Corrections
When we get something substantively wrong, we fix it and log the fix here; small copyedits live in each article's revision history instead. Found an error? Use the "Report issue" button on the article (no account needed). So far we have received 2 reports and reviewed 1, taking a median of 4.8 days. 1 is waiting, the oldest since September 26, 2026. That is our record rather than a promise. See the verifiability policy for how checking works.
321 articles were corrected in the last 30 days. Every correction note and resolved report is listed below, newest first. A fix applied to several articles with the same note is shown once, listing each article; "See the change" opens that version in the article's revision history.
- editorial
Bring the robotics survey current through July 2026. Add hardware and mechatronics, robot learning through teleoperation-at-scale and VLA models, foundation models and the division of labour with classical control, reproducibility failures, deployment-at-scale figures, and update the safety standards section for ISO 10218-1/-2:2025 absorbing ISO/TS 15066 including power-and-force limiting
- editorial
DeepSpeed is no longer a Microsoft-governed project. Records the February 2025 contribution to LF AI and Data, the May 2025 move to the PyTorch Foundation as a hosted project, and the repository move to deepspeedai/DeepSpeed.
- editorial
Rescope the Microsoft article to Microsoft as an AI company and rebuild it; add the missing January 2025 OpenAI amendment (right of first refusal) that enabled Stargate, correct the October 2025 terms, add the full Copilot product family with current GitHub AI Credits pricing, the Phi and MAI model tables, infrastructure, and the research-software provenance including projects no longer under Microsoft governance
- editorial
Update Brain2Qwerty metrics, citation, and canonical link
- editorial
Fact-check Brain2Qwerty protocol, metrics, citations, and scope
- editorial
The Atlas 950 SuperPoD section described the Ascend 950DT as "upcoming" with no date anchor. Records the actual staged rollout: Ascend 950PR and Atlas 350 in March 2026, 950DT on Huawei Cloud in August 2026 with commercial launch in Q4 2026.
- editorial
Rebuild the Google DeepMind hub with founding record, 2014 acquisition and ICO ruling detail, full Gemini/Gemma/TPU tables, science programmes and their critiques; fix Lila Ibrahim's title, record Shazeer and Jumper departures, correct the AlphaChip retraction (a News and Views commentary, not the paper) and the AlphaStar 99.8% claim
- editorial
The HBM4 section stated SK hynix "is targeting mass production in the second half of 2025", a forward-looking claim that had expired and contradicted the next paragraph, which already recorded the completed September 2025 mass-production system.
- editorial
Rebuild the OpenAI hub to cover the full model lineage, products, infrastructure, business and litigation; fix the 900M WAU announcement date (Feb 27 2026 not March), the 2019 profit cap terms, the 2023 board-crisis outcome (Altman returned as CEO only, rejoined the board March 8 2024), current executive roles, and the Musk verdict details
- editorial
LeCun left Meta after about twelve years, not ten (FAIR founded December 2013, departure confirmed November 2025). The page stated both facts and the arithmetic contradicted itself.
- editorial
Cover both senses of Meta AI (the research organisation and the assistant product); add FAIR/GenAI/MSL org history, full Llama release table, licensing debate, research output, MTIA generations, infrastructure capex; correct OSI post byline and a miscited Reuters/CNBC reference
- editorial
Expand generative AI survey to cover all modelling families, current modality practice, and 2026 legal status; fix EU AI Act applicability (Digital Omnibus Regulation (EU) 2026/1744 deferred Annex III high-risk duties to 2 Dec 2027)
- editorial
Rebuilt from primary documents. Fixes the tariff's scope (Proclamation 11002, 91 FR 2443, covers only HTS 8471.50/8471.80/8473.30 goods inside two narrow TPP and DRAM-bandwidth windows under new U.S. note 39, headings 9903.79.01-09), states the BIS rule's real citation and conditions (91 FR 1684, effective Jan 15 2026), resolves the disputed H200 revenue-share figure (Trump announced 25 percent; the rule contains no revenue-share provision; CNAS reads the tariff as the collection vehicle), corrects the 50 percent export condition to an aggregate-TPP test, and brings the article current to August 2026 with Kessler's July 14 2026 testimony that H200 shipments to China have begun but are 'very few'. Adds Section 301, Section 5949 and VEU context and links the China semiconductor cluster.
- reader report
Confirmed and fixed. A reader reported that the Tensor Core throughput figures in the H100 performance table were roughly double the correct values. They were: the dense column had been populated with NVIDIA's sparsity-assisted numbers, and the sparsity column with those numbers doubled again. The table now reports NVIDIA's shipping-product figures with dense and 2:4-structured-sparsity values in separate columns for both H100 SXM5 and H100 PCIe. Corrected dense peaks for the SXM5 are 494.7 TFLOPS TF32, 989.4 TFLOPS BF16 and FP16, and 1,978.9 TFLOPS/TOPS FP8 and INT8, with the sparsity-assisted figures at twice each value. The DGX H100 FP8 aggregate was corrected in the same pass. Thanks to the reader who spotted the inconsistency against our own NVIDIA B200 page.
- editorial
Rewrite from a MediaWiki GPT-directory import into an encyclopedic article. Verified the listing is still live (August 2026) and corrected its status, dates and configuration against the store record and Internet Archive captures: created 2023-11-06 (not first seen 2023-11-08), last edited 2026-04-09 (not 2024-01-12), description changed GPT-4 to GPT-4o on 2024-06-13 and to model-free wording on 2026-04-09 after GPT-4o was fully retired from ChatGPT on 2026-04-03. Removed the unsourced system-prompt text (the record shows instructions: null), the duplicate system-prompt section, four empty scaffold sections, the dead <comments /> tag and directory filler. Added the separate July 2026 ChatGPT Classic desktop application.
- editorial
Rewrote SEO FAQ headings as encyclopedic sections and brought the article current to August 1, 2026. Added MiniMax H3 (July 31, 2026): 2K, native stereo audio, 4-15s, API identifier MiniMax-H3. Corrected Video-01 launch to August 31, 2024 (MiniMax), I2V-01-Live to January 12, 2025, and Master plan price to $94.99 (was $119.99); added Ultra and Max tiers. Removed unsourced claims (Hailuo 02 MoE backbone, 92.1/87.3 benchmark scores, 70+ languages, traffic rankings). Replaced stale March 2026 leaderboard with current Artificial Analysis standings and added third-party availability, copyright litigation and limitations.
- editorial
Two of the three Check Point advisories were described wrongly. GHSA-ph6w-f82w-28w6 concerns the startup trust warning's wording, not hooks in settings.json, and CVE-2025-59536 is a startup trust dialog code injection, not an MCP consent bypass. Descriptions, fixed-in versions and advisory citations now come from the primary records. Currency pass to August 2026: rewrote model lineup (Opus 5/Sonnet 5/Fable 5 aliases, account-type defaults, effort levels, fast mode), rebuilt the interfaces section around all five first-party surfaces, added permission modes including auto mode, adde
- editorial
Rewrote as an encyclopedic article verified against current code.claude.com documentation (Claude Code 2.1.220). Replaced the SEO FAQ headings, deleted the Conclusion, External Links and "official Anthropic sources" sections, and merged four overlapping safety sections into two. Corrected the bypass-mode status bar string, the plan-mode description, the acceptEdits command list, auto mode's plan/model/provider requirements (Team and Enterprise no longer require Owner opt-in; CLAUDE_CODE_ENABLE_AUTO_MODE was removed in v2.1.207), the classifier model and its branch-push rules, and removed unverified version and incident claims. Cut from 34,570 to 21,471 characters.
- editorial
Attribute the SSA800 immersion scanner design to SMEE, which transferred the technology to Shanghai Yuliangsheng in 2025, rather than leaving it to read as Yuliangsheng's or SiCarrier's own design
- editorial
Rebuild the V4-Pro benchmark tables from the model card. The competitor column labelled GPT-5.5 was GPT-5.4 xHigh; its SWE-bench Verified 88.7% is not on the card, its SWE-bench Pro 58.6% was Kimi K2.6's score and its Terminal-Bench 82.7 was GPT-5.4's BrowseComp score. Two Claude Opus 4.6 cells and one MMLU-Pro cell were also wrong, and an unsourced GPT-5.2 column is replaced.
- editorial
DeepSeek published open weights for DeepSeek-V4-Flash-0731 on July 31, 2026 (MIT, 48 shards, ~167 GB), so the hub no longer describes the official Flash build as API-only. Added a correctly-framed note that the V4-Pro-DSpark and V4-Flash-DSpark repositories are the same checkpoints with a speculative decoding module attached, confirmed V4-Pro is still the April preview with no official build, re-verified API pricing and the still-inactive peak/off-peak policy against DeepSeek docs, and cross-linked the new DeepSeek V4-Pro page.
- editorial
Rewrite Supervised fine-tuning with corrected stage boundaries, repaired objective, and protocol-bounded evidence.
- editorial
Independently reverify Jensen Huang against SEC filings, institutional records, technical papers, and strong reporting; correct fiscal 2026 compensation and OSU timing, remove disputed or volatile claims, and rebuild explicit citations.
- editorial
Independently reverify and update Information Retrieval foundations, evaluation, neural architectures, benchmarks, RAG, fairness, and operations using primary and academic sources through July 28, 2026
- editorial
Independently reverify and update Amazon history, AI products, investments, robotics, financial scale, and governance using primary and authoritative sources through July 28, 2026