As of snapshot 27745 under ledger-default@0.3.0. Counts by assessment state, descriptive only: no topic is scored.
2 Contradicted3 Insufficient evidence4 Leans contradicted11 Leans supported13 Not assessed67 Supported
| Claim | Type | Assessment |
|---|
| The AGENT framework used high-throughput experiments and Bayesian optimization to identify thermostable solid-state mRNA-lipid nanoparticle vaccine formulations in six iterations over one month. | OBSERVATIONAL | Supported |
| SAIL succeeded in five of six physical block-placement trials using a LeRobot SO-101 arm with a budget of 15 candidate trajectories. | QUANTITATIVE | Supported |
| In simulation, SAIL increased average robot trajectory success from 25% with one generated trajectory to 73% when testing and refining up to 45 candidates, without updating model weights. | QUANTITATIVE | Supported |
| Across 26 data- and discovery-focused papers, Paper2Agent’s resource layer achieved 89.0 ± 3.1% accuracy on 100 synthesis questions and was reported as cheaper and faster than a browser-use baseline. | QUANTITATIVE | Leans supported |
| Across 42 execution-based tasks from 10 non-biology computational papers, Paper2Agent achieved 98.1 ± 0.8% accuracy across five independent runs. | QUANTITATIVE | Supported |
| On 300 tutorial-derived benchmark questions, Paper2Agent with Sonnet 4 achieved 91.2 ± 1.6% accuracy, compared with 80.3 ± 2.3% for direct-repository Claude Code using Sonnet 4. | QUANTITATIVE | Leans supported |
| In a corpus of 100 computational-biology papers, Paper2Agent successfully agentified 74 papers and produced 599 proposed tools, of which 593 passed automated validation. | QUANTITATIVE | Leans supported |
| Paper2Agent validates each generated tool against reference-code results and figures using example datasets, then locks validated tools to support reproducibility. | OBSERVATIONAL | Leans supported |
| Paper2Agent is an automated framework that converts research papers and their associated code, data and workflows into interactive AI agents exposed through MCP servers. | OBSERVATIONAL | Leans supported |
| Among Texas graduates who found work, first-year earnings for more AI-exposed majors fell approximately 5% from 2021 to 2024 relative to less-exposed majors. | QUANTITATIVE | Supported |
| Among recent graduates from Texas four-year universities, a 10-percentage-point higher share of automatable tasks was associated with a 1.7-percentage-point relative decline in employment rates after late 2022. | QUANTITATIVE | Supported |
| The study trained GPN-Star for five model organisms beyond humans, demonstrating robustness and generalizability across those species. | OBSERVATIONAL | Supported |
| In the study, GPN-Star achieved state-of-the-art performance across coding and non-coding human variant-effect prediction tasks. | OBSERVATIONAL | Supported |
| GPN-Star is a phylogeny-aware genomic language model that uses whole-genome alignments and species trees to model evolutionary relationships. | DEFINITIONAL | Supported |
| The study reported that Harris County magistrate judges' misdemeanor bail decisions could generally be captured by small, interpretable formulas. | OBSERVATIONAL | Supported |
| In the study’s 14,644-patient RW2 cohort, EAGLE-Plus reduced false positives by 72.7% while maintaining comparable sensitivity. | QUANTITATIVE | Supported |
| In the study’s eight-center external CT test cohorts, EAGLE achieved 98.5% specificity, 90.0% cancer sensitivity and 52.5% precancerous-lesion sensitivity. | QUANTITATIVE | Supported |
| The EAGLE model was trained on 6,813 patients and validated across 12 centers in three countries involving 80,612 patients. | QUANTITATIVE | Supported |
| GenFocal is an AI framework that generates fine-scale weather from coarse climate projections without requiring temporally paired simulated and observed events during training, and it synthesizes long-lived hazards such as heatwaves and tropical cyclones. | OBSERVATIONAL | Insufficient evidence |
| In a triple-masked randomized trial of 75 adults with subacromial pain syndrome, clinician-curated ChatGPT-4o and DeepSeek-R1 exercise programs added to standard physiotherapy did not improve six-week shoulder disability more than conventional programming, although an exploratory resting-pain difference favored DeepSeek-R1. | QUANTITATIVE | Supported |
| In that study, GPT-4-derived predictions were strongly correlated with observed treatment effects at a level similar to pooled human forecasts, but systematically overestimated effect sizes. | COMPARATIVE | Supported |
| A 2026 study evaluated LLM predictions using an archive of 70 preregistered US survey experiments containing 469 experimental effects and 119,330 participants. | QUANTITATIVE | Supported |
| An AI-guided Bayesian-optimization framework identified solid-state mRNA-lipid nanoparticle formulations that retained 100% bioactivity after more than two months at 37 °C, while inducing non-inferior antigen-specific responses in rodents and non-human primates versus fresh soluble vaccines. | QUANTITATIVE | Supported |
| In a catalyst-design application, MatBrain generated 30,000 candidate structures and identified 38 promising materials within 48 hours. | QUANTITATIVE | Supported |
| MatBrain is a dual-model crystal-materials research agent composed of a 30-billion-parameter analytical model and a 14-billion-parameter executive model. | DEFINITIONAL | Supported |
| In the same randomized evaluation, mean chart-review time was similar for human-plus-AI and human-alone prescreening (37.4 versus 37.8 minutes). | QUANTITATIVE | Supported |
| In a randomized evaluation of 355 oncology charts, human-plus-AI prescreening had higher chart-level accuracy than human-alone prescreening (76.1% versus 71.5%). | QUANTITATIVE | Supported |
| In a randomized controlled trial of 249 physicians in Indonesia, Kenya, and the Netherlands, GPT-4o access improved clinical-performance scores relative to control by 18 percentage points in Kenya, 10.7 in Indonesia, and 7.2 in the Netherlands under controlled conditions; harms were not assessed. | QUANTITATIVE | Supported |
| In a 12-month randomized HealthMPowerment trial among young Black and Latinx men who have sex with men and transgender women, the smartphone intervention did not improve routine HIV testing; apparent viral-suppression benefits were no longer significant after adjustment for baseline suppression, age and income. | QUANTITATIVE | Supported |
| The open-source GOLEM humanoid architecture demonstrated EV-battery fastener grasping on a real Hyundai Ioniq 5 pack, with success declining from 97% while tethered to 87% free-standing and 37% under navigation-induced pose variance. | QUANTITATIVE | Supported |
| In the study’s modeled channels, educational capacity partly mitigated the AI–inequality relationship, with heterogeneous roles for basic, higher, and vocational education. | OBSERVATIONAL | Supported |
| Using Chinese provincial panel data from 2010 to 2024, the study found that AI diffusion was positively associated with both the Theil index and the urban–rural income gap in fixed-effects and dynamic panel system-GMM estimates. | QUANTITATIVE | Leans supported |
| In a pragmatic cluster-randomized trial across 16 Kenyan primary-care facilities, LLM-assisted clinical officers did not significantly change 14-day treatment-failure rates versus control: 2.2% versus 2.0% (adjusted odds ratio 0.77, 95% CI 0.55–1.08, P=0.13). | QUANTITATIVE | Supported |
| In the reported study, higher media literacy was associated with improved detection of political deepfakes and reduced liking and sharing intentions. | COMPARATIVE | Supported |
| In a study of 1,124 participants from the United Kingdom, Slovenia, and Italy, high-plausibility political deepfakes were less likely to be detected and received more positive evaluations than low-plausibility deepfakes. | COMPARATIVE | Supported |
| In the study's MIRA-v2 evaluation, diagnostic behavioral consistency had AUC 0.860 for correctness discrimination and 0.875 under stress testing; a 0.90 threshold retained 49.4% of cases at 98.9% diagnostic accuracy. | QUANTITATIVE | Supported |
| In two MIMIC-IV-derived benchmarks, the evaluated on-premise clinical AI agent achieved 90.04% accuracy on a seven-disease task and 83.8% accuracy on a four-disease task. | QUANTITATIVE | Supported |
| Under combined semantics-preserving lexical perturbations, CausalVul's F1 dropped 5.1%, compared with 12.1–22.3% for the baselines evaluated in the study. | COMPARATIVE | Supported |
| Under project-level group-aware evaluation on the Devign vulnerability-detection benchmark, CausalVul achieved 76.4% F1, 18.0 points above its fine-tuned CodeBERT backbone and 4.8 points above the UniXcoder baseline. | QUANTITATIVE | Leans supported |
| In the same experiment, treated employers who posted a job were 15% less likely to make a hire than control employers who posted a job. | QUANTITATIVE | Supported |
| In a large randomized field experiment on an online labor market, offering employers AI-written first drafts made them 19% more likely to post a job, but did not increase the overall probability that a match was formed. | QUANTITATIVE | Supported |
| The study concluded that teacher expertise and critical engagement remain essential for ensuring the accuracy, depth, and pedagogical value of GenAI-generated science lessons. | INTERPRETIVE | Supported |
| In the study, ChatGPT-generated lesson-plan outputs typically lacked specificity and contained inaccuracies, while iterative prompting did not substantively elevate systems-thinking complexity. | COMPARATIVE | Supported |
| In the study, ChatGPT-generated high-school socioscientific-issue lesson plans provided opportunities mainly at basic levels of systems thinking. | OBSERVATIONAL | Supported |
| An independent August 2026 report said OpenAI's Astra solved ten previously open problems in mathematics and theoretical computer science, and that formal Lean proofs were published for machine checking. | HISTORICAL | Supported |
| The FourCastNet paper reported that its neural-operator weather model generated a week-long forecast in under two seconds and was about 45,000 times faster than traditional numerical weather prediction on a node-hour basis. | QUANTITATIVE | Supported |
| A September 2026 guest post said the group used Neural Operators to train a high-resolution AI weather model more than five years earlier and reported it was tens of thousands of times faster. | TEXTUAL | Supported |
| A September 2026 preprint reported evidence of a stable finite-time singularity in the three-dimensional Euler equations on R³ using a physics-informed neural network and self-similar ansatz. | TEXTUAL | Supported |
| The Decoder reported that GPT-6 Astra used one-third as many compute steps as GPT-5.6 Sol and one-fifth as many as Claude Opus 5 on coding tasks. | COMPARATIVE | Supported |
| The Decoder reported that maximum-reasoning ARC-AGI-3 testing for GPT-6 Astra cost about $26,098 versus $49,791 with no reasoning, while the score rose from 35.2% to 62.7%. | QUANTITATIVE | Supported |
| The Decoder reported GPT-6 Astra at 62.7% on ARC-AGI-3, compared with 7.78% for GPT-5.6 Sol and 30.16% for Claude Opus 5. | COMPARATIVE | Supported |
| The Decoder reported that GPT-6 Astra reached 62.7% on ARC-AGI-3 at a test cost of roughly $26,000. | QUANTITATIVE | Supported |
| In OpenAI’s reported DeepSWE v1.1 results, GPT-6 Astra scored 74.1% versus 72.7% for GPT-5.6 Sol. | QUANTITATIVE | Leans supported |
| OpenAI reported that GPT-6 Astra’s top configuration cut estimated API cost per DeepSWE v1.1 task by about 57% versus GPT-5.6 Sol. | QUANTITATIVE | Supported |
| Agricultural Bank of China and American Express launched a Kimi AI-focused credit card in mainland China on July 10, 2026. | HISTORICAL | Supported |
| Ping An Bank cardholders can redeem AI token packages and memberships with points earned from spending. | OBSERVATIONAL | Supported |
| Demis Hassabis shared half of the 2024 Nobel Prize in Chemistry for developing AlphaFold2, an AI system for protein structure prediction. | HISTORICAL | Supported |
| Dario Amodei has described an end state in which a single company such as Anthropic constitutes all private enterprise. | ATTRIBUTED_BELIEF | Leans contradicted |
| Sam Altman posted on social media within 24 hours of this recording expressing surprise that OpenAI had solved Navier-Stokes. | OBSERVATIONAL | Supported |
| Sam Altman said that in the future an AI should be running OpenAI. | ATTRIBUTED_BELIEF | Supported |
| A frontier AI lab is already running its own market-beating hedge fund strategy on an internal model. | OBSERVATIONAL | Contradicted |
| Frontier labs walling off their most capable internal models pushes toward larger firms, the opposite of the Coasean prediction. | INTERPRETIVE | Leans supported |
| Legal entities will eventually be able to be owned by AI agents in Argentina and elsewhere. | FORECAST | Not assessed |
| AI does not merely automate the firm but attacks the economic reason firms exist, turning the firm into a protocol coordinating a community of agents and humans. | INTERPRETIVE | Supported |
| MIT and Harvard researchers have published work on what happens when AI agents make transactions nearly free, which they call the Coasean singularity. | OBSERVATIONAL | Supported |
| There are no stationary competitive moats during a singularity, only temporary dynamical ones that must be sequenced. | INTERPRETIVE | Leans contradicted |
| Within a few years humanoid robots will make HVAC engineering as automatable as spreadsheet management and accounting. | FORECAST | Not assessed |
| Every profession as currently constituted will be automated, and only the sequencing differs. | FORECAST | Not assessed |
| Emad Mostaque put forward a Champion proposal in which every child owns equity in the robots at a nominal pre-money valuation. | OBSERVATIONAL | Insufficient evidence |
| AI will not create jobs fast enough to offset the ones it displaces, though it can create abundance. | FORECAST | Not assessed |
| AI models have reached a capability level at which an entire digital workforce could be replicated within a year. | FORECAST | Not assessed |
| Within the million AI-related jobs, co-pilot roles roughly double a worker's efficiency while a separate subset manages tens to hundreds of agents as if they were employees. | INTERPRETIVE | Supported |
| Eric Schmidt said that AI-exposed jobs are growing faster and paying better. | ATTRIBUTED_BELIEF | Supported |
| Occupations expected to face AI disruption, including paralegals and market research analysts, have continued to grow. | OBSERVATIONAL | Contradicted |
| LinkedIn estimated that 640,000 AI-specific jobs were created between 2023 and 2025. | QUANTITATIVE | Supported |
| Roughly one million professional positions in the United States are now classified as AI jobs. | QUANTITATIVE | Supported |
| The AI economy is now large and self-contained enough that it will mostly leave the legacy economy alone rather than destroying white collar jobs. | FORECAST | Not assessed |
| Anthropic and OpenAI going public will create a wave of new centimillionaires and billionaires who become the largest source of seed and Series A capital. | FORECAST | Not assessed |
| Abundance in healthcare, utility pricing, food, shelter and education is downstream of giving everyone universal basic computing. | INTERPRETIVE | Leans supported |
| South Korea has announced universal AI access for all its people through a consortium. | OBSERVATIONAL | Supported |
| China is the bellwether for universal basic compute, and state-issued redistribution of AI tokens will eventually extend to all of humanity. | FORECAST | Not assessed |
| AI will never become too cheap to meter, because use cases expand as fast as unit costs fall. | FORECAST | Not assessed |
| Intelligence is becoming infrastructure, a foundational layer beneath other activity. | INTERPRETIVE | Supported |
| The average Chinese person together with their AI will be smarter than the average American together with theirs. | FORECAST | Not assessed |
| A rampant anti-AI counterculture will emerge in the United States, probably led by Bernie Sanders. | FORECAST | Not assessed |
| Restaurants in China hand customers compute credits after a meal. | OBSERVATIONAL | Supported |
| China Telecom sells access to 142 AI models in the manner of a mobile data plan. | OBSERVATIONAL | Supported |
| Banks in China give AI tokens as credit card rewards. | OBSERVATIONAL | Supported |
| The AI investment cycle is roughly in its third inning of nine. | INTERPRETIVE | Supported |
| 26,000 teams entered the Build with Gemini XPRIZE. | QUANTITATIVE | Insufficient evidence |
| The Build with Gemini XPRIZE asked teams worldwide to pick a problem affecting 100,000 people and build, from a clean sheet in under 90 days, the company with the greatest revenue. | OBSERVATIONAL | Supported |
| A model trained on 100,000 GPUs cannot be served at the speeds users see from Astra, so frontier labs hold a larger internal model that they distil down into the served one. | INTERPRETIVE | Leans supported |
| OpenAI's next DevDay falls on 29 September 2026. | OBSERVATIONAL | Supported |
| Frontier labs are releasing models faster because Chinese open source models are at near parity with them and the labs need separation to win enterprise deals. | INTERPRETIVE | Supported |
| OpenAI has offered corporations a revenue-sharing deal under which it becomes their permanent AI partner. | OBSERVATIONAL | Leans contradicted |
| Palantir's Alex Karp publicly told companies to become an AI company or die. | OBSERVATIONAL | Leans contradicted |
| Every corporation will fundamentally need to become an AI company. | FORECAST | Not assessed |
| The modern use of GPUs in AI began with ImageNet and the ILSVRC computer vision competition, and reinforcement learning as a field began with DeepMind's deep reinforcement learning for winning games. | HISTORICAL | Supported |
| Social media rumour holds that a version of Astra offering real-time interactive control will appear before the end of 2026. | FORECAST | Not assessed |
| GPT-6 Astra has won Pokemon, Portal and other visually intensive interactive games. | OBSERVATIONAL | Supported |
Galedra