[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"kb-article-when-nonfiction-lies-engineering-lessons-from-ai-fabricated-quotes-in-the-future-of-truth-en":3,"ArticleBody_xQBPRRgrM9neW8OWI6OOOy1e0oAK9CqvSC1sEjhjM":197},{"article":4,"relatedArticles":167,"locale":54},{"id":5,"title":6,"slug":7,"content":8,"htmlContent":9,"excerpt":10,"category":11,"tags":12,"metaDescription":10,"wordCount":13,"readingTime":14,"publishedAt":15,"sources":16,"sourceCoverage":46,"transparency":48,"seo":51,"language":54,"featuredImage":55,"featuredImageCredit":56,"isFreeGeneration":60,"trendSlug":61,"niche":62,"geoTakeaways":65,"geoFaq":74,"entities":84},"6a11fbf252421694669491e9","When Nonfiction Lies: Engineering Lessons from AI‑Fabricated Quotes in “The Future of Truth”","when-nonfiction-lies-engineering-lessons-from-ai-fabricated-quotes-in-the-future-of-truth","An author publishing AI‑fabricated quotes in a nonfiction book is not a quirky misuse of [ChatGPT](\u002Fentities\u002F6a0e316d07a4fdbfcf5ea647-chatgpt). It is a production incident.\n\nYou have:\n\n- A generative model that invents sources.\n- An operator who treats outputs as ground truth.\n- No guardrails, [provenance](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FProvenance), or verification loop.\n\nThis is the exact failure pattern enterprises are warned about when deploying LLMs into workflows that touch customers, regulators, or the public record.[1][2]\n\nThis article treats the scandal around *The Future of Truth* as an engineering post‑mortem. We map symptoms—fabricated quotes, fake citations, undetected errors—to known LLM behaviors, then outline an architecture that makes this class of failure far less likely.\n\n---\n\n## From literary scandal to post‑mortem: what AI‑fabricated quotes reveal\n\nAI‑invented quotes in a nonfiction book mirror LLMs fabricating academic references that look real.[7] The model generated plausible content; a human over‑trusted it; the process lacked systematic checks, so fabrications shipped.\n\nLLMs hallucinate with conviction:\n\n- They output authoritative‑sounding text with no grounding in real documents or events.[1][7]\n- When this text flows into “high‑stakes” surfaces—books, reports, public FAQs—the risk becomes reputational and potentially regulatory.[1][2]\n\n💼 **Incident framing**\n\nAs an internal post‑mortem, the timeline might read:\n\n- **T‑0**: Author uses an LLM for research.\n- **T‑1**: Model invents a plausible quote and attribution.[7]\n- **T‑2**: Author pastes it into the manuscript, unlabeled.\n- **T‑3**: Editorial review checks style, not facts.\n- **T‑4**: Book is published; readers discover fabrications.\n\nEach step exposes missing controls. Guardrail guidance highlights three layers—input control, output moderation, governance—that should have broken this chain.[1]\n\nA close parallel: a product manager asked a model for “three recent peer‑reviewed studies” on their niche. The LLM returned formatted citations with real authors and journals—but the papers did not exist.[7] The PM nearly presented them to customers before checking Google Scholar.\n\n⚠️ **Lesson:** once AI‑generated text enters your pipeline without provenance, downstream reviewers cannot reliably separate truth from [hallucination](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FHallucination).[2][6] As agentic systems and OS‑level copilots spread, unmarked AI content can silently propagate across drafts, wikis, and release notes unless you design for transparency.[3][4][5]\n\n---\n\n## Why LLMs hallucinate quotes and citations\n\nLLMs are trained to predict the next token, not to retrieve only verifiable facts.[7] They model what text *should look like*, not whether it is true.\n\nConsequences:\n\n- If a prompt implies a quote *should* exist, the model may invent one that fits the narrative.\n- When asked for citations, it fabricates convincing references—authors, journals, titles—that have never existed.[7]\n\n📊 **Typical hallucination pattern**\n\nAnalyses of hallucinated citations show that:[7]\n\n- Authors are real researchers.\n- Journals or conferences are real venues.\n- Titles look stylistically correct.\n- The specific article is nonexistent.\n\nThis is analogous to fabricating a political quote that matches a speaker’s views but was never said.\n\nResearch on hallucinations highlights root behaviors:[7]\n\n- **Accepting false premises** – treats fabricated prompt context as true and builds around it.\n- **Misleading context** – mirrors wrong information in provided documents, even against pre‑training.\n- **Sycophancy** – optimizes for user satisfaction, agreeing rather than defending correctness.\n\nEnterprise guidance is blunt: even top models—GPT‑class, [Claude](\u002Fentities\u002F6a0a74001f0b27c1f426a613-claude), LLaMA, [Mistral](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FMistral_(missile))—emit inaccurate or fabricated content without guardrails and process controls.[1]\n\n💡 **Systemic, not a bug**\n\nGovernance‑focused literature treats hallucination as inherent, to be *managed* with policy, oversight, and architecture, not eliminated.[2][1] In autonomous or agentic setups, risk compounds: an agent tasked with “complete the research chapter” will fill gaps with invented bridges or quotes if rewarded for task completion rather than factual accuracy.[4][5]\n\n---\n\n## Guardrails for nonfiction: input, output, and governance\n\nThe standard guardrail model—input control, output moderation, governance—used for chatbots and assistants maps cleanly onto nonfiction editorial workflows.[1]\n\n### Input control: constrain what the model is allowed to do\n\nInput control governs allowed prompt types.[1]\n\nFor nonfiction:\n\n- **Allow**:\n  - “Rewrite this verified quote for clarity; keep attribution.”\n  - “Summarize these interview notes into three bullet points.”\n- **Block or warn**:\n  - “Invent a plausible quote by X supporting Y.”\n  - “Find three sources that say Z” if you cannot ground them in a reference corpus.\n\nBecause hallucinations often appear as made‑up references and quotes,[7] limiting the model to transformations of verified material shrinks the attack surface.[1]\n\n⚡ **Practical pattern**\n\nWrap model access in a thin SDK instead of direct API calls:\n\n```python\ndef rewrite_quote(quote_id: str, style: str) -> str:\n    quote = get_verified_quote(quote_id)\n    prompt = f\"Rewrite this quote in a {style} style, keep meaning and attribution.\\n\\n{quote.text}\"\n    return call_llm(prompt, system=\"You must not introduce new facts or quotes.\")\n```\n\nNo endpoint exists for raw “generate quotes.”\n\n### Output moderation: scan drafts for suspicious content\n\nOutput moderation uses a secondary model or rules engine to inspect generated text and flag likely hallucinations.[1]\n\nFor nonfiction:\n\n- Use regex + NER to detect quotes and citations.\n- Require every quote to link to a record in your notes or transcript system.\n- Run retrieval over your corpus for each quote; if no fuzzy match exists, flag as high‑risk.[7]\n\nTreat this like CI for manuscripts: fail the build if any quote lacks provenance.\n\n💼 **Callout: hallucination signals**\n\n- Specific dates, venues, or titles not present in your corpus.\n- Many direct quotes from figures with sparse source material.\n- Sudden appearance of perfectly formatted references late in the workflow.[7]\n\n### Governance: process, policy, and accountability\n\nGovernance provides policies and oversight.[2]\n\nFor nonfiction, define that:\n\n- Every quote carries provenance metadata (source, location, verification status).\n- Any “AI‑assisted research” section requires sign‑off from a designated editor.[2][6]\n- For regulated markets, you maintain documentation of hallucination‑mitigation measures aligned with AI Act expectations on risk management and transparency.[6][1]\n\nAgentic AI in content workflows must be explicitly constrained:\n\n- Agents cannot inject new factual claims or quotes without verifiable sources or explicit “model‑generated” labels.[4][5]\n- As OS‑level platforms like [Ubuntu](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FUbuntu) prompt users about enabling AI features, editorial stacks should similarly force explicit choices on where AI is allowed in the process.[3][2]\n\n---\n\n## Architectures to prevent fabricated quotes: [RAG](\u002Fentities\u002F69d15a4e4eea09eba3dfe1b0-rag), tools, and verification loops\n\nGuardrails specify *what* should happen; architecture handles *how* it happens.\n\n### Grounding with RAG\n\nFor nonfiction, a Retrieval‑Augmented Generation (RAG) pipeline should be standard:\n\n1. Index verified sources: transcripts, emails, books, reports.\n2. For any request that might produce a quote, retrieve candidate passages.\n3. Constrain generation: the model may paraphrase but not invent text unsupported by retrieved snippets.[1][7]\n\n💡 **Key rule**\n\n> “No source, no quote.”\n\nIf retrieval returns nothing relevant, the model must refuse or ask for more data, not fabricate.[7]\n\nBecause hallucinated citations often look highly plausible,[7] your RAG stack needs a verification pass:\n\n- Extract every quote and citation from the output.\n- Match against your index (exact + fuzzy search).\n- Flag low‑similarity items for human review.\n\n### Agentic verification loops\n\nAgentic AI patterns can enforce verification instead of bypassing it.[4][5]\n\nDesign a narrow verification agent:\n\n- **Input**: manuscript section.\n- **Tools**:\n  - `search_corpus(query) -> list[passages]`\n  - `search_web(query) -> list[urls]`\n- **Behavior**:\n  - For each quote or factual claim, seek a supporting passage.\n  - Label each as `verified`, `partially_verified`, or `unverified`.\n  - Refuse “publishable” status if any high‑impact claim is `unverified`.\n\nRegulatory commentary on the AI Act stresses explainability and documentation: such agents should leave an auditable trail showing which tools they used and why a quote was accepted or rejected.[6][2]\n\n📊 **Infrastructure choices**\n\nWith OS‑level inference endpoints like Ubuntu’s OpenAI‑compatible local APIs, you can:\n\n- Run drafting and verification locally.\n- Keep sensitive interviews and drafts on your own infrastructure.[3][2]\n\nThis is crucial for investigative journalism and regulated industries with confidential sources.\n\n---\n\n## Trust, compliance, and sovereign AI: why provenance matters beyond books\n\nThis incident is not just about one book. It illustrates how organizations can build or erode public trust when relying on AI in knowledge work.\n\nEnterprise AI guidance insists that progress must be tied to clear principles and ethical discipline, especially where systems shape public understanding.[2] Nonfiction, long‑form reports, and policy documents fall squarely here.\n\n⚠️ **Regulatory pressure is rising**\n\nThe EU AI Act treats hallucinations, bias, and opacity as concrete risks that can trigger obligations, penalties, and documentation requirements.[6][1] Using AI‑generated factual content in products or public communication without controls is becoming a compliance issue, not just a PR risk.\n\nHallucinations and misinformation can:[1][7]\n\n- Undermine trust among users and stakeholders.\n- Create legal exposure when decisions rely on false information.\n- Force costly retractions and public incident responses.\n\n“Sovereign” and “trusted” AI narratives therefore stress knowing where models run, what data they use, and how outputs are controlled and audited.[2][6] The same mindset should govern AI in research, whitepapers, analyst notes, and nonfiction writing.\n\n💼 **Multi‑agent editorial model**\n\nAgentic AI guidance recommends orchestrating specialized agents under clear constraints and oversight.[5][4] For publishing, define:\n\n- **Research agent** – retrieves and clusters documents; cannot generate quotes.\n- **Summarization agent** – compresses source material; must attach citations.\n- **Verification agent** – cross‑checks claims and blocks unsupported content.\n\nHuman editors remain accountable, but the system adds guardrails and logs.\n\nAs general‑purpose platforms like Ubuntu embed AI into common tools, organizations need consistent provenance policies:[3][1]\n\n- When AI assistance is allowed.\n- How it must be disclosed.\n- What extra checks are mandatory before AI‑touched content goes public.\n\n---\n\n## Designing production‑grade editorial pipelines with AI and humans in the loop\n\nAssume hallucinations *will* occur. Then design editorial workflows like production systems.\n\n### 1. Inventory and classification\n\nWebinar guidance on AI compliance notes many organizations already use AI without grasping regulatory implications.[6]\n\nSteps:\n\n- Inventory every touchpoint where LLMs touch content:\n  - Research: search, summarization.\n  - Drafting: co‑writing, outlining.\n  - Editing: style, tone, translation.\n  - Review: fact‑checking, QA.\n- Classify each use case by risk: internal notes vs public reports vs legally sensitive docs.\n\n### 2. Policy + technical enforcement\n\nBest practices emphasize policies, monitoring, and feedback loops.[1][2] Policies alone are insufficient; enforce technically where possible:\n\n- Restrict “high‑risk” content to tools that log prompts, models, and outputs.\n- Block direct paste from general‑purpose chatbots into authoring tools for certain document types.\n- Require internal flags on AI‑assisted sections for reviewers.\n\n💡 **Minimal telemetry to log**\n\n- `model_id`, `version`\n- `prompt_template` and task type\n- `source_documents` used (IDs)\n- `verification_status` of claims and quotes\n\n### 3. Automated checks before human review\n\nHallucinated citations slip through when reviewers trust surface plausibility.[7] Shift initial checks to the system:\n\n- CI step: run quote and citation verification on each merge request or manuscript milestone.\n- Dashboard: show editors “unverified” sections requiring deeper review.\n- Metrics: track hallucination incidents, false positives, and time to resolve.[1]\n\nAgentic AI guidance holds that multi‑step agents need defined goals and constraints, with humans validating high‑impact outcomes.[5][4] In editorial work, agents may propose, cluster, and verify—but not approve publication.\n\n📊 **On‑prem and local models**\n\nEmbedding local LLMs into developer and content environments, as OS vendors begin to do, enables:[3][2]\n\n- Keeping drafts and corpora on your own infrastructure.\n- Rich logging and audit without third‑party data sharing.\n- Consistent policy enforcement via a shared local API.\n\n---\n\n## Conclusion: treat AI‑assisted facts like production systems, not writing toys\n\nAI‑fabricated quotes in a nonfiction book show how easily hallucinations can cross from model output into public record when guardrails, provenance, and verification are absent.[1][7] The failure pattern matches known hallucination behaviors, enterprise warnings about misinformation and trust erosion, and growing regulatory expectations for responsible AI in decision‑impacting contexts.[2][6]\n\nFor ML and platform engineers, the takeaway is direct: any AI‑assisted factual content is a production system. It needs grounding architectures like RAG, strict input constraints, agentic verification, audit trails, and human oversight—not just clever prompting.[1][4][7]\n\nBefore you ship another AI‑assisted article, report, or product surface:\n\n- Map your workflow against the guardrail, governance, and verification patterns outlined here.\n- Identify every step where a model could invent a quote or citation without detection.\n- Add concrete controls—retrieval, cross‑checks, logging, human review—to close those gaps and avoid becoming the next cautionary tale.","\u003Cp>An author publishing AI‑fabricated quotes in a nonfiction book is not a quirky misuse of \u003Ca href=\"\u002Fentities\u002F6a0e316d07a4fdbfcf5ea647-chatgpt\">ChatGPT\u003C\u002Fa>. It is a production incident.\u003C\u002Fp>\n\u003Cp>You have:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>A generative model that invents sources.\u003C\u002Fli>\n\u003Cli>An operator who treats outputs as ground truth.\u003C\u002Fli>\n\u003Cli>No guardrails, \u003Ca href=\"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FProvenance\" class=\"wiki-link\" target=\"_blank\" rel=\"noopener\">provenance\u003C\u002Fa>, or verification loop.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>This is the exact failure pattern enterprises are warned about when deploying LLMs into workflows that touch customers, regulators, or the public record.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>This article treats the scandal around \u003Cem>The Future of Truth\u003C\u002Fem> as an engineering post‑mortem. We map symptoms—fabricated quotes, fake citations, undetected errors—to known LLM behaviors, then outline an architecture that makes this class of failure far less likely.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>From literary scandal to post‑mortem: what AI‑fabricated quotes reveal\u003C\u002Fh2>\n\u003Cp>AI‑invented quotes in a nonfiction book mirror LLMs fabricating academic references that look real.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa> The model generated plausible content; a human over‑trusted it; the process lacked systematic checks, so fabrications shipped.\u003C\u002Fp>\n\u003Cp>LLMs hallucinate with conviction:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>They output authoritative‑sounding text with no grounding in real documents or events.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>When this text flows into “high‑stakes” surfaces—books, reports, public FAQs—the risk becomes reputational and potentially regulatory.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>💼 \u003Cstrong>Incident framing\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>As an internal post‑mortem, the timeline might read:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>T‑0\u003C\u002Fstrong>: Author uses an LLM for research.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>T‑1\u003C\u002Fstrong>: Model invents a plausible quote and attribution.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Cstrong>T‑2\u003C\u002Fstrong>: Author pastes it into the manuscript, unlabeled.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>T‑3\u003C\u002Fstrong>: Editorial review checks style, not facts.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>T‑4\u003C\u002Fstrong>: Book is published; readers discover fabrications.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Each step exposes missing controls. Guardrail guidance highlights three layers—input control, output moderation, governance—that should have broken this chain.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>A close parallel: a product manager asked a model for “three recent peer‑reviewed studies” on their niche. The LLM returned formatted citations with real authors and journals—but the papers did not exist.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa> The PM nearly presented them to customers before checking Google Scholar.\u003C\u002Fp>\n\u003Cp>⚠️ \u003Cstrong>Lesson:\u003C\u002Fstrong> once AI‑generated text enters your pipeline without provenance, downstream reviewers cannot reliably separate truth from \u003Ca href=\"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FHallucination\" class=\"wiki-link\" target=\"_blank\" rel=\"noopener\">hallucination\u003C\u002Fa>.\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa> As agentic systems and OS‑level copilots spread, unmarked AI content can silently propagate across drafts, wikis, and release notes unless you design for transparency.\u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>\u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>Why LLMs hallucinate quotes and citations\u003C\u002Fh2>\n\u003Cp>LLMs are trained to predict the next token, not to retrieve only verifiable facts.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa> They model what text \u003Cem>should look like\u003C\u002Fem>, not whether it is true.\u003C\u002Fp>\n\u003Cp>Consequences:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>If a prompt implies a quote \u003Cem>should\u003C\u002Fem> exist, the model may invent one that fits the narrative.\u003C\u002Fli>\n\u003Cli>When asked for citations, it fabricates convincing references—authors, journals, titles—that have never existed.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>📊 \u003Cstrong>Typical hallucination pattern\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Analyses of hallucinated citations show that:\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Authors are real researchers.\u003C\u002Fli>\n\u003Cli>Journals or conferences are real venues.\u003C\u002Fli>\n\u003Cli>Titles look stylistically correct.\u003C\u002Fli>\n\u003Cli>The specific article is nonexistent.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>This is analogous to fabricating a political quote that matches a speaker’s views but was never said.\u003C\u002Fp>\n\u003Cp>Research on hallucinations highlights root behaviors:\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Accepting false premises\u003C\u002Fstrong> – treats fabricated prompt context as true and builds around it.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Misleading context\u003C\u002Fstrong> – mirrors wrong information in provided documents, even against pre‑training.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Sycophancy\u003C\u002Fstrong> – optimizes for user satisfaction, agreeing rather than defending correctness.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Enterprise guidance is blunt: even top models—GPT‑class, \u003Ca href=\"\u002Fentities\u002F6a0a74001f0b27c1f426a613-claude\">Claude\u003C\u002Fa>, LLaMA, \u003Ca href=\"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FMistral_(missile)\" class=\"wiki-link\" target=\"_blank\" rel=\"noopener\">Mistral\u003C\u002Fa>—emit inaccurate or fabricated content without guardrails and process controls.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>💡 \u003Cstrong>Systemic, not a bug\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Governance‑focused literature treats hallucination as inherent, to be \u003Cem>managed\u003C\u002Fem> with policy, oversight, and architecture, not eliminated.\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa> In autonomous or agentic setups, risk compounds: an agent tasked with “complete the research chapter” will fill gaps with invented bridges or quotes if rewarded for task completion rather than factual accuracy.\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>\u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>Guardrails for nonfiction: input, output, and governance\u003C\u002Fh2>\n\u003Cp>The standard guardrail model—input control, output moderation, governance—used for chatbots and assistants maps cleanly onto nonfiction editorial workflows.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fp>\n\u003Ch3>Input control: constrain what the model is allowed to do\u003C\u002Fh3>\n\u003Cp>Input control governs allowed prompt types.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>For nonfiction:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Allow\u003C\u002Fstrong>:\n\u003Cul>\n\u003Cli>“Rewrite this verified quote for clarity; keep attribution.”\u003C\u002Fli>\n\u003Cli>“Summarize these interview notes into three bullet points.”\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Block or warn\u003C\u002Fstrong>:\n\u003Cul>\n\u003Cli>“Invent a plausible quote by X supporting Y.”\u003C\u002Fli>\n\u003Cli>“Find three sources that say Z” if you cannot ground them in a reference corpus.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Because hallucinations often appear as made‑up references and quotes,\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa> limiting the model to transformations of verified material shrinks the attack surface.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>⚡ \u003Cstrong>Practical pattern\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Wrap model access in a thin SDK instead of direct API calls:\u003C\u002Fp>\n\u003Cpre>\u003Ccode class=\"language-python\">def rewrite_quote(quote_id: str, style: str) -> str:\n    quote = get_verified_quote(quote_id)\n    prompt = f\"Rewrite this quote in a {style} style, keep meaning and attribution.\\n\\n{quote.text}\"\n    return call_llm(prompt, system=\"You must not introduce new facts or quotes.\")\n\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>No endpoint exists for raw “generate quotes.”\u003C\u002Fp>\n\u003Ch3>Output moderation: scan drafts for suspicious content\u003C\u002Fh3>\n\u003Cp>Output moderation uses a secondary model or rules engine to inspect generated text and flag likely hallucinations.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>For nonfiction:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Use regex + NER to detect quotes and citations.\u003C\u002Fli>\n\u003Cli>Require every quote to link to a record in your notes or transcript system.\u003C\u002Fli>\n\u003Cli>Run retrieval over your corpus for each quote; if no fuzzy match exists, flag as high‑risk.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Treat this like CI for manuscripts: fail the build if any quote lacks provenance.\u003C\u002Fp>\n\u003Cp>💼 \u003Cstrong>Callout: hallucination signals\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Specific dates, venues, or titles not present in your corpus.\u003C\u002Fli>\n\u003Cli>Many direct quotes from figures with sparse source material.\u003C\u002Fli>\n\u003Cli>Sudden appearance of perfectly formatted references late in the workflow.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3>Governance: process, policy, and accountability\u003C\u002Fh3>\n\u003Cp>Governance provides policies and oversight.\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>For nonfiction, define that:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Every quote carries provenance metadata (source, location, verification status).\u003C\u002Fli>\n\u003Cli>Any “AI‑assisted research” section requires sign‑off from a designated editor.\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>For regulated markets, you maintain documentation of hallucination‑mitigation measures aligned with AI Act expectations on risk management and transparency.\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Agentic AI in content workflows must be explicitly constrained:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Agents cannot inject new factual claims or quotes without verifiable sources or explicit “model‑generated” labels.\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>\u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>As OS‑level platforms like \u003Ca href=\"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FUbuntu\" class=\"wiki-link\" target=\"_blank\" rel=\"noopener\">Ubuntu\u003C\u002Fa> prompt users about enabling AI features, editorial stacks should similarly force explicit choices on where AI is allowed in the process.\u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Chr>\n\u003Ch2>Architectures to prevent fabricated quotes: \u003Ca href=\"\u002Fentities\u002F69d15a4e4eea09eba3dfe1b0-rag\">RAG\u003C\u002Fa>, tools, and verification loops\u003C\u002Fh2>\n\u003Cp>Guardrails specify \u003Cem>what\u003C\u002Fem> should happen; architecture handles \u003Cem>how\u003C\u002Fem> it happens.\u003C\u002Fp>\n\u003Ch3>Grounding with RAG\u003C\u002Fh3>\n\u003Cp>For nonfiction, a Retrieval‑Augmented Generation (RAG) pipeline should be standard:\u003C\u002Fp>\n\u003Col>\n\u003Cli>Index verified sources: transcripts, emails, books, reports.\u003C\u002Fli>\n\u003Cli>For any request that might produce a quote, retrieve candidate passages.\u003C\u002Fli>\n\u003Cli>Constrain generation: the model may paraphrase but not invent text unsupported by retrieved snippets.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>💡 \u003Cstrong>Key rule\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cblockquote>\n\u003Cp>“No source, no quote.”\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Cp>If retrieval returns nothing relevant, the model must refuse or ask for more data, not fabricate.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>Because hallucinated citations often look highly plausible,\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa> your RAG stack needs a verification pass:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Extract every quote and citation from the output.\u003C\u002Fli>\n\u003Cli>Match against your index (exact + fuzzy search).\u003C\u002Fli>\n\u003Cli>Flag low‑similarity items for human review.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3>Agentic verification loops\u003C\u002Fh3>\n\u003Cp>Agentic AI patterns can enforce verification instead of bypassing it.\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>\u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>Design a narrow verification agent:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Input\u003C\u002Fstrong>: manuscript section.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Tools\u003C\u002Fstrong>:\n\u003Cul>\n\u003Cli>\u003Ccode>search_corpus(query) -> list[passages]\u003C\u002Fcode>\u003C\u002Fli>\n\u003Cli>\u003Ccode>search_web(query) -> list[urls]\u003C\u002Fcode>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Behavior\u003C\u002Fstrong>:\n\u003Cul>\n\u003Cli>For each quote or factual claim, seek a supporting passage.\u003C\u002Fli>\n\u003Cli>Label each as \u003Ccode>verified\u003C\u002Fcode>, \u003Ccode>partially_verified\u003C\u002Fcode>, or \u003Ccode>unverified\u003C\u002Fcode>.\u003C\u002Fli>\n\u003Cli>Refuse “publishable” status if any high‑impact claim is \u003Ccode>unverified\u003C\u002Fcode>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Regulatory commentary on the AI Act stresses explainability and documentation: such agents should leave an auditable trail showing which tools they used and why a quote was accepted or rejected.\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>📊 \u003Cstrong>Infrastructure choices\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>With OS‑level inference endpoints like Ubuntu’s OpenAI‑compatible local APIs, you can:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Run drafting and verification locally.\u003C\u002Fli>\n\u003Cli>Keep sensitive interviews and drafts on your own infrastructure.\u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>This is crucial for investigative journalism and regulated industries with confidential sources.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>Trust, compliance, and sovereign AI: why provenance matters beyond books\u003C\u002Fh2>\n\u003Cp>This incident is not just about one book. It illustrates how organizations can build or erode public trust when relying on AI in knowledge work.\u003C\u002Fp>\n\u003Cp>Enterprise AI guidance insists that progress must be tied to clear principles and ethical discipline, especially where systems shape public understanding.\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa> Nonfiction, long‑form reports, and policy documents fall squarely here.\u003C\u002Fp>\n\u003Cp>⚠️ \u003Cstrong>Regulatory pressure is rising\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>The EU AI Act treats hallucinations, bias, and opacity as concrete risks that can trigger obligations, penalties, and documentation requirements.\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa> Using AI‑generated factual content in products or public communication without controls is becoming a compliance issue, not just a PR risk.\u003C\u002Fp>\n\u003Cp>Hallucinations and misinformation can:\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Undermine trust among users and stakeholders.\u003C\u002Fli>\n\u003Cli>Create legal exposure when decisions rely on false information.\u003C\u002Fli>\n\u003Cli>Force costly retractions and public incident responses.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>“Sovereign” and “trusted” AI narratives therefore stress knowing where models run, what data they use, and how outputs are controlled and audited.\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa> The same mindset should govern AI in research, whitepapers, analyst notes, and nonfiction writing.\u003C\u002Fp>\n\u003Cp>💼 \u003Cstrong>Multi‑agent editorial model\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Agentic AI guidance recommends orchestrating specialized agents under clear constraints and oversight.\u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa> For publishing, define:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Research agent\u003C\u002Fstrong> – retrieves and clusters documents; cannot generate quotes.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Summarization agent\u003C\u002Fstrong> – compresses source material; must attach citations.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Verification agent\u003C\u002Fstrong> – cross‑checks claims and blocks unsupported content.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Human editors remain accountable, but the system adds guardrails and logs.\u003C\u002Fp>\n\u003Cp>As general‑purpose platforms like Ubuntu embed AI into common tools, organizations need consistent provenance policies:\u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>When AI assistance is allowed.\u003C\u002Fli>\n\u003Cli>How it must be disclosed.\u003C\u002Fli>\n\u003Cli>What extra checks are mandatory before AI‑touched content goes public.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Chr>\n\u003Ch2>Designing production‑grade editorial pipelines with AI and humans in the loop\u003C\u002Fh2>\n\u003Cp>Assume hallucinations \u003Cem>will\u003C\u002Fem> occur. Then design editorial workflows like production systems.\u003C\u002Fp>\n\u003Ch3>1. Inventory and classification\u003C\u002Fh3>\n\u003Cp>Webinar guidance on AI compliance notes many organizations already use AI without grasping regulatory implications.\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>Steps:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Inventory every touchpoint where LLMs touch content:\n\u003Cul>\n\u003Cli>Research: search, summarization.\u003C\u002Fli>\n\u003Cli>Drafting: co‑writing, outlining.\u003C\u002Fli>\n\u003Cli>Editing: style, tone, translation.\u003C\u002Fli>\n\u003Cli>Review: fact‑checking, QA.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>Classify each use case by risk: internal notes vs public reports vs legally sensitive docs.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3>2. Policy + technical enforcement\u003C\u002Fh3>\n\u003Cp>Best practices emphasize policies, monitoring, and feedback loops.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa> Policies alone are insufficient; enforce technically where possible:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Restrict “high‑risk” content to tools that log prompts, models, and outputs.\u003C\u002Fli>\n\u003Cli>Block direct paste from general‑purpose chatbots into authoring tools for certain document types.\u003C\u002Fli>\n\u003Cli>Require internal flags on AI‑assisted sections for reviewers.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>💡 \u003Cstrong>Minimal telemetry to log\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Ccode>model_id\u003C\u002Fcode>, \u003Ccode>version\u003C\u002Fcode>\u003C\u002Fli>\n\u003Cli>\u003Ccode>prompt_template\u003C\u002Fcode> and task type\u003C\u002Fli>\n\u003Cli>\u003Ccode>source_documents\u003C\u002Fcode> used (IDs)\u003C\u002Fli>\n\u003Cli>\u003Ccode>verification_status\u003C\u002Fcode> of claims and quotes\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3>3. Automated checks before human review\u003C\u002Fh3>\n\u003Cp>Hallucinated citations slip through when reviewers trust surface plausibility.\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa> Shift initial checks to the system:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>CI step: run quote and citation verification on each merge request or manuscript milestone.\u003C\u002Fli>\n\u003Cli>Dashboard: show editors “unverified” sections requiring deeper review.\u003C\u002Fli>\n\u003Cli>Metrics: track hallucination incidents, false positives, and time to resolve.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Agentic AI guidance holds that multi‑step agents need defined goals and constraints, with humans validating high‑impact outcomes.\u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa> In editorial work, agents may propose, cluster, and verify—but not approve publication.\u003C\u002Fp>\n\u003Cp>📊 \u003Cstrong>On‑prem and local models\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Embedding local LLMs into developer and content environments, as OS vendors begin to do, enables:\u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Keeping drafts and corpora on your own infrastructure.\u003C\u002Fli>\n\u003Cli>Rich logging and audit without third‑party data sharing.\u003C\u002Fli>\n\u003Cli>Consistent policy enforcement via a shared local API.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Chr>\n\u003Ch2>Conclusion: treat AI‑assisted facts like production systems, not writing toys\u003C\u002Fh2>\n\u003Cp>AI‑fabricated quotes in a nonfiction book show how easily hallucinations can cross from model output into public record when guardrails, provenance, and verification are absent.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa> The failure pattern matches known hallucination behaviors, enterprise warnings about misinformation and trust erosion, and growing regulatory expectations for responsible AI in decision‑impacting contexts.\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>For ML and platform engineers, the takeaway is direct: any AI‑assisted factual content is a production system. It needs grounding architectures like RAG, strict input constraints, agentic verification, audit trails, and human oversight—not just clever prompting.\u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>Before you ship another AI‑assisted article, report, or product surface:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Map your workflow against the guardrail, governance, and verification patterns outlined here.\u003C\u002Fli>\n\u003Cli>Identify every step where a model could invent a quote or citation without detection.\u003C\u002Fli>\n\u003Cli>Add concrete controls—retrieval, cross‑checks, logging, human review—to close those gaps and avoid becoming the next cautionary tale.\u003C\u002Fli>\n\u003C\u002Ful>\n","An author publishing AI‑fabricated quotes in a nonfiction book is not a quirky misuse of ChatGPT. It is a production incident.\n\nYou have:\n\n- A generative model that invents sources.\n- An operator who...","hallucinations",[],1974,10,"2026-05-23T19:15:20.413Z",[17,22,26,30,34,38,42],{"title":18,"url":19,"summary":20,"type":21},"Des garde-fous en IA générative : prévenir les hallucinations et les émissions toxiques","https:\u002F\u002Ffr.linkedin.com\u002Fpulse\u002Fguardrails-generative-ai-preventing-hallucinations-toxic-amit-kharche-amv2f?tl=fr","Introduction : Pourquoi les garde-fous définissent l’avenir de l’IA\n\nL’IA générative transforme les entreprises. De l’automatisation du contenu marketing à l’assistance aux développeurs de logiciels e...","kb",{"title":23,"url":24,"summary":25,"type":21},"INTELLIGENCE ARTIFICIELLE D’ENTREPRISE:  DÉVELOPPER UNE IA DE CONFIANCE DANS LE CLOUD SOUVERAIN","https:\u002F\u002Fwww.opentext.com\u002Ffr\u002Fmedia\u002Febook\u002Fenterprise-artificial-intelligence-building-trusted-ai-with-secure-data-ebook-fr.pdf","INTELLIGENCE ARTIFICIELLE D’ENTREPRISE:  DÉVELOPPER UNE IA DE CONFIANCE DANS LE CLOUD SOUVERAIN\n\n# INTELLIGENCE ARTIFICIELLE D’ENTREPRISE \n\n# Développer une IA de confiance dans le cloud souverain \n\nP...",{"title":27,"url":28,"summary":29,"type":21},"Canonical va foutre de l'IA partout dans Ubuntu","https:\u002F\u002Fkorben.info\u002Fubuntu-ia-canonical-roadmap-2026.html","Canonical va foutre de l'IA partout dans Ubuntu\n\n27 avril 2026 – Par Korben\n\nCe qu’il faut retenir\n1) Canonical intègre l'IA partout dans Ubuntu via des Inference Snaps (modèles locaux pré-optimisés c...",{"title":31,"url":32,"summary":33,"type":21},"Qu'est-ce que l'Agentic AI ?","https:\u002F\u002Fwww.trendmicro.com\u002Ffr_fr\u002Fwhat-is\u002Fai\u002Fagentic-ai.html","Qu'est-ce que l'Agentic AI ? par Fernando Cardoso\nDernière mise à jour Mar 27, 2026\n\nL’IA agentique est une forme avancée d’intelligence artificielle (IA) qui utilise des « agents » d’IA autonomes pou...",{"title":35,"url":36,"summary":37,"type":21},"Comment utiliser l'IA agentique dans les entreprises en 2026: un plan","https:\u002F\u002Fwww.truefoundry.com\u002Ffr\u002Fblog\u002Fagentic-ai-in-enterprise","Par Abhishek Choudhary\nMis à jour: June 10, 2025\n\nL'intelligence artificielle (IA) a évolué, passant de l'automatisation de tâches simples à la mise en place de processus décisionnels complexes. Au cœ...",{"title":39,"url":40,"summary":41,"type":21},"Comment utiliser l'Intelligence artificielle en conformité avec l'AI Act ?","https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=1QvpL7Dp9pI","Description\nComment utiliser l'Intelligence artificielle en conformité avec l'AI Act ?\n\nBonjour à toutes et à tous ! 🚀 Plongez avec nous dans ce webinaire intelligence artificielle captivant où nous ...",{"title":43,"url":44,"summary":45,"type":21},"Hallucinations de l'IA en 2026 : où en est-on vraiment ?","https:\u002F\u002Fvocaneo.com\u002Farticles\u002Fhallucinations-ia-mars-2026","William\n\nPublié le 11\u002F03\u002F2026\n\nDans Études & statistiques\n\nQuand l'IA invente avec aplomb (et que personne ne s'en rend compte)\n\nVous demandez à ChatGPT de vous citer les trois études les plus récente...",{"totalSources":47},7,{"generationDuration":49,"kbQueriesCount":47,"confidenceScore":50,"sourcesCount":47},145918,100,{"metaTitle":52,"metaDescription":53},"AI Fabricated Quotes: Engineering Fixes for Hallucinations","AI‑fabricated quotes in nonfiction are production failures: this postmortem links AI hallucinations to workflow gaps and lists practical fixes to stop them.","en","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1583443920098-6b56d6aabdb1?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxfHxub25maWN0aW9uJTIwbGllcyUyMGVuZ2luZWVyaW5nJTIwbGVzc29uc3xlbnwxfDB8fHwxNzc5NTcyNTcwfDA&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60",{"photographerName":57,"photographerUrl":58,"unsplashUrl":59},"Pawel Czerwinski","https:\u002F\u002Funsplash.com\u002F@pawel_czerwinski?utm_source=coreprose&utm_medium=referral","https:\u002F\u002Funsplash.com\u002Fphotos\u002Fi-am-a-good-man-i-am-a-good-man-3DT_d9ttbv0?utm_source=coreprose&utm_medium=referral",false,null,{"key":63,"name":64,"nameEn":64},"ai-engineering","AI Engineering & LLM Ops",[66,68,70,72],{"text":67},"AI‑fabricated quotes in nonfiction are a production incident following a 5‑step failure pattern (T‑0 through T‑4) where a generative model invents sources, an operator treats outputs as ground truth, and no verification loop exists.",{"text":69},"Hallucinations are inherent to next‑token LLMs: models will produce authoritative‑sounding but unsupported quotes and citations unless constrained by retrieval and provenance; “No source, no quote” must be enforced.",{"text":71},"Preventing these incidents requires a RAG grounding layer, an automated verification pass that fails the build on unverified quotes, and minimal telemetry (model_id, prompt_template, source_document_ids, verification_status).",{"text":73},"Editorial pipelines must treat AI‑assisted factual content as production systems with input control, output moderation, agentic verification, auditable logs, and human sign‑off for high‑risk releases.",[75,78,81],{"question":76,"answer":77},"How did AI‑fabricated quotes make it into a published nonfiction book?","The fabrication happened because a generative model produced plausible but unsupported quotes, a human operator accepted them as factual, and the editorial process lacked provenance checks and verification. In practice this follows a T‑0 to T‑4 timeline: an author uses an LLM for research; the model invents a quote; the author inserts it unlabeled; editorial review focuses on style not sourcing; the book is published and the fabrication is discovered. This incident illustrates systemic gaps: no retrieval‑augmented generation (RAG) to ground claims, no automated scan to match quotes against transcripts or indexed sources, and no required metadata or verification sign‑off that would have caused the manuscript CI to fail before publication.",{"question":79,"answer":80},"What technical controls prevent fabricated quotes in publishing workflows?","Implement RAG so every quote request retrieves candidate passages from an indexed corpus, require the model to paraphrase only retrieved text, and build an automated verification pass that extracts quotes and citations and performs exact + fuzzy matching against the index. Add an SDK that blocks raw “generate quote” endpoints, use regex\u002FNER to detect quotes, fail CI when provenance is missing, and log minimal telemetry (model_id, prompt_template, source_document_ids, verification_status) for audits.",{"question":82,"answer":83},"What governance and policy measures should organizations adopt?","Organizations must require provenance metadata for every quote, mandate editor sign‑off for AI‑assisted research, classify use cases by risk, and enforce technical constraints for high‑risk documents. They should maintain auditable trails for agentic verification agents, document mitigation measures for regulators (e.g., AI Act expectations), and ensure humans remain the final approver for any publishable factual claim.",[85,91,98,103,110,115,121,127,131,135,140,146,152,157,162],{"id":86,"name":87,"type":88,"confidence":89,"wikipediaUrl":61,"slug":90,"mentionCount":47},"6a0e39b007a4fdbfcf5ea778","Agentic AI","concept",0.98,"6a0e39b007a4fdbfcf5ea778-agentic-ai",{"id":92,"name":93,"type":88,"confidence":94,"wikipediaUrl":95,"slug":96,"mentionCount":97},"69d15a4e4eea09eba3dfe1b0","RAG",0.97,"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FRag","69d15a4e4eea09eba3dfe1b0-rag",5,{"id":99,"name":100,"type":88,"confidence":89,"wikipediaUrl":61,"slug":101,"mentionCount":102},"6a0b8ac41f0b27c1f426f70c","LLMs","6a0b8ac41f0b27c1f426f70c-llms",4,{"id":104,"name":105,"type":88,"confidence":106,"wikipediaUrl":107,"slug":108,"mentionCount":109},"6a11fc89a2d594d36d2240c5","hallucination",0.99,"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FHallucination","6a11fc89a2d594d36d2240c5-hallucination",2,{"id":111,"name":112,"type":88,"confidence":113,"wikipediaUrl":61,"slug":114,"mentionCount":109},"69d08f184eea09eba3dfd050","guardrails",0.96,"69d08f184eea09eba3dfd050-guardrails",{"id":116,"name":117,"type":88,"confidence":118,"wikipediaUrl":119,"slug":120,"mentionCount":109},"69d15a504eea09eba3dfe1bc","provenance",0.95,"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FProvenance","69d15a504eea09eba3dfe1bc-provenance",{"id":122,"name":123,"type":88,"confidence":124,"wikipediaUrl":61,"slug":125,"mentionCount":126},"6a11fcfaa2d594d36d2240e4","editorial workflows",0.92,"6a11fcfaa2d594d36d2240e4-editorial-workflows",1,{"id":128,"name":129,"type":88,"confidence":106,"wikipediaUrl":61,"slug":130,"mentionCount":126},"6a11fcf9a2d594d36d2240e1","AI‑fabricated quotes","6a11fcf9a2d594d36d2240e1-ai-fabricated-quotes",{"id":132,"name":133,"type":88,"confidence":106,"wikipediaUrl":61,"slug":134,"mentionCount":126},"6a11fcfaa2d594d36d2240e2","fabricated citations","6a11fcfaa2d594d36d2240e2-fabricated-citations",{"id":136,"name":137,"type":138,"confidence":118,"wikipediaUrl":61,"slug":139,"mentionCount":126},"6a11fcf9a2d594d36d2240df","The Future of Truth","other","6a11fcf9a2d594d36d2240df-the-future-of-truth",{"id":141,"name":142,"type":143,"confidence":118,"wikipediaUrl":144,"slug":145,"mentionCount":102},"6a0a74001f0b27c1f426a613","Claude","product","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FClaude","6a0a74001f0b27c1f426a613-claude",{"id":147,"name":148,"type":143,"confidence":118,"wikipediaUrl":149,"slug":150,"mentionCount":151},"6a0a74001f0b27c1f426a610","Llama","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FLlama","6a0a74001f0b27c1f426a610-llama",3,{"id":153,"name":154,"type":143,"confidence":118,"wikipediaUrl":155,"slug":156,"mentionCount":151},"6a0e316d07a4fdbfcf5ea647","ChatGPT","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FChatGPT","6a0e316d07a4fdbfcf5ea647-chatgpt",{"id":158,"name":159,"type":143,"confidence":118,"wikipediaUrl":160,"slug":161,"mentionCount":109},"6a11fc89a2d594d36d2240c7","Mistral","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FMistral_(missile)","6a11fc89a2d594d36d2240c7-mistral",{"id":163,"name":164,"type":143,"confidence":165,"wikipediaUrl":61,"slug":166,"mentionCount":126},"6a11fcf9a2d594d36d2240e0","GPT‑class",0.9,"6a11fcf9a2d594d36d2240e0-gpt-class",[168,175,183,190],{"id":169,"title":170,"slug":171,"excerpt":172,"category":11,"featuredImage":173,"publishedAt":174},"6a11fa635242169466949187","When AI Invents Sources: What the ‘Future of Truth’ Quote Scandal Teaches Us About LLM Hallucinations and Editorial Guardrails","when-ai-invents-sources-what-the-future-of-truth-quote-scandal-teaches-us-about-llm-hallucinations-and-editorial-guardrails","A nonfiction author publishing AI‑fabricated quotes is not just a publishing disaster; it is a failure of system design around truth. The core problem was not only that a model hallucinated, but that...","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1638866412987-e4663ec0ab8a?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxfHxpbnZlbnRzJTIwc291cmNlcyUyMGZ1dHVyZSUyMHRydXRofGVufDF8MHx8fDE3Nzk1NzI1NzN8MA&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-05-23T19:13:40.594Z",{"id":176,"title":177,"slug":178,"excerpt":179,"category":180,"featuredImage":181,"publishedAt":182},"6a10f51f5242169466943e82","How Commercial AI Models Are Scaling and Automating Cyber Attacks","how-commercial-ai-models-are-scaling-and-automating-cyber-attacks","1. From Human-Driven Hacking to AI-Scaled Campaigns  \n\nCyberattacks are shifting from human operators to automated, AI-orchestrated campaigns spanning the full kill chain.[2] ML and generative models...","trend-radar","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1701777892936-84cfa8a67cdd?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxfHx1c2UlMjBjb21tZXJjaWFsfGVufDF8MHx8fDE3Nzk0OTYyMjN8MA&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-05-23T00:37:10.739Z",{"id":184,"title":185,"slug":186,"excerpt":187,"category":11,"featuredImage":188,"publishedAt":189},"6a0fd62b035a091ce258268e","Linus Torvalds vs AI Bug Hunters: How to Stop Duplicate Linux Vulnerability Reports from Overwhelming Security Teams","linus-torvalds-vs-ai-bug-hunters-how-to-stop-duplicate-linux-vulnerability-reports-from-overwhelming-security-teams","AI-powered vulnerability scanners are now good enough to find serious Linux bugs at scale—but that success risks turning into a denial-of-service attack on security teams’ attention.\n\nLinus Torvalds h...","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1652174834052-119f4d8f8448?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxfHxsaW51cyUyMHRvcnZhbGRzfGVufDF8MHx8fDE3Nzk0NDIzMTl8MA&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-05-22T04:11:38.998Z",{"id":191,"title":192,"slug":193,"excerpt":194,"category":11,"featuredImage":195,"publishedAt":196},"6a0f81bf035a091ce25801a8","AI-Enabled Cyber Attacks Hit 600+ Firewalls: The 9 Autonomous Breaches That Redefined Security in 2026","ai-enabled-cyber-attacks-hit-600-firewalls-the-9-autonomous-breaches-that-redefined-security-in-2026","In Q1 2026, nine coordinated intrusion campaigns crossed more than 600 enterprise firewalls before defenders realized the “operator” was a mesh of large‑language‑model (LLM)–driven agents executing fu...","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1614064641938-3bbee52942c7?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxfHxlbmFibGVkJTIwY3liZXIlMjBhdHRhY2tzJTIwaGl0fGVufDF8MHx8fDE3Nzk0MjE4NjB8MA&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-05-21T22:10:25.898Z",["Island",198],{"key":199,"params":200,"result":202},"ArticleBody_xQBPRRgrM9neW8OWI6OOOy1e0oAK9CqvSC1sEjhjM",{"props":201},"{\"articleId\":\"6a11fbf252421694669491e9\",\"linkColor\":\"red\"}",{"head":203},{}]