AI Wiki

Explore AI Wiki

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 3 reports and reviewed 3, taking a median of 8.8 days. That is our record rather than a promise. See the verifiability policy for how checking works.

1,628 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.

  1. editorial

    Removed a fabricated Simon Willison quote ('share button for code') and fabricated details of his October 2024 post, restored two altered Anthropic quotes, corrected the file-upload date (July 17, 2025), the Pragmatic Engineer date (August 27, 2024), the help-center content-type claim, the CDN allowlist, a composite traffic statistic, competitor rows (ChatGPT Canvas renders HTML and React since January 2025), the claim that viewers of shared Artifacts need an account, and nonexistent monthly inference budgets, re-cited quotes to their real sources, and other errors (plus 28 minor fixes)

  2. editorial

    Corrected the Claude 3.7 Sonnet launch date (24 February 2025), the origin of the effort parameter (Claude Opus 4.5, not 4.6), a misread Sonnet 4.6 cost comparison, the Opus 4.6 benchmark effort setting (max, not high), a customer testimonial presented as Anthropic's claim, an unsupported 'less prone to overthinking' quote, Gemini's sixfold figure (a price gap, not a measured saving), Mythos Preview's manual-thinking support and the prompt-caching behavior on mode switches, and updated the model and effort tables for the Claude 5-series models released June to October 2026 (plus 65 minor fixes)

  3. editorial

    Updated Yaskawa's leadership (Ogasawara became chairman and president in May 2026; former president Ogawa stepped down and died in October 2026) and its latest results (FY ended Feb 2026), and corrected the US-dollar revenue conversion, the segment shares and the description of the Other segment, the Big Four market-share figures, and an unsupported claim about the installed base doubling (plus 27 minor fixes)

  4. editorial

    Corrected Ōkuma's prime-ministerial number and Takata Sanae's role, the WL-series dynamic-walking milestone (WL-10RD, 1984), WABOT-1's static gait, the description of WHL-11 (a hydraulic biped shown at Expo '85, not a vehicle-mounted leg), and two alumni entries (Murakami's graduation year; Lee Hae-chan, who did not attend Waseda) (plus 29 minor fixes)

  5. editorial

    Corrected the DEX-EE unveiling date (9 May 2024, not 30 January 2025) and removed the claim that Shadow still sells pneumatic hands (its current line is all motor-driven) (plus 22 minor fixes)

  6. editorial

    Corrected the CoreWeave position history (opened in Q4 2025 at about $230 million; no separate $326 million addition), removed the false claim that the Anthropic Series H was the largest startup financing in history, removed unsupported total-AUM figures, and updated the 13F portfolio to the June 30, 2026 filing (plus 13 minor fixes)

  7. editorial

    Corrected the paper's headline perplexity comparison (5.40 is StreamingLLM, not a recomputation oracle, on a 65K-token book), the BERT/ViT sink finding ([SEP], not [CLS]), a speculative gpt-oss quantisation claim presented as reported, and the stale description of SinkCache, which was removed from Hugging Face Transformers in v4.53.0 (plus 16 minor fixes)

  8. editorial

    Corrected the Arctic benchmark table to Snowflake's published (corrected) figures, removed a quote misattributed to Snowflake's founders, and updated stale product status: Snowflake Intelligence renamed Snowflake CoWork, Cortex Analyst superseded by Cortex Agents, Cortex Fine-tuning scheduled for deprecation, and the February 2026 OpenAI deal (plus 19 minor fixes)

  9. editorial

    Corrected a composite ActAdd quote; wrong authors and years for REMEDI, In-Context Vectors, the sycophancy post, Steering Awareness and the collateral-damage paper; misstated findings of the Tan et al. reliability paper, Anthropic's feature-steering bias study, the sentiment-direction paper, the refusal-direction paper and the sycophancy post (base model, not chat; no capability numbers); a false 'first' claim; and DeepMind credit for academic and Microsoft work (plus 15 minor fixes)

  10. editorial

    Corrected altered or composite quotes from the Bletchley Declaration, Seoul Declaration and Seoul Ministerial Statement; fixed framework dates and titles (Google DeepMind, Microsoft, NVIDIA, Magic); removed an invented Singapore/Israel signatory, unsourced lists of attending leaders and executives, a Bletchley-era open letter misattributed to Seoul, an invented CSIS warning and unsourced CEO statements; corrected China's participation, the safety-institute network (renamed December 2025) and the India AI Impact Summit details (plus 20 minor fixes)

  11. editorial

    Replaced a fabricated ablation table with the paper's actual Table 2 figures, corrected fabricated descriptions of the paper's Figure 1 and of a nonexistent three-critic ablation and Figure 4 delay result, and corrected the claim that TD3 won six of seven MuJoCo tasks (baseline DDPG scored higher on InvertedDoublePendulum) (plus 19 minor fixes)

  12. editorial

    Updated Sutton's affiliation (he left Keen Technologies in July 2026 to co-found Oak Lab), corrected a list of 'former students' that wrongly included faculty colleagues Michael Bowling and Csaba Szepesvari, removed a false description of Jonathan Schaeffer as 'the late' author of a computer-Go program, corrected the attribution of the 1995 TD-dopamine study, and replaced an unsourced quote attributed to Sutton and Carmack (plus 24 minor fixes)

  13. editorial

    Corrected Inspire Robots' 2026 shipment target (30,000 to 50,000 hands, not 50,000 to 100,000), added its February 2026 C1/C2 funding rounds, and removed an unsupported claim that Agility and Skild AI are customers (plus 7 minor fixes)

  14. editorial

    Corrected the Unitree G1 list price (from about USD 16,000 to about USD 13,500 per Unitree's page) and removed several claims the cited sources do not support, including an Apptronik dual purchase/service model and Tesla's supposed rejection of a service model (plus 16 minor fixes)

  15. editorial

    Replaced HotpotQA results that do not appear in the Reflexion paper with the paper's own Table 5 figures (CoT (GT) 60% to 77%, ReAct 30% to 55%) (plus 9 minor fixes)

  16. editorial

    Corrected the stated typical range of CLIP Score values (observed scores fall between 0 and about 1, not 0 to 2.5) (plus 9 minor fixes)

  17. editorial

    Corrected a Vinyals quote that described Gemini Live, not Project Astra, and the date Gemini Live was announced (I/O May 2024, launched August 13, 2024) (plus 8 minor fixes)

  18. editorial

    Corrected the EU AI Act description (the fundamental rights impact assessment falls on certain deployers, and Annex III obligations now apply from December 2, 2027), updated the Colorado AI Act (repealed and replaced by SB 26-189 in May 2026), and added the 2025 federal executive order against disparate-impact liability (plus 16 minor fixes)

  19. editorial

    Corrected the keyword-priority table to match the published DOCTOR script, Weizenbaum's stated reason for the name ELIZA, the 2021 date and finders of the original code, the title and venue of the 2024 Jones and Bergen Turing-test paper, and the details of the BBN vice president anecdote (plus 18 minor fixes)

  20. editorial

    Corrected the Goyal et al. (2017) result: 23.74% top-1 error was the large-minibatch run's error, versus 23.60% for the small-batch baseline (plus 6 minor fixes)

  21. editorial

    Corrected the description of the original LSI test results (MED and CISI only, no gain on CISI; Berry et al.'s TREC run used a 70,000-document sample, not 1 million documents) and replaced a wrong Howard and Kahana citation with their 2002 LSA free-recall paper (plus 6 minor fixes)

  22. editorial

    Corrected the paper honored by her 2021 CVPR Longuet-Higgins Prize (the 2011 CVPR "Baby Talk" paper, not 2013 work) (plus 23 minor fixes)

  23. editorial

    Corrected the false statement that Mike Lewis is not an author of the RAG paper (he is its eighth author), and updated his affiliation and role, which cited sources described only through April 2025, to note his 2026 Meta Superintelligence Labs affiliation (plus 21 minor fixes)

  24. editorial

    Corrected the claim that NumPy and scikit-learn solve least squares by QR decomposition (both use SVD-based LAPACK solvers; only R's lm() uses QR) (plus 12 minor fixes)

  25. editorial

    Corrected the K1 launch-event date (October 24, 2025), the Series A and A+ funding amounts and leads, the K1 edition specs, the Unitree G1 height, and unsupported RoboCup partnership terms and RoboCup 2025 team-country list; added the April 2026 funding round, the Booster T2 launch and RoboCup 2026 results (plus 17 minor fixes)