Corrections log
Every substantive correction to the glossary is recorded here, with its date and the entries concerned. Changes of form (typography, style) are not listed.
This log lists corrections published since 14 October 2026. Editorial method
Five entries added: deepfake, deep research, chain of thought, Agent Skills, Google AI Overviews and AI Mode.
Entries concerned: Deepfake, Deep research, Chain of thought, Agent Skills, AI Overviews and AI Mode
Error corrected in the example: the study cited concerns Japan's Cancer Information Service, not the US one. Connected to this reliable source, RAG chatbots fall to 0% (GPT-4) and 6% (GPT-3.5) hallucinations; the 19% and 35% rates concerned questions outside the source's scope.
Entries concerned: RAG (Retrieval-Augmented Generation)
Several errors corrected: the Digital Omnibus adds two prohibitions, not one; the fine in the high-risk example is 15 million euros or 3% of turnover, not 35 million or 7%. Three clarifications added: biometric categorisation that infers sensitive data is prohibited, a human operator is not an obligation for a simple chatbot, and obligations are split between provider and deployer.
Entries concerned: AI Act
The 5 to 15% range is now tied to its source (Vectara leaderboard, September 2026) and to what it measures: information added when summarising a provided document, not factual questions in general.
Entries concerned: Hallucination
Prices updated after Claude Haiku 5.5 was released on 7 October 2026 ($0.10 input and $0.50 output per million tokens, for prompts up to 100,000 tokens). Sonnet 5.5 cache reads corrected: 5% of the input price.
Entries concerned: Cost per token, Token, AI API, Context window, LLM (Large Language Model)
Errors and unsourced figures corrected: calculation error in the ROI example (233%, not 240%); typical breakdown of TCO costs removed for lack of a source; GDPR fine and mechanisms for transfers outside the EU clarified; GPT-4 parameter count removed (OpenAI never published it); the French Armed Forces agreement with Mistral AI reported from the ministry's announcement.
Entries concerned: AI ROI, AI TCO, GDPR and AI, LLM (Large Language Model), MoE (Mixture of Experts), Digital sovereignty
Facts updated or reported more accurately: LangChain observability statistic (89% of respondents to its late-2025 survey); Pinecone European region opened in May 2026; InstructGPT study result; roughly 24-hour delay stated by OpenAI for changes to take effect; token use of multi-agent systems according to Anthropic; authors of the GitHub Copilot study; current Claude Code commands; open-weight models as of October 2026.
Entries concerned: Agentic loop, Vector database, Alignment, Conversational indexing, Agent orchestration, Copilot, Claude Code, Open-source model
Full review of the 95 existing entries against primary sources. Figures we could not find were removed or replaced with sourced figures, and invented examples are now marked "Illustrative case".