Metatron Intelligence
Today: the five AI headlines of the day → and the AI Wiki
What a domain is How to read this page

One of twenty-four connected subject areas

Domains are how claims relate to one another inside the Atlas. Desks are how readers find them. A claim can sit in more than one domain without being counted twice.

Coverage here is not importance

A domain with more claims has been worked on more, which is a fact about this publication rather than a fact about the world.

Everything returns to the record

Each entry links to its canonical claim record, with its source, its non-claim and its update condition intact.

New to this publication?

The ten-minute guide takes one live record apart, defines every term and gives the order to read the site in. Start here →

Horizontal · D11

AI Models, Agents & Human–AI Interaction

367 connected claims · 336 source-domain links · Priority P0

Search and filter this domain →
  1. MWITA-CR-2026-0008Metatron sollte Health-Aussagen als versionierte, juristiktions- und populationsgebundene Claims führen und AI auf Suche, Extraktion und Widerspruchserkennung begrenzen, während medizinische Schlussfolgerung und Freigabe unter fachlicher menschlicher Autorität bleiben.Evidence B · P0
  2. MWITA-CR-2026-0017In METRs randomisierter Studie benötigten 16 erfahrene Open-Source-Entwickler mit frühen 2025er AI-Tools für 246 reale Aufgaben 19 % mehr Zeit, obwohl sie eine Beschleunigung erwarteten.Evidence A · P0
  3. MWITA-CR-2026-00182025 nutzten 20,0 % der EU-Unternehmen irgendeine AI-Technologie, aber 55,0 % der großen Unternehmen und 18,9 % der KMU; diese Statistik misst weder Agentenadoption noch Produktivitätswirkung.Evidence A · P1
  4. MWITA-CR-2026-0026Lokale Sprache ist Infrastruktur: SEA-LION, BHASHINI, Masakhane und Latam-GPT adressieren Daten-, Benchmark-, Code-Switching-, Rechte- und Repräsentationslücken, die eine nachträgliche Übersetzung nicht zuverlässig löst.Evidence B · P0
  5. MWITA-KI-2026-001South Korea's Framework Act on AI, Act No. 20676, was promulgated on 21 January 2025 and entered into force on 22 January 2026.Evidence A · P0
  6. MWITA-KI-2026-003Article 24 of Presidential Decree No. 36053 applies the safety threshold only when the system cumulatively meets at least 10^26 training FLOP, state-of-the-art capability, and broad or significant risk to life, safety or fundamental rights.Evidence A · P0
  7. MWITA-KI-2026-021Government of India releases reported IndiaAI compute capacity exceeding 45,000 GPUs in 2026.Evidence B · P1
  8. MWITA-KI-2026-022BHASHINI's official repositories describe ULCA data and model infrastructure covering machine translation, automatic speech recognition, text-to-speech and OCR.Evidence B · P1
  9. MWITA-AP-2026-004BHASHINI's NMT evaluation workshop documented that Indian-language evaluation needs language-specific treatment, including multiword-expression problems in Bengali and explicit attention to data quality, bias and hallucination.Evidence B · P1
  10. MWITA-AP-2026-006India's FREE-AI committee recommends financial-sector AI governance that combines board policy and lifecycle controls with indigenous multilingual models, including narrow language-task-domain combinations for local financial use cases.Evidence A · P1
  11. MWITA-AP-2026-017Thailand's public research agency NECTEC describes Pathumma 1.0 as an open-source Thai-focused family spanning text, vision and audio, while characterizing future agentic functionality as a plan rather than a released outcome.Evidence B · P1
  12. MWITA-AP-2026-022Singapore's IMDA reported MERaLiON support for Malay, Tamil, Thai, Bahasa Indonesia, Vietnamese, English, Mandarin and Singlish with code-switching, while its first release had more than 90,000 downloads; these are program and distribution signals, not independent quality proof.Evidence B · P1
  13. MWITA-AF-2026-006Kenya's AI Strategy 2025–2030 is a national roadmap organized around AI digital infrastructure, data and AI governance, and AI research, innovation and commercialization.Evidence A · P0
  14. MWITA-AF-2026-008IrokoBench's peer-reviewed evaluation of 17 human-translated low-resource African languages found the best tested open model, Gemma 2 27B, reached 63% of the best tested proprietary model GPT-4o across its benchmark, while translate-to-English improved some larger English-centric models.Evidence A · P0
  15. MWITA-AF-2026-009AfriHuBERT reports continued self-supervised pretraining on more than 10,000 hours spanning 1,226 African languages or dialects plus four widely used non-African languages, with gains on FLEURS language identification and speech recognition relative to mHuBERT-147.Evidence A · P1
  16. MWITA-AF-2026-011AfroXLMR-Social reports task-specific F1 improvements of 1–30% after domain-adaptive pretraining on AfriSocial social-media text across 19 languages, and identifies religious-corpus domain bias in prior language models.Evidence A · P1
  17. MWITA-AF-2026-014SAHPRA published version 1 of MD08-2025/2026 on 19 August 2025, setting regulatory requirements for AI/ML-enabled medical devices in South Africa.Evidence A · P0
  18. MWITA-AF-2026-021UNESCO and the KIX Africa 21 Hub began supporting national AI/digital competency frameworks and translation of digital learning resources in four pilots—Benin, Burundi, Côte d’Ivoire, and Senegal—in 2025; Côte d’Ivoire had begun workshops and translation work and Benin had committed to a national framework.Evidence B · P1
  19. MWITA-FT-2026-003In the prospective randomized MASAI study, AI-supported mammography screening detected more cancers and reduced screen-reading workload by 44.2% without a statistically significant increase in false positives in the reported secondary analysis.Evidence A · P1
  20. MWITA-FT-2026-004Rentosertib, whose target and molecule were described as generated through AI methods, completed a multicentre randomized placebo-controlled phase 2a trial in China; the primary endpoint was safety and tolerability, while lung-function findings were exploratory.Evidence B · P1
  21. MWITA-FT-2026-006ECMWF began operational deterministic AIFS forecasts in February 2025 and ensemble forecasts in July 2025, disseminating daily products while verifying them alongside physics-based IFS; WMO still identifies gaps for local high-impact weather and hydrology.Evidence A · P1
  22. MWITA-FT-2026-008MLCommons' AILuminate v0.5 jailbreak evaluation reported that 35 of 39 de-identified systems received a lower grade under jailbreak attacks than on the naive safety benchmark; the benchmark is English, single-turn and content-hazard focused.Evidence B · P1
  23. MWITA-FT-2026-009China's MIIT created a humanoid-robot and embodied-intelligence standardization committee in December 2025 covering components, complete systems, applications and safety.Evidence A · P1
  24. MWITA-FT-2026-017A staggered enterprise deployment across 5,172 customer-support agents increased issues resolved per hour by about 15% on average, with larger gains for less experienced and lower-skill workers.Evidence A · P1
  25. MWITA-FT-2026-018METR's randomized controlled trial of 16 experienced open-source developers completing 246 tasks in familiar repositories found 19% longer completion time with early-2025 AI tools, despite participants expecting and perceiving speedups.Evidence B · P1
  26. MWITA-FT-2026-020A six-week randomized trial in Nigeria using Microsoft Copilot with teacher facilitation improved an assessment covering English and related skills by 0.31 standard deviations.Evidence A · P1
  27. MWITA-LC-2026-002Brazil's Ministry of Health reported that Meu SUS Digital's digital child-health booklet had more than 2 million accesses in 2025 and that 250 million supplementary-health records were available for viewing in the app.Evidence A · P1
  28. MWITA-LC-2026-007Colombia's Ministry of Health reported that, less than two months after the national electronic-health-record interoperability mechanism began operation on 15 April 2026, 744 providers across 2,456 sites had exchanged more than 7.7 million digital care summaries concerning about 2.7 million patients.Evidence A · P1
  29. MWITA-ME-2026-001Saudi Arabia's National Data Governance Platform offers an AI Service Provider Accreditation process requiring entity registration, appointment of an AI officer, a compliance questionnaire and supporting files before committee review and certificate issuance.Evidence A · P0
  30. MWITA-ME-2026-002SDAIA published separate generative-AI guidance for government entities and the public on 15 May 2025, emphasizing governance, risk identification, privacy, human rights and responsible use.Evidence A · P0
  31. MWITA-ME-2026-003Egypt's second National AI Strategy covers 2025–2030 and organizes action around governance, technology, data, infrastructure, ecosystem and talent.Evidence A · P0
  32. MWITA-ME-2026-004Jordan's adopted AI Strategy and Implementation Plan 2023–2027 contains 68 projects tied to five objectives spanning capacity, research, investment, regulation/data and public-sector applications.Evidence A · P1
  33. MWITA-ME-2026-017Morocco's Digital Morocco 2030 strategy has two axes—digital public services and a value- and job-producing digital economy—with talent, cloud and connectivity accelerators and AI as a cross-cutting technology.Evidence A · P1
  34. MWITA-ME-2026-019Abu Dhabi Department of Health's 2025 Responsible AI Standard sets minimum lifecycle requirements for DoH and all licensed healthcare entities developing, procuring or deploying AI across clinical, financial, administrative and related functions.Evidence A · P0
  35. MWITA-ME-2026-023Dubai's 2025 State of AI Report assessed more than 100 high-impact government AI use cases that were deployed or in development across urban planning, health, mobility, finance, procurement, hiring, compliance and decision support.Evidence B · P1
  36. MWITA-ME-2026-025DialectalArabicMMLU manually adapted 3,000 question-answer pairs into each of five dialects—Syrian, Egyptian, Emirati, Saudi and Moroccan—across 32 domains and found substantial performance variation among 19 tested open-weight models.Evidence A · P0
  37. MWITA-ME-2026-026TounsiBench evaluated ten widely used LLMs on 744 Tunisian-Arabic instructions with human-written references and found that most tested models struggled to recognize and respond in Tunisian Arabic.Evidence A · P1
  38. MWITA-ME-2026-027Atlas-Chat reports that its 9B Moroccan-Darija-adapted model outperformed a larger 13B comparator by 13% on the paper's DarijaMMLU suite and improved both discriminative and generative tasks.Evidence A · P1
  39. MWITA-ME-2026-028AraDiCE introduced seven post-edited dialect datasets plus Modern Standard Arabic and a cultural benchmark for Gulf, Egypt and Levant contexts; Arabic-specific models performed better on dialectal tasks but still showed material identification, generation and translation gaps.Evidence A · P1
  40. MWITA-ME-2026-029Morocco launched a national digital and AI training program for children in 2025 through a partnership among three ministries and the UNESCO-category-II AI Movement centre.Evidence B · P2
  41. MWITA-RD-2026-016FDA issued draft guidance on 7 January 2025 recommending lifecycle risk-management and marketing-submission information for AI-enabled device software functions; it is explicitly draft, non-binding and not for implementation.Evidence A · P1
  42. MWITA-WD-2026-001The European Commission published a mandatory-format template on 24 July 2025 establishing a common minimum baseline for public summaries of content used to train general-purpose AI models under the AI Act.Evidence A · P0
  43. MWITA-WD-2026-002The European Commission published the General-Purpose AI Code of Practice on 10 July 2025 and endorsed it as an adequate voluntary route for signatories to demonstrate compliance with relevant AI Act transparency, copyright, safety and security obligations.Evidence A · P0
  44. MWITA-WD-2026-003The European Commission released an open-source white-label age-verification blueprint in July 2025; initial pilots in Denmark, France, Greece, Italy and Spain use anonymous proof-of-age patterns compatible with future European Digital Identity Wallets.Evidence A · P1
  45. MWITA-WD-2026-004The ECB's 2025 digital-euro innovation platform involved more than 70 public and private participants testing or exploring capabilities including conditional payments for mobility and e-commerce, while issuance remained contingent on legislation and a later Governing Council decision.Evidence A · P1
  46. MWITA-WD-2026-005From 19 May 2026, the US TAKE IT DOWN Act requires covered platforms to provide a notice-and-removal process and remove validly reported nonconsensual intimate imagery, including AI-created digital forgeries, and known identical copies within 48 hours.Evidence A · P0
  47. MWITA-WD-2026-006The US Copyright Office's May 2025 pre-publication Part 3 report analyzes generative-AI training under existing copyright law and licensing markets; it is an expert agency policy analysis, not legislation or a final court ruling that all training is fair use or infringement.Evidence A · P0
  48. MWITA-WD-2026-007FDA's January 2025 draft guidance proposes lifecycle and marketing-submission recommendations for AI-enabled device software, while expressly remaining draft, non-binding and not for implementation.Evidence A · P0
  49. MWITA-WD-2026-008The July 2025 White House AI Action Plan lists more than 90 federal policy actions across innovation, infrastructure and international diplomacy/security.Evidence A · P1
  50. MWITA-WD-2026-010Canada's Budget 2025 framework proposes Bank of Canada-supervised consumer-driven banking with accreditation, consent, liability and security rules, while later work would consider write access such as payment initiation.Evidence A · P0
  51. MWITA-WD-2026-012Federal, provincial and territorial health governments except Quebec endorsed Pan-Canadian AI for Health guiding principles published in January 2025, covering privacy/security, lifecycle safety, accountability, transparency and AI literacy.Evidence A · P1
  52. MWITA-WD-2026-013Canada launched its first AI Strategy for the federal public service in March 2025; it is a government-operating roadmap rather than a law governing all Canadian private-sector AI.Evidence A · P1
  53. MWITA-WD-2026-014Canada's Digital Regulators Forum published a September 2025 cross-regulator analysis mapping synthetic-media risks across privacy, competition, communications, copyright and automated-decision frameworks without creating a new unified synthetic-media law.Evidence A · P1
  54. MWITA-WD-2026-015The UK Data (Use and Access) Act 2025 received Royal Assent on 19 June 2025 and creates statutory machinery for digital verification services, Smart Data schemes and changes to data-protection/privacy law.Evidence A · P0
  55. MWITA-WD-2026-016From 25 July 2025, in-scope UK user-to-user and search services likely to be accessed by children had to apply Ofcom's Protection of Children Codes or effective alternatives, including risk assessment and proportionate measures around age assurance, moderation, recommender systems, reporting and complaints.Evidence A · P0
  56. MWITA-WD-2026-019Ofsted's June 2025 study of 21 early-adopter schools, colleges and trusts found workload-oriented AI uses and governance practices, but concluded that robust evidence of educational impact remained limited, explorative and short-term.Evidence A · P1
  57. MWITA-WD-2026-020From 10 December 2026, Australian Privacy Principle entities using personal information in computer-assisted decisions that could significantly affect individual rights or interests must disclose specified kinds of information about those automated decisions in their privacy policies.Evidence A · P0
  58. MWITA-WD-2026-021Australia launched a National AI Plan on 2 December 2025 with goals to build capability and infrastructure, spread adoption and worker/public-service benefits, and mitigate harms through regulatory and responsible-practice measures.Evidence A · P1
  59. MWITA-WD-2026-022The ACCC's final Digital Platform Services Inquiry report, released in June 2025, identified emerging competition and consumer issues in generative AI alongside messaging, app stores, mobile operating systems, ad tech, marketplaces, cloud and gaming, and recommended continued monitoring and broader reform.Evidence A · P1
  60. MWITA-WD-2026-023Project Acacia's 2026 final report records 12 real-money/real-asset pilots and eight simulated proofs of concept conducted from August 2025 to February 2026, finding potential benefits but unresolved coordination, legal and regulatory barriers to scaling tokenised markets.Evidence A · P1
  61. MWITA-WD-2026-024New Zealand's Customer and Product Data Act 2025 received assent on 29 March 2025 and establishes a statutory framework for designated customer data, data holders and accredited requestors, with practical operation dependent on commencement and sector designations.Evidence A · P0
  62. MWITA-WD-2026-025New Zealand's Biometric Processing Privacy Code 2025 came into force on 3 November 2025; Amendment No. 1 reflected new notification principle IPP3A from May 2026, and the transition for existing biometric users ended on 3 August 2026.Evidence A · P0
  63. MWITA-WD-2026-026New Zealand released its first national AI strategy in July 2025, explicitly prioritising adoption and application over foundational-model development and pairing it with voluntary responsible-AI guidance for businesses.Evidence A · P1
  64. MWITA-ST-2026-007Google reported that an automated quota-policy update containing unintended blank fields replicated globally within seconds and drove Service Control binaries into crash loops, producing a three-hour global incident across multiple Cloud and Workspace products.Evidence A · P1
  65. MWITA-ST-2026-008During the Google Cloud incident, the provider's service-health infrastructure and some customer monitoring on Google Cloud also failed, delaying the first incident report and leaving some customers without a signal.Evidence A · P1
  66. MWITA-SI-2026-010The OpenAPI Initiative released OpenAPI Specification 3.2.0 on 19 September 2025.Evidence A · P1
  67. MWITA-SI-2026-012The GraphQL project published the September 2025 edition of its language and execution specification.Evidence A · P1
  68. MWITA-SI-2026-013The 28 July 2026 MCP specification introduced a stateless protocol core and retired the initialize/initialized exchange and Mcp-Session-Id header.Evidence A · P1
  69. MWITA-SI-2026-014The MCP project reported close to half a billion monthly downloads across Tier-1 SDKs in July 2026, with TypeScript and Python SDKs each exceeding one billion cumulative downloads.Evidence B · P1
  70. MWITA-SI-2026-016The Linux Foundation launched the Agent2Agent protocol project on 23 June 2025 and reported support from more than 100 technology companies.Evidence B · P1
  71. MWITA-SI-2026-017The Linux Foundation reported A2A support from more than 150 organizations and production deployments across multiple industries by April 2026.Evidence B · P1
  72. MWITA-SST-2026-003ISRO reported that SpaDeX satellites docked on 16 January and undocked on 13 March 2025 in a 460 km circular orbit.Evidence A · P1
  73. MWITA-SST-2026-004ISRO reported autonomous redocking from 15 metres on 20 April 2025 and approximately four minutes of bidirectional power transfer on 21 April.Evidence A · P1
  74. MWITA-SST-2026-005CNSA reported that Tianwen-2 launched on 29 May 2025, entered the transfer orbit toward asteroid 2016 HO3 and deployed its solar arrays normally.Evidence A · P1
  75. MWITA-SST-2026-006CNSA describes asteroid sampling and Earth return, followed by exploration of comet 311P, as mission objectives rather than completed outcomes.Evidence B · P1
  76. MWITA-SST-2026-008The UAE Space Agency reported that the Emirates Mission to the Asteroid Belt completed its critical design review in February 2025.Evidence A · P1
  77. MWITA-SST-2026-009The official mission plan describes a six-year development phase, launch, seven asteroid flybys and a final landing on Justitia; these are scheduled objectives, not observed outcomes.Evidence B · P1
  78. MWITA-SST-2026-010JAXA reported HTV-X1 launched on H3 F7 on 26 October 2025 and confirmed communications and attitude control after separation.Evidence A · P1
  79. MWITA-QE-2026-001USTC reports a 105-qubit superconducting prototype and benchmark fidelities for Zuchongzhi 3.0.Evidence ungraded · P1
  80. MWITA-QE-2026-002USTC reports one million random-circuit samples in hundreds of seconds and estimates 5.9 billion years for a classical supercomputer on that benchmark.Evidence ungraded · P1
  81. MWITA-QE-2026-003RIKEN launched collaborative operation linking Fugaku with an IBM Quantum System Two in July 2025.Evidence ungraded · P1
  82. MWITA-QE-2026-004India's National Quantum Mission was approved with a ₹6,003.65 crore envelope for 2023-24 through 2030-31.Evidence ungraded · P1
  83. MWITA-QE-2026-005India's Department of Science and Technology reports that eight quantum startups were selected for support under the National Quantum Mission.Evidence ungraded · P1
  84. MWITA-QE-2026-006The Chinese Academy of Sciences reports EAST sustained high-confinement plasma operation for 1,066 seconds in January 2025.Evidence ungraded · P1
  85. MWITA-QE-2026-007ITER reported in June 2026 that the final module of the central solenoid had been stacked.Evidence ungraded · P1
  86. MWITA-QE-2026-008The US Department of Energy announced selections totaling $107 million for Fusion Innovation Research Engine projects.Evidence ungraded · P1
  87. MWITA-QE-2026-009DOE reported $46 million initially obligated under its milestone program and early company milestone progress, while the underlying roadmaps remain company-specific and pre-commercial.Evidence ungraded · P1
  88. MWITA-QE-2026-014India's Department of Atomic Energy reported continued integrated commissioning and fuel loading for the prototype fast breeder reactor as of January 2026.Evidence ungraded · P1
  89. MWITA-QE-2026-027The European Innovation Council selected 71 companies in February 2025, with up to €161 million in grants and approximately €226 million in proposed equity investments.Evidence ungraded · P1
  90. MWITA-QE-2026-028The EIC selected 40 startups in June 2025 for almost €230 million in proposed combined support.Evidence ungraded · P1
  91. MWITA-QE-2026-029The EIC 2025 Impact Report states that the EIC Fund had invested more than €1 billion cumulatively.Evidence ungraded · P1
  92. MWITA-QE-2026-030Japan's Ministry of Economy, Trade and Industry selected 31 companies for the J-Startup program in March 2025.Evidence ungraded · P1
  93. MWITA-SE-2026-003ISRO reported successful SpaDeX docking and undocking in a 460 km orbit by March 2025.Evidence A · P1
  94. MWITA-SE-2026-004SpaDeX completed a second docking and about four minutes of bidirectional power transfer in April 2025.Evidence A · P1
  95. MWITA-SE-2026-005India formalized a public-private agreement in January 2025 to establish a 12-satellite Earth-observation constellation.Evidence A · P1
  96. MWITA-SE-2026-006ISRO reported NISAR launched successfully on 30 July 2025.Evidence A · P1
  97. MWITA-SE-2026-007ISRO reported CMS-03, a multi-band communications satellite, placed into geosynchronous transfer orbit in November 2025.Evidence A · P1
  98. MWITA-SE-2026-008India launched the commercial BlueBird Block-2 communications satellite into its intended orbit in December 2025.Evidence A · P1
  99. MWITA-SE-2026-009ISRO reported seven successful missions out of ten launch-vehicle and spacecraft missions in 2025.Evidence A · P1
  100. MWITA-SE-2026-010ISRO's registry listed NISAR as operational after its July 2025 launch.Evidence A · P1
  101. MWITA-SE-2026-011ISRO's registry labels the May 2025 EOS-09 launch unsuccessful.Evidence A · P1
  102. MWITA-SE-2026-012ISRO's registry labels the January 2026 EOS-N1 launch unsuccessful.Evidence A · P1
  103. MWITA-SE-2026-014The UAE asteroid-belt mission passed critical design review in March 2025.Evidence A · P1
  104. MWITA-SE-2026-015The UAE mission's launch, asteroid encounters and 2034 landing remained prospective after design review.Evidence A · P1
  105. MWITA-SE-2026-021ITU's 2025 satellite-broadband report identifies direct-to-device connectivity as an emerging route to ordinary-handset coverage.Evidence A · P1
  106. MWITA-SE-2026-033Brazil and China agreed CBERS-06 calibration, validation, commissioning-test and SAR operations planning in May 2026.Evidence A · P1
  107. MWITA-WS-2026-023TSMC reported reclaimed-water supply capacity of 67,000 cubic metres per day in southern Taiwan in 2024.Evidence A · P1
  108. MWITA-WS-2026-024TSMC reported consuming 19.65 million cubic metres of reclaimed water at its Tainan sites in 2024, reducing municipal tap-water use there by 31 percent.Evidence A · P1
  109. MWITA-WS-2026-025TSMC's additional reclaimed-water facilities and 2030 replacement-rate objectives are targets rather than completed outcomes.Evidence A · P1
  110. MWITA-ER-2026-003China's official series reported 18,581,081 service robots produced in 2025, up 16.1% year on year.Evidence A · P1
  111. MWITA-ER-2026-008Automotive applications accounted for 4,070 of India's industrial-robot installations in 2024, 45% of the national total.Evidence A · P1
  112. MWITA-ER-2026-015METI and NEDO selected seven R&D themes in 2025 for open robot-software and platform development across two projects running through 2027 or 2029.Evidence A · P1
  113. MWITA-ER-2026-016Japan's open-robotics projects explicitly target common software functions and development environments to reduce duplicated development effort.Evidence A · P1
  114. MWITA-ER-2026-017South Korea's 2025 industrial-AI R&D plan covered 445 projects with KRW 478.7 billion, including 31 intelligent-robotics projects with KRW 29.6 billion.Evidence A · P1
  115. MWITA-ER-2026-019South Korea opened a 2025 call for shared AI infrastructure spanning autonomous laboratories, humanoids and autonomous manufacturing, with support of up to KRW 10 billion per project over five years.Evidence A · P1
  116. MWITA-ER-2026-021MIIT described development of an embodied-intelligence open-source community and a comprehensive standardization system as policy priorities.Evidence A · P1
  117. MWITA-ER-2026-022A Chinese government portal reported 130 welding robots operating at FAW Jiefang's Qingdao production site in 2025.Evidence A · P1
  118. MWITA-ER-2026-024Google DeepMind introduced Gemini Robotics, a vision-language-action model, and Gemini Robotics-ER, a spatial-reasoning model, in March 2025.Evidence A · P1
  119. MWITA-ER-2026-025DeepMind reported that Gemini Robotics more than doubled average performance over other state-of-the-art VLA models on its generalization benchmark.Evidence A · P1
  120. MWITA-ER-2026-027NVIDIA released the 2-billion-parameter GR00T N1 model and sample training code for generalist humanoid development in March 2025.Evidence A · P1
  121. MWITA-ER-2026-028NVIDIA reported a 76.8% average success rate for GR00T N1 across its real-world task evaluation with the full training-data mix.Evidence A · P1
  122. MWITA-ER-2026-029NVIDIA reported that adding 750,000 synthetic trajectories to real data improved GR00T N1 performance by 40% relative to its real-data-only condition.Evidence A · P1
  123. MWITA-ER-2026-030Physical Intelligence evaluated pi0.5 on long-horizon manipulation across multiple robot embodiments and both laboratory and previously unseen household environments.Evidence A · P1
  124. MWITA-ER-2026-031The pi0.5 study combines heterogeneous robot data with web-scale vision-language pretraining to test open-world generalization.Evidence A · P1
  125. MWITA-ER-2026-040LIBERO-Plus evaluates VLA robustness under controlled changes in object layout, camera viewpoint, robot initial state, language, lighting, background texture and sensor noise.Evidence A · P1
  126. MWITA-ER-2026-041LIBERO-Plus reports that high aggregate VLA benchmark scores can coexist with material failures under realistic visual, language and sensor variation.Evidence A · P1
  127. MWITA-ER-2026-045UBTECH reported participation in drafting Chinese technical requirements for humanoid operational tasks, positioning/navigation and human-computer interaction.Evidence A · P1
  128. MWITA-AD-2026-005Alibaba launched a Qwen-powered version of Quark on 13 March 2025 as an all-in-one AI assistant with chat, reasoning, research and task-execution functions.Evidence B · P1
  129. MWITA-AD-2026-019Google began rolling AI Mode out to all users in India in English on 8 July 2025 without requiring Labs enrollment.Evidence B · P1
  130. MWITA-AD-2026-020The India rollout of Google AI Mode allowed text, voice and image input and follow-up questions.Evidence B · P1
  131. MWITA-AD-2026-021Google began rolling AI Mode in Hindi out globally on 8 September 2025 after its English launch in India.Evidence B · P1
  132. MWITA-AD-2026-022Google described Hindi AI Mode as using a custom Gemini 2.5 version and supporting text, voice and image queries.Evidence B · P1
  133. MWITA-BMF-2026-007India received 284 letters of intent for Bio-AI hubs and recommended 15 hubs for support by July 2026.Evidence A · P1
  134. MWITA-EN-2026-001Intel's Hala Point research system packages 1,152 Loihi 2 processors and supports up to 1.15 billion neurons and 128 billion synapses.Evidence B · P1
  135. MWITA-EN-2026-002Intel reported Hala Point characterization of up to 20 peta-operations per second for conventional deep neural networks.Evidence B · P1
  136. MWITA-EN-2026-003Intel reported efficiency above 15 TOPS/W on a sparse, low-activity multilayer-perceptron characterization of Hala Point.Evidence B · P1
  137. MWITA-EN-2026-004Hala Point was delivered to Sandia National Laboratories as a research system, not released as a commercial general-purpose accelerator.Evidence B · P1
  138. MWITA-EN-2026-005IBM's NorthPole research chip eliminates off-chip memory for inference by interweaving compute and memory on-chip.Evidence A · P1
  139. MWITA-EN-2026-006On ResNet-50, NorthPole achieved 25 times the frames-per-second per watt of a GPU fabricated on a comparable 12-nm process in IBM's study.Evidence A · P1
  140. MWITA-EN-2026-007On the same ResNet-50 comparison, IBM reported NorthPole at 22 times lower latency and five times higher frames per second per transistor than the matched-process GPU.Evidence A · P1
  141. MWITA-EN-2026-008Tsinghua researchers demonstrated Tianjic as a many-core chip able to run both conventional artificial neural networks and brain-inspired spiking models.Evidence A · P1
  142. MWITA-EN-2026-009The Tianjic study integrated multiple neural networks on one chip to control an unmanned bicycle in a research demonstration.Evidence A · P1
  143. MWITA-EN-2026-010The Tianjic chip used a 28-nm process and contained 156 reconfigurable functional cores.Evidence A · P1
  144. MWITA-EN-2026-011Darwin3 implements a neuromorphic instruction-set architecture with ten primary instructions plus extensions for programmable neuron models and local learning rules.Evidence A · P1
  145. MWITA-EN-2026-012Darwin3 supports up to 2.35 million neurons on a 24-by-24 mesh of computing nodes.Evidence A · P1
  146. MWITA-EN-2026-013Darwin3 experiments reported code-density improvement up to 28.3 times and effective fan-in and fan-out improvements up to 4,096 and 3,072 times relative to physical memory depth.Evidence A · P1
  147. MWITA-EN-2026-014Darwin3 reported 6.8-to-200.8-times memory savings when mapping selected convolutional spiking neural networks.Evidence A · P1
  148. MWITA-EN-2026-015A Taiwan-led 2025 Nature study fabricated an edge processor combining memristor-CIM, SRAM-CIM and small digital units with integer and floating-point computation.Evidence A · P1
  149. MWITA-EN-2026-016The heterogeneous CIM processor measured 40.91 TFLOPS/W on ResNet-20 with CIFAR-100 and 28.63 TFLOPS/W on MobileNet-v2 with ImageNet.Evidence A · P1
  150. MWITA-EN-2026-017The 2025 heterogeneous CIM study reported less than 0.45% accuracy degradation on its tested ResNet-20 and MobileNet-v2 tasks and a 373.52-microsecond wake-to-response time.Evidence A · P1
  151. MWITA-EN-2026-018KAIST researchers demonstrated a selector-less 32-by-32 memristor crossbar system performing self-supervised video foreground-background separation with on-device self-calibration.Evidence A · P1
  152. MWITA-EN-2026-019KAIST reported that self-calibration brought real-time image-processing accuracy close to ideal computer simulation despite memristor non-idealities.Evidence A · P1
  153. MWITA-EN-2026-020Renesas launched the RZ/V2H MPU with a DRP-AI3 accelerator rated by the vendor at up to 10 TOPS/W and 80 sparse TOPS.Evidence B · P1
  154. MWITA-EN-2026-021RZ/V2H integrates four Cortex-A55 application cores, two Cortex-R8 real-time cores, one Cortex-M33 core and the AI accelerator on one MPU.Evidence A · P1
  155. MWITA-EN-2026-022Renesas states RZ/V2H can run its target vision-AI workloads without a cooling fan.Evidence B · P1
  156. MWITA-EN-2026-023Sony's IMX500 combines a CMOS image sensor, image processing and AI processing on one stacked sensor device.Evidence A · P1
  157. MWITA-EN-2026-024IMX500 can output processed metadata without requiring raw image data to leave the chip.Evidence A · P1
  158. MWITA-EN-2026-025Sony specifies IMX500 AI processing and metadata output at up to 30 frames per second, TensorFlow Lite support and a maximum AI input tensor of 640 by 480 pixels.Evidence A · P1
  159. MWITA-EN-2026-026MediaTek launched Dimensity 9400 in October 2024 with an NPU and a developer framework for on-device agentic AI.Evidence B · P1
  160. MWITA-EN-2026-027MediaTek claims Dimensity 9400 supports on-device LoRA training and on-device video generation.Evidence B · P1
  161. MWITA-EN-2026-028MediaTek reports the Dimensity 9400 NPU uses 35% less power than the Dimensity 9300 NPU under its internal tests.Evidence B · P1
  162. MWITA-EN-2026-029Google designed Gemma 3n for phones, laptops and tablets with text, image, video and audio input support.Evidence A · P1
  163. MWITA-EN-2026-030Google states Gemma 3n E2B can reduce its effective loaded parameter set from more than five billion to 1.91 billion through PLE caching and parameter skipping.Evidence A · P1
  164. MWITA-EN-2026-031Gemma 3n documentation says training covered more than 140 languages and provides a 32K-token context window.Evidence A · P1
  165. MWITA-EN-2026-032Apple described its 2024 on-device foundation model as approximately three billion parameters and optimized for Apple silicon.Evidence B · P1
  166. MWITA-EN-2026-033Apple reported mixed 2-bit and 4-bit palletization averaging 3.7 bits per weight achieved the same accuracy as its uncompressed on-device model in internal evaluation.Evidence B · P1
  167. MWITA-EN-2026-034Apple measured about 30 generated tokens per second and roughly 0.6 milliseconds per prompt token to first token on an iPhone 15 Pro before speculative decoding.Evidence B · P1
  168. MWITA-EN-2026-035MLPerf Tiny fixes models, datasets and minimum accuracy targets while measuring latency and energy per inference for constrained systems.Evidence A · P1
  169. MWITA-EN-2026-036MLPerf Tiny v1.3 added a streaming workload because prior energy tests did not accurately represent duty-cycled always-on deployments.Evidence A · P1
  170. MWITA-EN-2026-037SynSense's Speck integrates a dynamic vision sensor and spiking-neural-network processor on one chip and activates computation on detected visual changes.Evidence A · P1
  171. MWITA-EN-2026-038SynSense specifies Speck 2f at about one milliwatt and its development platform at 320,000 neurons with support for convolutional networks up to nine layers.Evidence A · P1
  172. MWITA-EN-2026-039Google introduced Coral NPU in October 2025 as an open, RISC-V-based reference architecture and software stack for low-power edge AI.Evidence B · P1
  173. MWITA-EN-2026-040Google says the Coral NPU base design targets roughly 512 GOPS at a few milliwatts, while noting its matrix execution unit was still under development at announcement.Evidence B · P1
  174. MWITA-AM-2026-001Hong Kong's autonomous-vehicle ordinance and subsidiary regulation took effect on 1 March 2024.Evidence A · P1
  175. MWITA-AM-2026-002Hong Kong requires pilot licensing, vehicle certification, examination, registration, insurance, maintenance and incident reporting for autonomous-vehicle trials.Evidence A · P1
  176. MWITA-AM-2026-003During 2024 Hong Kong licensed two autonomous-vehicle pilot projects involving 12 vehicles at two locations.Evidence A · P1
  177. MWITA-AM-2026-004Hong Kong's current trial register lists a Baidu licence covering 30 private autonomous cars in North Lantau, valid from 9 December 2024 to 8 December 2029.Evidence A · P1
  178. MWITA-AM-2026-005Hong Kong's trial register lists separate Baidu authorizations for ten private cars in Southern District from August 2025 and ten in Kowloon East from November 2025.Evidence A · P1
  179. MWITA-AM-2026-007Beijing approved paid autonomous shuttle services to Beijing South Railway Station on 13 March 2025.Evidence A · P1
  180. MWITA-AM-2026-008The two approved Beijing South Railway Station shuttle routes were reported as 16 kilometres and 29.3 kilometres within the pilot zone.Evidence A · P1
  181. MWITA-AM-2026-009Beijing E-Town approved WeRide's GXR for fully driverless commercial ride-hailing in February 2025.Evidence A · P1
  182. MWITA-AM-2026-010The approved Beijing E-Town GXR service specified operating windows of 07:30–22:00 on weekdays and 09:00–17:00 on weekends and holidays.Evidence A · P1
  183. MWITA-AM-2026-012Beijing's regulation provides for autonomous vehicles in passenger transport, public buses and urban operations.Evidence A · P1
  184. MWITA-AM-2026-014Wuhan's commercial-pilot rules require warning functions and redundant systems for autonomous vehicles.Evidence A · P1
  185. MWITA-AM-2026-015Wuhan requires a commercial-pilot autonomous vehicle to transition to a minimum-risk safe stop when a relevant failure occurs or it leaves its operational design domain.Evidence A · P1
  186. MWITA-AM-2026-017Singapore's Land Transport Authority announced three autonomous shuttle routes in Punggol in September 2025.Evidence A · P1
  187. MWITA-AM-2026-018Singapore's January 2025 autonomous-vehicle application requirements call for multiple operator control methods, including braking, acceleration and steering.Evidence A · P1
  188. MWITA-AM-2026-019Singapore's application requirements specify that an autonomous vehicle must operate safely under conditions representative of Singapore public roads.Evidence A · P1
  189. MWITA-AM-2026-020Japan's transport ministry listed eight locations with Level 4 road-vehicle approvals as of May 2025.Evidence A · P1
  190. MWITA-AM-2026-021Japan's May 2025 Level 4 inventory states that most approved operations covered only part of a route or a dedicated space, including private premises.Evidence A · P1
  191. MWITA-AM-2026-022Japan's Eiheiji Level 4 service in the May 2025 inventory covered a two-kilometre route at about 12 km/h and used remote, unmanned supervision plus guide infrastructure.Evidence A · P1
  192. MWITA-AM-2026-023South Korea's transport ministry amended its administrative rules on autonomous-vehicle safe-operation requirements and test operation on 26 January 2026.Evidence A · P1
  193. MWITA-AM-2026-025Dubai launched a public robotaxi pilot bookable through the Uber app in Umm Suqeim and Jumeirah on 11 December 2025.Evidence A · P1
  194. MWITA-AM-2026-026Dubai's December 2025 public robotaxi pilot operated with a safety driver ahead of a planned driverless phase.Evidence A · P1
  195. MWITA-AM-2026-027Dubai's RTA described successive testing between the April 2025 partnership announcement and the December 2025 public pilot.Evidence A · P1
  196. MWITA-AM-2026-028California autonomous-vehicle testing permit holders reported more than nine million public-road test miles from 1 December 2024 through 30 November 2025.Evidence A · P1
  197. MWITA-AM-2026-031Zoox reported a 2025 recall covering ADS software used on 270 drivered test vehicles and driverless robotaxis.Evidence A · P1
  198. MWITA-AM-2026-032The Zoox recall identified inaccurate prediction of a vehicle creeping from a perpendicular driveway and stopping as a condition that could delay collision avoidance.Evidence A · P1
  199. MWITA-AM-2026-033Zoox stated that all public-road vehicles had received remedied software by 17 April 2025.Evidence A · P1
  200. MWITA-AM-2026-034A 2025 peer-reviewed Waymo-authored retrospective study examined 56.7 million rider-only miles through January 2025.Evidence A · P1
  201. MWITA-AM-2026-035The Waymo study found statistically lower crashed-vehicle rates than aligned human benchmarks for any-injury-reported, airbag-deployment and suspected-serious-injury-or-worse outcomes.Evidence A · P1
  202. MWITA-AM-2026-036The Waymo study reported no statistically significant disbenefit in any of 11 analyzed crash-type groups.Evidence A · P1
  203. MWITA-AM-2026-037Serve Robotics reported 2,000 deployed sidewalk delivery robots across 20 US cities at year-end 2025.Evidence A · P1
  204. MWITA-AM-2026-039Serve disclosed that remote human supervisors monitor its delivery fleet and can assist at intersections or when robots cannot navigate conditions.Evidence A · P1
  205. MWITA-AM-2026-041Participants with mobility disabilities in the study described having to compete for sidewalk space when public robots are introduced, while practitioners reported that accessibility is often addressed after problems arise.Evidence A · P1
  206. MWITA-SC-2026-001India reported more than 45,000 GPUs in its shared national compute pool as of June 2026.Evidence A · P1
  207. MWITA-SC-2026-002India reported that 237 projects had accessed subsidized AI computing by August 2026.Evidence A · P1
  208. MWITA-SC-2026-003India reported cumulative subsidized usage of 9.318 million GPU-hours by August 2026.Evidence A · P1
  209. MWITA-SC-2026-004India reported selecting 20 indigenous foundation-model proposals from 506 applications by August 2026.Evidence A · P1
  210. MWITA-SC-2026-005India's July 2025 disclosure listed 34,381 GPUs provisioned through the IndiaAI Compute Portal.Evidence A · P1
  211. MWITA-SC-2026-007India reported an average portal price of about ₹67 per GPU-hour in July 2025.Evidence A · P1
  212. MWITA-SC-2026-008India's portal offered multiple accelerator types, including Nvidia H100/H200/B200, Intel Gaudi 2/3 and AMD MI300X.Evidence A · P1
  213. MWITA-SC-2026-014Brazil's plan allocates R$1.8 billion for a specialized AI supercomputer action during 2024–2028.Evidence B · P1
  214. MWITA-SC-2026-015Brazil's plan targets upgrading national AI supercomputing into the TOP500 top five within five years.Evidence B · P1
  215. MWITA-SC-2026-019Kenya's strategy also identifies the number of local standard-compliant data centres and GPUs as infrastructure KPIs.Evidence A · P1
  216. MWITA-SC-2026-020Kenya's strategy includes edge-computing capability as a separate infrastructure objective.Evidence B · P1
  217. MWITA-SC-2026-021Kenya's implementation roadmap specifies at least 400 PFLOPS for its first planned AI data centre.Evidence B · P1
  218. MWITA-SC-2026-022Kenya's roadmap targets total computing power above 1,000 PFLOPS across three centres by 2030.Evidence B · P1
  219. MWITA-SC-2026-025Japan's AIST started operating ABCI 3.0 on 20 January 2025.Evidence A · P1
  220. MWITA-SC-2026-026AIST reported ABCI 3.0 AI processing performance of 6.22 EFLOPS.Evidence A · P1
  221. MWITA-SC-2026-027AIST described ABCI 3.0 performance as more than seven times the preceding system.Evidence A · P1
  222. MWITA-SC-2026-029Korea revised project requirements after private-sector concerns about public ownership, put-option risk and mandatory domestic-chip adoption.Evidence A · P1
  223. MWITA-SC-2026-030Korea planned more than 50,000 GPUs through public-private collaboration across procurement, a sixth supercomputer and a national centre.Evidence B · P1
  224. MWITA-SC-2026-031Korea's revised plan assigned more than 15,000 GPUs to the National AI Computing Center by 2028.Evidence B · P1
  225. MWITA-SC-2026-034The Korean centre is intended to support industry, academia and research and promote domestic AI semiconductors.Evidence B · P1
  226. MWITA-SC-2026-035China's National Data Administration reported total intelligent computing capacity of 780,000 PFLOPS in August 2025.Evidence A · P1
  227. MWITA-SC-2026-037China reported that eight national computing hubs concentrated more than 60 percent of new computing capacity and 620,000 PFLOPS of intelligent compute.Evidence A · P1
  228. MWITA-SC-2026-047Abu Dhabi announced Stargate UAE as a planned 1-GW compute cluster within a planned 5-GW UAE–US AI campus.Evidence B · P1
  229. MWITA-SC-2026-048The first 200-MW Stargate UAE cluster was expected to go live in 2026.Evidence B · P1
  230. MWITA-SC-2026-050Saudi Arabia's PIF launched HUMAIN in May 2025 as a company spanning data centres, cloud, models and AI applications.Evidence A · P1
  231. MWITA-SC-2026-051PIF stated that HUMAIN would streamline data-centre initiatives and procure hardware.Evidence B · P1
  232. MWITA-SC-2026-057China's National Energy Administration counted 42 intelligent-computing clusters with at least 10,000 accelerators by May 2026.Evidence A · P1
  233. MWITA-ASVC-2026-001JD.com reported that its multimodal AI customer-service system handled more than 4.2 billion inquiries during the 2025 11.11 promotion.Evidence B · P1
  234. MWITA-ASVC-2026-002JD.com described that customer-service system as both multi-scenario and multimodal.Evidence B · P1
  235. MWITA-ASVC-2026-003JD.com reported more than 50,000 AI agents across internal systems by the end of 2025.Evidence B · P1
  236. MWITA-ASVC-2026-015Safaricom reported 38 NLP Smart-IVR journeys developed in FY2024.Evidence B · P1
  237. MWITA-ASVC-2026-016Grab unveiled its LLM-based Merchant AI Assistant inside GrabMerchant in April 2025.Evidence B · P1
  238. MWITA-ASVC-2026-017Grab says the Merchant AI Assistant can proactively suggest sales improvements and help create ad campaigns or update menus.Evidence B · P1
  239. MWITA-ASVC-2026-018Grab reported that its pilot voice-reporting tool received more than 16,000 road reports per day from over 900,000 drivers.Evidence B · P1
  240. MWITA-ASVC-2026-019Grab's voice-reporting feature accepts multilingual verbal reports for traffic, roadworks and flooding.Evidence B · P1
  241. MWITA-ASVC-2026-020Grab built the Merchant AI Assistant with OpenAI and Anthropic models, while the Driver Companion used OpenAI models.Evidence B · P1
  242. MWITA-ASVC-2026-023Nubank reported that AI agents handled more than 60% of customer-support conversations in Brazil in Q2 2026.Evidence B · P1
  243. MWITA-ASVC-2026-024Nubank characterized the Q2 2026 AI-handled support conversations as operating at or above human parity.Evidence B · P1
  244. MWITA-ASVC-2026-025Nubank said its NuFormer model powers customer-service and growth decisions across the company.Evidence B · P1
  245. MWITA-ASVC-2026-026Nubank's Precog system predicts why a customer is contacting support and prepares the human agent with relevant information.Evidence B · P1
  246. MWITA-ASVC-2026-030Nubank reported in late 2024 that AI resolved 40% of demands arriving through chat.Evidence B · P1
  247. MWITA-ASVC-2026-031Nubank disclosed a copilot that assists its human customer-service agents.Evidence B · P1
  248. MWITA-ASVC-2026-036Mercado Libre said model routing, evaluation, observability and cache improvement reduced its cost per token as AI scaled.Evidence B · P1
  249. MWITA-ASVC-2026-038Mercado Libre's 2024 Form 10-K identifies nondeterministic generative-AI outputs as a risk to product reliability and trustworthiness.Evidence B · P1
  250. MWITA-ASVC-2026-039Mercado Libre's 2024 Form 10-K says AI failures could cause customer dissatisfaction, reputational harm or legal and regulatory liabilities.Evidence B · P1
  251. MWITA-ASVC-2026-040Mercado Libre's 2024 Form 10-K identifies data sourcing, bias, security and privacy as AI-adoption risks.Evidence B · P1
  252. MWITA-ASVC-2026-041NAVER's Shopping AI Agent moved from a February 2026 beta focused on discovery and summaries to proactively initiating conversations in June 2026.Evidence B · P1
  253. MWITA-ASVC-2026-042NAVER says its Shopping AI Agent analyzes clicks, saved items, cart additions and trends to suggest product-discovery directions.Evidence B · P1
  254. MWITA-ASVC-2026-044NAVER labels the Shopping AI Agent action-oriented, but the cited June 2026 release documents suggestions rather than completed purchases.Evidence B · P1
  255. MWITA-ASVC-2026-045NAVER's CLOVA LiveChat analyzes Smart Store inquiries and recommends seller-specific FAQs.Evidence B · P1
  256. MWITA-ASVC-2026-046CLOVA LiveChat can automatically answer configured FAQs through NAVER TalkTalk around the clock after a seller supplies answers.Evidence B · P1
  257. MWITA-ASVC-2026-048CLOVA LiveChat expanded FAQ recommendations from store-level topics to about 4,000 product categories.Evidence B · P1
  258. MWITA-ASVC-2026-049Rakuten said its generative-AI customer-service agent Raptor was being deployed across its ecosystem.Evidence B · P1
  259. MWITA-ASVC-2026-051Rakuten says Raptor combines AI chat, AI voice and AI-augmented human service.Evidence B · P1
  260. MWITA-ASVC-2026-052stc launched a voice-of-customer transcription co-worker using its speech-recognition and large-language-model capabilities.Evidence B · P1
  261. MWITA-ASVC-2026-053stc says its call-transcription co-worker extracts signals for churn, CSAT and fraud analysis.Evidence B · P1
  262. MWITA-ASVC-2026-054stc reported more than 5,000 stc Brain conversations across over ten sectors and more than 200 users during 2024.Evidence B · P1
  263. MWITA-ASVC-2026-058etisalat by e& soft-launched a voice-capable Consumer Assistant at GITEX in October 2023.Evidence B · P1
  264. MWITA-ASVC-2026-059e& said its Consumer Assistant can join voice calls and take actions such as scheduling appointments.Evidence B · P1
  265. MWITA-SR-2026-001GPT-3 conditioned on thousands of sociodemographic backstories from real US survey participants produced response patterns the authors described as algorithmic fidelity.Evidence A · P0
  266. MWITA-SR-2026-005The benchmark found exaggerated ideological out-group antipathy in some persona-conditioned outputs.Evidence A · P0
  267. MWITA-SR-2026-006Turing Experiments replicated the established direction in the Ultimatum Game, Garden Path and Milgram studies for tested recent models.Evidence A · P0
  268. MWITA-SR-2026-007The Wisdom of Crowds Turing Experiment exposed a hyper-accuracy distortion in some models including GPT-4.Evidence A · P0
  269. MWITA-SR-2026-008Agents built from qualitative interviews with 1,052 individuals were evaluated against those same individuals across surveys, traits and experiments.Evidence B · P0
  270. MWITA-SR-2026-012Fine-tuning on response distributions outperformed tested prompting and zero-shot baselines on seen and unseen benchmark splits.Evidence A · P0
  271. MWITA-SR-2026-016Default model responses were closer to opinions in the United States and some European and South American populations than to all populations equally.Evidence B · P0
  272. MWITA-SR-2026-017Country prompting shifted outputs toward the prompted population but could also produce harmful cultural stereotypes.Evidence B · P0
  273. MWITA-SR-2026-018Translating questions into a target language did not necessarily make model responses most similar to speakers of that language.Evidence B · P0
  274. MWITA-SR-2026-019Across the benchmark's four models, no LLM beat the strongest non-LLM baseline for individual-level prediction.Evidence B · P0
  275. MWITA-SR-2026-024The framework separates agent embodiment, scenario exposure and sentiment generation with rationales.Evidence B · P0
  276. MWITA-SR-2026-025A China–US comparison evaluated DeepSeek, Qwen, GPT and Llama families on ANES and China's Zuobiao public-opinion data.Evidence B · P0
  277. MWITA-SR-2026-026A preliminary experiment tested whether LLMs reproduce personality differences between US and South Korean targets.Evidence B · P0
  278. MWITA-SR-2026-027Google DeepMind's Concordia implements entities, modular components, simulation engines and Game Master-mediated environments for generative social simulation.Evidence B · P0
  279. MWITA-SR-2026-028Concordia includes questionnaire, open-ended interview, marketplace and psychology-experiment components.Evidence B · P0
  280. MWITA-SR-2026-029Synthetic Users documents audience definitions spanning demographics, psychographics, behaviors and context, with AI-generated or extended personas.Evidence C · P0
  281. MWITA-SR-2026-030The platform documents structured or semi-structured synthetic interviews, dynamic follow-ups and full transcripts.Evidence C · P0
  282. MWITA-SR-2026-032Qualtrics stated in February 2026 that its generally available synthetic panel covered the US general population in English.Evidence C · P0
  283. MWITA-SR-2026-037The documented LINK AI pipeline decomposes media and extracts up to 20,000 features before predicting creative-effectiveness metrics.Evidence C · P0
  284. MWITA-SR-2026-039Ipsos PersonaBots are presented as a way to converse with existing research segments about concepts, messages and competitive scenarios.Evidence C · P0
  285. MWITA-SR-2026-040NIST's Generative AI Profile organizes suggested risk actions across Govern, Map, Measure and Manage functions.Evidence A · P0
  286. MWITA-SR-2026-048ASEAN's expanded guide identifies mistakes and anthropomorphism, factual inaccuracy, impersonation, IP infringement, privacy/confidentiality and embedded bias as generative-AI risks.Evidence A · P0
  287. MWITA-SR-2026-055The Chinese measures require steps to prevent discrimination across ethnicity, belief, nationality, region, sex, age, occupation and health.Evidence A · P0
  288. MWITA-SR-2026-056The Chinese measures call for transparency and improved accuracy and reliability of generated content.Evidence A · P0
  289. MWITA-SR-2026-057Japan's METI and MIC consolidated earlier R&D, utilization and implementation-governance guidance into AI Guidelines for Business Ver 1.0.Evidence A · P0
  290. MWITA-SR-2026-068Temporal validation is required because a model calibrated to a past survey wave can drift from current preferences even if its original benchmark passed.Evidence C · P0
  291. MWITA-SR-2026-071Synthetic interview transcripts and quotes must be visibly labeled because anthropomorphic language can make generated testimony look like consented lived experience.Evidence C · P0
  292. MWITA-NA-2026-009Japan's Moonshot program is developing a cybernetic-avatar concept that combines internal brain activity, brain-surface information, interaction data and AI, with ethical issues stated as part of the program.Evidence A · P1
  293. MWITA-SPC-2026-005AIST announced in January 2025 that it had begun R&D on robot foundation models for manufacturing, logistics and retail as well as household and office service settings.Evidence A · P0
  294. MWITA-SPC-2026-006AIST stated that the program seeks collaboration with robot users, manufacturers and system integrators around tasks that remain difficult to automate with conventional teaching or programming.Evidence A · P0
  295. MWITA-SPC-2026-007AIST built a laboratory system combining AI-based human detection with a user intervention interface for a robot arm and a transfer lift.Evidence A · P0
  296. MWITA-SPC-2026-008The AIST project explicitly treats AI misjudgment and uncertainty as conditions that require a separate high-reliability path for users to suppress, interrupt or reverse robot motion.Evidence A · P0
  297. MWITA-SPC-2026-012Samsung specifies Galaxy XR interaction through voice, gaze and gesture and lists two passthrough cameras, six world-tracking cameras, four eye-tracking cameras, five IMUs and a depth sensor.Evidence B · P0
  298. MWITA-SPC-2026-013KAIST researchers reported a glove prototype that delivers static pressure, dynamic vibration and variable-stiffness feedback across finger joints.Evidence A · P0
  299. MWITA-SPC-2026-014The reported glove was used for real-time VR grasping and haptic-assisted robot teleoperation intended to convey robot contact sensations to the operator.Evidence A · P0
  300. MWITA-SPC-2026-022The Dubai housing platform describes an AI assistant for finding housing models and access through VR headsets, computers, tablets and mobile phones.Evidence A · P0
  301. MWITA-SPC-2026-025India's MeitY 2025–2026 annual report says its Emerging Technologies Division supports policy and strategy work in areas including AR/VR, robotics, computer vision, IoT and drones.Evidence A · P0
  302. MWITA-SPC-2026-035The collaborative-shopping experiment used functional near-infrared spectroscopy hyperscanning alongside shopping-satisfaction measures.Evidence A · P0
  303. MWITA-SPC-2026-041NVIDIA describes GR00T N1 as a generalist vision-language-action model trained on egocentric human videos, real and simulated robot trajectories and synthetic data.Evidence B · P0
  304. MWITA-SPC-2026-042NVIDIA released GR00T N1 model weights under the NVIDIA Open Model License.Evidence B · P0
  305. MWITA-SPC-2026-044DeepMind describes a layered robotics-safety approach in which embodied reasoning can connect to embodiment-specific low-level controllers for collision avoidance, contact-force limits and dynamic stability.Evidence B · P0
  306. MWITA-SPC-2026-045Unitree markets the G1 as a commercial humanoid platform and publishes manufacturer specifications and pricing language on its product page.Evidence B · P1
  307. MWITA-SPC-2026-046Unitree's product page warns users to keep a sufficient safe distance from humanoid robots and acknowledges limitations of the current industry.Evidence B · P0
  308. MWITA-SPC-2026-047AIST's Artificial Intelligence Research Center describes 3D object, movement and environment recognition as foundational to situational understanding in manufacturing, service workplaces and homes.Evidence A · P0
  309. MWITA-SPC-2026-048AIRC states that it built a cloud-based virtual-reality environment and is accumulating interaction data between humans and robots for future robot behavior planning research.Evidence A · P0
  310. MWITA-PE-2026-001Adding GPT-extracted listing-text features improved an Oslo XGBoost AVM from 10.53% to 7.97% RMSE and from 6.36% to 5.39% MAPE.Evidence A · P1
  311. MWITA-SEC-2026-001InjecAgent found ReAct-prompted GPT-4 vulnerable to indirect prompt injection in 24% of tests.Evidence A · P1
  312. MWITA-SEC-2026-002AgentDojo measured a strong security-utility trade-off: for GPT-4o, targeted attack success fell from 57.69% without defense to 7.95% with a detector or 6.84% with tool filtering, with different benign utility costs.Evidence A · P1
  313. MWITA-SEC-2026-007OS-Harm documented unsafe execution traces across all tested frontier computer-use model families.Evidence A · P1
  314. MWITA-SEC-2026-008A month-long public red-team competition recorded more than 60,000 policy violations from 1.8 million attempts across 22 frontier LLMs.Evidence A · P1
  315. MWITA-CO-2026-008Three laboratory experiments found that advertising disclosure attached to the product explicitly sought in a generative-AI answer reduced trust in both system and product, whereas disclosure on a later non-target recommendation reduced recommendation trust without a significant system-trust penalty.Evidence B · P1
  316. MWITA-CO-2026-012In a randomized online-travel-planning field experiment, enthusiastic GenAI elicited longer prompts, while enthusiastic and neutral variants were more likely than an unconstrained-tone control to generate web-service subscription purchases.Evidence B · P1
  317. MWITA-AGC-2026-001Access to an AI livestream-shopping assistant increased sales 3.00% and reduced product-return rates 12.55%.Evidence A · P1
  318. MWITA-AGC-2026-002In a custom Alexa shopping experiment, evaluating fewer alternatives was associated with greater decision satisfaction, trust and willingness to delegate shopping decisions.Evidence A · P1
  319. MWITA-AGC-2026-003Scarcity cues increased experimental selection of a high-autonomy shopping assistant, although ordinary conditions favored lower autonomy.Evidence A · P1
  320. MWITA-AGC-2026-004Across five experiments, high AI autonomy generally reduced shopping-assistant adoption, while explicit scarcity modestly increased high-autonomy choice.Evidence A · P1
  321. MWITA-AGC-2026-005Controlled shopping-agent choices were model-dependent and causally sensitive to listing position, endorsements, price and description wording.Evidence B · P1
  322. MWITA-AGC-2026-006In simulated two-sided agent markets, response speed created a much larger selection advantage than service quality.Evidence B · P1
  323. MWITA-AGC-2026-007The best evaluated language agent still failed more than half of complex end-to-end shopping instructions in ShoppingBench.Evidence B · P1
  324. MWITA-AGC-2026-008On complex live-web shopping queries, evaluated web and deep-research agents achieved low holistic success, especially when filters and sorting had to be satisfied together.Evidence B · P1
  325. MWITA-AGC-2026-009Population-scale measurement found that raw x402 settlement count substantially overstated independently evidenced commerce adoption.Evidence B · P1
  326. MWITA-AGC-2026-010After excluding identified wash and test activity, Visa and Artemis reported measurable but differently scaled x402 and MPP machine-payment use.Evidence C · P1
  327. MWITA-AGC-2026-011A card-network agentic-payment path executed one disclosed live production purchase with explicit human approval and agent identification.Evidence C · P1
  328. MWITA-AGC-2026-012Surveyed monthly online shoppers reported substantial willingness to delegate shopping stages, while data control and purchase accuracy remained major constraints.Evidence C · P1
  329. MWITA-AIAO-2026-001A diffusion-based auto-bidder improved reported ROI by 3.36% in a one-week Alibaba online A/B test, alongside 2.09% more purchases and 2.81% higher GMV.Evidence A · P1
  330. MWITA-AIAO-2026-002A hierarchical on-policy bidding agent raised conversion value 8.1% and ROI 3.3% versus a production system in a seven-day test, while campaigns with fewer than three daily conversions remained a documented weakness.Evidence B · P1
  331. MWITA-AIAO-2026-003A CTR-optimized advertising-image pipeline reported 7.4% higher click-through rate than a pretrained multimodal model in adaptive online serving.Evidence B · P1
  332. MWITA-AIAO-2026-004Embedded undisclosed product placement in an experimental LLM chatbot was difficult to recognize and shifted product evaluations while reducing benchmark answer quality by no more than three percent.Evidence A · P1
  333. MWITA-AIAO-2026-005LLM-generated ads did not significantly outperform human ads on personality personalization but were preferred more often on general persuasion principles, with a large penalty when AI origin was correctly detected.Evidence B · P1
  334. MWITA-AIAO-2026-008In green advertising, disclosure that an image was AI-generated produced more favorable brand attitudes and purchase intentions than disclosure that the copy was AI-generated.Evidence A · P1
  335. MWITA-AIAO-2026-009Disclosing that both advertising copy and image were AI-produced reduced perceived advertising value and purchase intention, with credibility mediating the negative response.Evidence A · P1
  336. MWITA-AIAO-2026-010Consumer trust differed across five concealed in-app ad-generation modalities spanning AI, human, and user-generated advertisements and memes.Evidence A · P1
  337. MWITA-AIAO-2026-011Adding explanations or beneficiary framing to an AI disclosure did not repair advertising response in a preregistered replication; explanatory wording slightly worsened ad and brand attitudes while engagement intention was statistically equivalent.Evidence A · P1
  338. MWITA-AIAO-2026-012Product-relevant AIGC advertisements directed substantially more gaze toward the product region than divergent AIGC advertisements among Chinese university viewers.Evidence A · P1
  339. MWITA-SSCO-2026-001One generative image model outperformed professional stock imagery on clicks in one display-ad setting, while the other tested models did not.Evidence A · P1
  340. MWITA-SSCO-2026-006In controlled XR shopping, VR increased perceived enjoyment, while AR reduced perceived ease of use and did not significantly increase enjoyment or usefulness.Evidence A · P1
  341. MWITA-SSCO-2026-010A Chinese sign-language-avatar benchmark showed measurable improvement but limited absolute quality: only four of ten submitted systems achieved good readability.Evidence B · P1
  342. MWITA-HCO-2026-001In a preregistered cervical-screening messaging experiment, AI-targeted and AI-tailored narratives were perceived as lower quality and less believable and did not increase screening motivation beyond nonnarrative or generic narratives.Evidence A · P1
  343. MWITA-HCO-2026-002A clinician-validated AI avatar increased immediate semaglutide self-injection knowledge relative to nurse education but produced lower satisfaction and trust, with no significant self-efficacy difference.Evidence A · P1
  344. MWITA-HCO-2026-003Brief GPT-4o conversations increased immediate parental HPV-vaccination intent versus no message but did not outperform country-matched public-health materials, did not persist at 45 days, and did not increase self-reported vaccination uptake.Evidence A · P1
  345. MWITA-HCO-2026-004Labeling otherwise identical physicians as using AI for diagnosis substantially reduced US respondents’ trust and stated intention to seek care, with the lowest ratings under extensive AI use.Evidence A · P1
  346. MWITA-HCO-2026-005ChatGPT-assisted counseling before chemotherapy reduced measured anxiety and depression relative to clinician-led education in a two-center randomized trial, while 11.5% of reviewed AI responses were not rated appropriate.Evidence A · P1
  347. MWITA-HCO-2026-006A five-week motivational-interviewing AI vaccine assistant improved vaccine readiness and health literacy among vaccine-hesitant Hong Kong adults but did not produce a significant group-by-time reduction in vaccine hesitancy.Evidence A · P1
  348. MWITA-HCO-2026-007Adding a clinician-validated cataract chatbot to brochures did not improve knowledge gain, anxiety, or satisfaction, despite high usability among users and a 52% non-engagement rate in the chatbot arm.Evidence A · P1
  349. MWITA-HCO-2026-008Plausible AI misinformation cut novice medical students’ diagnostic accuracy from 21.0% to 9.2%, whereas correct AI explanations did not significantly improve accuracy over no explanation.Evidence A · P1
  350. MWITA-HCO-2026-009A minimal prompting boost improved the evidence-based quality of screening information produced by three LLMs for lay users, but responses remained below required health-communication standards.Evidence A · P1
  351. MWITA-HCO-2026-010Co-designed animations and images increased knowledge of ChatGPT limitations and selectively reduced intention to use it for higher-risk health questions, while leaving lower-risk use intentions similar to control.Evidence A · P1
  352. MWITA-HCO-2026-011Identical digital medical advice was judged less reliable and empathetic and elicited lower willingness to follow when labeled as AI-generated or human-plus-AI than when labeled human-generated.Evidence A · P1
  353. MWITA-HCO-2026-012In a randomized simulated-consultation crossover study, AMIE exceeded primary-care physicians on diagnostic accuracy and most rated communication dimensions, but the unfamiliar text-only OSCE setting prevents inference to clinical deployment.Evidence A · P1
  354. MWITA-EBRO-2026-004Telling directly sourced candidates that an AI application had been used to identify them significantly reduced their interest in a real public-sector job offer, even though an exploratory deployment analysis indicated that the tool's person-job-fit score could predict leaders' fit assessments.Evidence A · P1
  355. MWITA-EBRO-2026-005Combining premium vacancy promotion with applicant identity verification increased portal-sourced hiring by 8.2 percentage points, or 67.8% relative to control, and increased the probability that a vacancy was filled overall by about 11%; either service alone did not significantly increase hiring.Evidence A · P1
  356. MWITA-EBRO-2026-010Referred online workers were more likely to accept an offer, submit work on time and continue than observably similar nonreferred workers, including on work performed under a different firm name without their referrers, indicating that referrals carried otherwise hidden information about worker quality and persistence.Evidence A · P1
  357. MWITA-ELO-2026-010Adding problem-solving content to four monthly emails reduced rather than increased access to an audit-and-feedback tool by 12.9%, while anticipated-regret and material-incentive content did not significantly increase access; fewer than half opened and fewer than one in ten clicked.Evidence A · P1
  358. MWITA-ELO-2026-011Among people who had enrolled but not activated an e-cohort, three reminders mixing standard email, SMS and institutional email increased 30-day questionnaire completion 27% relative to a single-mode approach; SMS accelerated response and institutional/formal email appeared more effective than standard wording.Evidence A · P1
  359. MWITA-CUX-2026-006Among 2,605 recruitment-site visitors, one infographic landing page reduced interest-form completion from 16.8% to 12.8% (OR 0.73, p=.019), while a second infographic was null; neither changed eventual study randomization.Evidence A · P1
  360. MWITA-BCO-2026-002Across two Brazilian B2B field experiments, a WhatsApp-based chatbot generated significantly more total and qualified leads than a static landing page, establishing a top-of-funnel effect without evidence here for meetings, opportunities or revenue.Evidence A · P1
  361. MWITA-BCO-2026-008A two-stage cluster-plus-score model increased average observed lead conversion from 6.2% under the traditional method to 8.9% without using later behavioral features, a 43.5% relative uplift at lead creation.Evidence B · P1
  362. MWITA-BCO-2026-011A model-based account call-frequency planning aid increased six-month sales by 8.1% versus matched manual planning among United Airlines sales representatives.Evidence A · P1
  363. MWITA-NCO-2026-003Mobile VR was highly acceptable but did not produce statistically distinguishable six-month neonatal-resuscitation pass rates versus video or a digital guide.Evidence A · P1
  364. MWITA-NCO-2026-004Two weeks of immersive virtual-movement therapy reduced neuropathic pain more than a headset-based 2D-film control among spinal-cord-injury patients who completed the trial.Evidence A · P1
  365. MWITA-NCO-2026-005VR goggles offered no pain benefit over the same cartoon on a conventional screen; secondary pointwise adjusted heart-rate estimates favored the screen, while the omnibus longitudinal group interaction and raw change comparisons were null.Evidence A · P1
  366. MWITA-NCO-2026-007Immersive CPR training improved practical-skill retention at one month but not immediate theoretical knowledge, while imposing greater mental demand.Evidence A · P1
  367. MWITA-NCO-2026-008Children assigned by serial-number parity to passive VR reported less pain and had lower pulse rates than conventional behavior management, while oxygen saturation was unchanged.Evidence B · P1