samples and tasks. They support the three published checks without demonstrating that an AI tool caused all reported consequences, and do not establish any claim about the structural detection itself.
PRESENT. NiCE promotes end-to-end service resolution using AI agents and human handoffs. Consumers were explicitly asked about trusting AI when resolution is guaranteed. This is a promise and conditional preference, not proof the resolution occurs.
PRESENT ✓ · PROMISE / ASSIGNMENTCONTINUATION DOCUMENTED. GWI reports 88% daily AI use, 80% frequent exposure to misleading findings, and only 36% verifying AI insights most of the time. This is recurrent use accompanied by further checking and correction, not demonstrated durable arrival.
PRESENT ✓ · CONTINUATION DOCUMENTEDPRESENT IN SELF-REPORT. In GWI's same respondent group, 82% reported at least one negative consequence of AI inaccuracies, including rework (38%), delays (36%), and reported financial losses (27%). These are self-reports, not independently audited harms.
PRESENT ✓ · CONSEQUENCES SELF-REPORTED02 / RESULT · RUBRIC v2.1
HIT: Assignment, continuation-shaped output, and observable fallout are all supported in the public materials released during the selected publication window. The business outcome evidence comes from a single cross-sectional self-report survey and should not be interpreted as longitudinal proof of AI causation.
W1: Positive improvements in consumer AI trust, and successful calibration when accuracy feedback is provided, are material contrary data; neither by itself documents a stable, durable, completion-side outcome across the AI/Trust receiver. W2: not triggered; a promised endpoint is explicit. W4: not triggered; reported downstream costs are observable. W3: nonrepresentative sampling and release-date versus fieldwork distinctions are limiting, but do not erase the GWI within-sample outcome record.
03 / SOURCE WINDOW & PUBLICATION
Source-window qualification: November 7–13 is the ordinary seven-day review interval reconstructed for this report, not claimed as a separately published Signal 25 source-window commitment. Source release dates and earlier underlying fieldwork are labeled distinctly.
Public status: Evidence review prepared November 14 before the public upload; the scheduled public readout remains 3:30 PM EST. This file is not a production-deployment receipt. After public posting, attach the actual GitHub/Netlify publication timestamp rather than backdating.
04 / DATED PUBLIC SOURCE RECORD
Six identifiable source publications (released November 9–12) are listed below. Two are primary research-company releases, one is an industry-community report, one is a preprint, one is a magazine account of earlier research, and one validates a survey instrument. They are not six independent replications.
- 2026-11-09 · NiCE / Metrigy — Q3 2026 AI Consumer Experience Index · ASSIGNMENT; CONTRARY
Among 1,000 U.S. adults, 28.8% would choose an AI agent if issue resolution were guaranteed, up from 23% in Q1; accuracy is the leading choice factor (54.7%). AI Trust Index rose to 106.9 relative to Q1 baseline 100. Survey pertains to Q3, released Nov 9.
Limit: Vendor-commissioned study; conditional intentions are not actual guaranteed resolution.
- 2026-11-09 · Thinkers360 — 2026 AI Trust Index · OUTPUT; CONTRARY
More than 80% of surveyed AI users/providers report at least some concern about AI trust, versus 87% in 2025; 71% believe benefits outweigh risks. Survey conducted September–October 2026 and released Nov 9.
Limit: Community/industry sample with limited public denominator and sampling details; concern is not directly measured harm.
- 2026-11-10 · GWI — Artificial Confidence: The AI Trust Gap · OUTPUT; FALLOUT
Of 1,883 senior marketing, insights and agency professionals across U.K., U.S., Singapore: 88% used AI daily, 80% frequently met misleading outputs, 58% reported an important decision based on an AI insight later shown wrong, and 82% reported an adverse business consequence; 38% rework, 36% delayed decisions, 27% financial loss.
Limit: Research fieldwork in July 2026, released Nov 10; self-reported and nonrepresentative of all businesses; no independent causal audit.
- 2026-11-10 · Jones et al. — Do People Trust AI Assistants Even When They Are Consistently Wrong? (preprint) · OUTPUT; CONTRARY
Four controlled experiments: with accuracy feedback, participants selected reliable advice more than unreliable (67% versus 21%). Without feedback, 48% were persuaded by consistently wrong AI advice.
Limit: Preprint, not yet peer-reviewed; laboratory persuasion decisions, not real-world downstream business damage.
- 2026-11-10 · Nature — The public trusts scientists with AI · CONTRARY / CONTEXT
The article reports comparatively high public confidence in scientists using AI and contextualizes trust as role-dependent.
Limit: Reporting on a Reuters Institute study published Oct 29; not a fresh Nov 10 measurement.
- 2026-11-12 · Yankouskaya et al. — Perceived Trustworthiness of LLMs scale · METHODOLOGY / CONTEXT
Psychometric study introduced an eight-item perceived-trust measure, tested among 752 U.K. LLM users; published November 12.
Limit: Instrument validation is not evidence of system-level fallout or successful completion.
05 / CONTRARY EVIDENCE · NEGATIVE CONTROLS
Measured improvement: NiCE/Metrigy reports AI trust 106.9 on an index standardized to 100 in Q1 2026; the share preferring an AI agent with guaranteed resolution rose from 23% to 28.8%. These are genuine positive findings and may signal performance gains in some service settings.
Trust can be calibrated: Jones et al. found external correctness feedback led users to favor helpful AI more than unhelpful AI (67% versus 21%) in controlled trials. This counters any suggestion that users must remain stuck with unreliable AI.
Benefit and role differentiation: Thinkers360 reports 71% say AI benefits outweigh risks and that overall concern improved from 87% (2025) to over 80% in 2026. Nature's November 10 article notes comparatively high public confidence in scientific use of AI, based on a study issued October 29.
Not proven: None of these cross-sectional public sources measure structural Completion, establish that additional steps alone cause trust failure, or causally connect NiCE's promises to GWI's reported harms. These limitations remain part of the record.
06 / SOURCE REVIEW & PRESERVATION
Preparation: GPT-6-assisted review of named public web sources and Gravity's preserved v2.1 protocol, November 14, 2026. Source selection was performed after the original publication dates; historical data collection dates are distinguished from release dates. Independent second-model adjudication: NOT PERFORMED. Exact historical source snapshots: NOT RECONSTRUCTED. This is a prepared public-edition record; actual public deployment is confirmed separately.
SOURCE-CHECK VERIFICATION: The cited public releases were cross-checked against their published findings. All three checks support the HIT classification under rubric v2.1. Inspect verification addendum ↗. This checks the reported source evidence; it does not independently audit or establish causation.
Record preservation: Signal 24 remains OBSERVING / W3 with its source receipt and local P≈0.75 note untouched. Earlier Signals 01–24 are not rescored. The scan tracks predicted fallout, not proof of Completion → Sequence.