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
Rebuilt Artificial General Intelligence from primary and peer-reviewed sources; separated proposed definitions from consensus, bounded current capabilities and forecasts, and removed unsupported fast-aging claims
- editorial
Rebuilt Tool use from primary research and official specifications; corrected execution architecture, MCP governance, OWASP numbering, and tau-bench claims; removed duplicated and unstable score tables
- editorial
Replaced unsupported product-market claims with primary evidence, corrected study statistics, and distinguished task, firm, and economy-wide productivity
- editorial
Reverify the DALL-E family, replace an unsupported DALL-E 2 component total and nonexistent DALL-E 3 accuracy figure, remove unstable product claims, and rebuild the article around primary papers, dated official records, and independent academic audits.
- editorial
Independently fact-check Question answering, correct the LUNAR result, replace unsupported claims, add 35 explicit primary or academic references, and validate canonical internal links.
- editorial
Replace speculative SWE-bench material with verified protocol, variants, and audit evidence
- editorial
Independently verify MIT License history, terms, compatibility, adoption evidence, and AI scope
- editorial
Independently reverify Cursor features, security, current product scope, and academic citations
- editorial
Rebuilt the ResNet article from primary papers and official documentation; corrected the ImageNet evaluation table, the earlier human-reference chronology, the COCO baseline AP values, the original training duration, the ResNet-RS 83.4% model identity, and modern v1/v1.5/v2 distinctions.
- editorial
Independently revalidated Softmax and corrected temperature/filter ordering, made the PyTorch axis wording version-robust, and clarified additive logit identifiability; rechecked all claims, citations, equations, sources, links, tables, and revision history.
- editorial
Replaces the false GPT-2 redirect with a sourced DistilGPT2 article that establishes its distinct identity, reconciles parameter and evaluation labels, and documents architecture, distillation, training data, use, and limitations.
- editorial
Reverify GitHub Copilot features, billing, data terms, research, and litigation
- editorial
Rebuilds GPT-2 as the four-checkpoint 2019 OpenAI model family; reconciles corrected parameter names, distinguishes no-fine-tuning evaluations from demonstration-free prompting, qualifies benchmarks and detection, documents WebText and staged-release limits, and adds primary and peer-reviewed evidence on bias, toxicity, memorization, privacy, distribution, and legacy.
- editorial
Separated the April 2024 Llama 3 checkpoints from later releases, replaced unsupported chronology and evaluative wording with source-bounded facts, corrected launch-era training and license scope, and rebuilt continuous citations with complete PDF evidence.
- editorial
Removed speculative GPT-4-and-later architecture and parameter claims, corrected GPT-3 training-token and GPT-5 context figures, updated the release history through GPT-5.6, clarified trademark status and GPT product terminology, and rebuilt the article from primary, official, academic, procedural, and security sources.
- editorial
Reverify Object detection, replace unsupported market and cross-hardware benchmark claims, correct dataset and detector specifications, and add deterministic primary-source citations and canonical internal links.
- editorial
Corrects the R1 training-cost attribution and R1-Zero benchmark endpoint; rebuilds architecture, training, evaluation, licensing, update, safety, and impact coverage with protocol-specific citations and clear open-weight and hosted-service distinctions.
- editorial
Rebuild the Embeddings article from primary research, distinguish lookup, contextual, sentence, multimodal, graph, multi-vector, and hosted API representations, remove unsupported performance and cost claims, and add dated official interface facts, compatibility rules, evaluation guidance, privacy evidence, and complete citations.
- editorial
Rebuilt the biography from primary records and authoritative reporting; corrected the false 2025 OpenAI-equity claim, qualified financing and Stargate plans, removed volatile wealth claims, and clarified procedural legal outcomes.
- editorial
Narrowed the page to trained-model execution, removed unsupported forecasts and vendor comparisons, and added primary-source coverage of serving, LLM inference, optimization, benchmarking, and monitoring.
- editorial
Reverified Reasoning models with bounded terminology, historical precursors, disclosed training and inference methods, protocol-aware evaluation, and qualified limitations; replaced unsupported chronology and cross-model rankings with 13 primary and academic sources.
- editorial
Replaces a bloated and stale product account with a current, date-bounded description of Claude Code's interfaces, licensing, installation, plans, security controls, data handling, release history, and research evidence, while removing unsupported market and performance claims.
- editorial
Fixed mixed-condition benchmark values, separated prompted CoT from trained reasoning, qualified scale, theorem, faithfulness, and monitorability claims, removed unsupported product and application claims, and rebuilt citations from primary research.
- editorial
Reverify Backpropagation with corrected reverse-mode, layered-network, BPTT, clipping, cost, memory, framework, biological-scope, and historical claims; normalize citations and canonical internal links.
- editorial
Replace oversimplified and prescriptive claims with sourced definitions of explicit, Bayesian, algorithmic, data-based, and neural-network regularization; correct penalty scaling, Laplace-prior, dropout, early-stopping, weight-decay, label-smoothing, spectral-normalization, BatchNorm, and evaluation claims.