AI-Assisted Content Workflow

Originally published at thatdevpro.com. This framework reference is part of the 14-tier Engine Optimization stack from ThatDevPro, an SDVOSB-certified veteran-owned web + AI engineering studio. You are reading the dev.to mirror; the source-of-truth canonical version with embedded validation tools lives at the link above.

The Positive Playbook for Using AI Correctly in Editorial Content Production

A comprehensive installation and audit reference for the AI-assisted content workflow discipline. AI is in the editorial stack whether the agency endorses it or not (over 80 percent of marketers use AI for content creation per Content Marketing Institute and SQ Magazine 2025 surveys; 79 percent report quality lift per ColorWhistle 2025; 7 percent of people globally and 15 percent of under-25s use AI chatbots for news weekly per Reuters Institute 2025). The question is which tasks AI is suited to, which it is not, what review threshold separates AI-assisted human content from thin AI output with a human nameplate, and how that threshold integrates with Google's HCS, YMYL standards, SQRG, and the 2025-2026 regulatory layer. Dual purpose: installation manual and audit document.

Cross stack implementation note: the code samples in this framework are written in plain HTML for clarity. For React, Vue, Svelte, Next.js, Nuxt, SvelteKit, Astro, Hugo, 11ty, Remix, WordPress, Shopify, and Webflow equivalents of every pattern below, see framework-cross-stack-implementation.md. For pure client rendered SPAs (no SSR/SSG) see framework-react.md. For Tailwind specific concerns see framework-tailwind.md.

1. Document Purpose and How to Use This Document

1.1 What This Document Is

The canonical operational reference for AI-assisted content production. Three sibling frameworks set the negative constraints: framework-hcs.md on mass-produced AI failing the Helpful Content System (the December 2025 core update hit thin affiliate sites at 71 percent traffic loss and YMYL sites at 67 percent per ALM Corp's December 2025 rollout analysis); framework-ymyl.md on credentialed reviewer sign-off for YMYL AI content; framework-sqrg.md on the September 2025 SQRG AI evaluation criteria. This document is the positive playbook those negatives need.

The position: AI is a class of editorial assistance whose appropriate use depends on task, topic, reviewer credentials, and disclosure posture. Some tasks are unambiguously suited to AI (transcription, brainstorming, summarizing the writer's own notes, grammar editing). Some are unambiguously hostile (bulk drafting outside the entity's expertise, fabricating quotes or statistics, masking AI as human-authored). Most fall between and require a workflow that brings human expertise back into the loop at the right stages.

1.2 Three Operating Modes

Mode A, Install Mode. Establish AI-assisted workflow infrastructure. Sections 2 through 14 in order.

Mode B, Audit Mode. Evaluate an existing AI practice. Skip to Section 13.

Mode C, Hybrid Mode. Audit first, then install for failing items.

1.3 How Claude Code CLI Should Consume This Document

Section 2 variables; Section 3 spectrum; Section 4 acceptable tasks; Section 5 unacceptable tasks; Section 6 stages and gates; Section 7 threshold; Section 8 citation verification; Section 9 disclosure; Section 10 YMYL; Section 11 detectors; Section 12 anti-patterns; Section 13 audit; Section 14 maintenance and reports.

1.4 Conflict Resolution Rules

1.5 Required Tools and Validators

GSC and GA4 for performance baseline; Originality.ai or comparable detector (Section 11 accuracy caveats); Copyscape or comparable plagiarism checker; primary source databases per topic (PubMed for medical, SEC EDGAR for financial, CourtListener for legal, Google Scholar for academic, government source registries for civic); editorial workflow tool (Notion, Airtable) for stage tracking; published Editorial Policy and AI Use Policy pages (see framework-eeat.md Section 5.2 for editorial policy; Section 9 here for AI policy).

1.6 Four Pillars Context

The workflow operates across all four pillars. SEO: December 2025 core update extended E-E-A-T weighting beyond YMYL to "practically all competitive searches" per ALM Corp December 2025. AEO: featured snippet eligibility correlates with byline credibility and citation quality. AIO: Google AI Overviews preferentially cite sources with verifiable expertise and current update history per framework-aioverviews.md. GEO: ChatGPT, Claude, Perplexity, Gemini, Bing Copilot citation logic discounts publishers with detectable AI bulk patterns per framework-aicitations.md. Meeting this framework's standards earns citations across all four; shipping thin AI bulk loses them across all four.

1.7 Relationship to Neighboring Frameworks

Operational standard for the workflow. General quality bar: framework-hcs.md. YMYL elevated bar: framework-ymyl.md Section 5.8. SQRG perspective: framework-sqrg.md Section 9. Originality AI cannot produce alone: framework-infogain.md. Credentialing for bylined responsibility: framework-eeat.md Section 4. Two Phase 2 siblings (brand voice in AI production; content brief structure for AI handoff) scheduled.

2. Client Variables Intake

# AI-ASSISTED CONTENT WORKFLOW CLIENT VARIABLES

# --- Business and Site Identity (REQUIRED) ---
business_name: ""
primary_domain: ""
business_industry: ""
ymyl_classification: ""              # full_ymyl, partial_ymyl, lite_ymyl, non_ymyl
cms_or_stack: ""
total_indexable_pages: 0

# --- Current AI Use Posture (REQUIRED, answer truthfully) ---
ai_currently_in_use: false
ai_tools_used: []                    # ChatGPT, Claude, Gemini, Copilot, Jasper, Writesonic, custom
ai_use_policy_page_exists: false
ai_use_policy_url: ""
ai_use_disclosed_in_articles: false

# --- AI Task Distribution (REQUIRED) ---
ai_used_for_research: false
ai_used_for_brainstorming: false
ai_used_for_outlining: false
ai_used_for_first_draft: false
ai_used_for_full_draft: false        # Honest answer; the bar gets higher with each task
ai_used_for_grammar_editing: false
ai_used_for_summarization: false
ai_used_for_transcription: false
ai_used_for_bulk_publication: false  # The red line. See Section 5.
ai_used_for_translation: false

# --- Volume and Review Capacity (REQUIRED) ---
articles_published_per_week: 0
percentage_ai_assisted: 0
percentage_fully_ai_generated: 0
in_house_editors_count: 0
hours_per_article_for_review_avg: 0
named_reviewers_with_credentials: []  # For YMYL, list reviewers by name and credential
review_quality_documented: ""         # "minimal", "moderate", "rigorous", "expert_review_required"
review_signed_off_per_article: false

# --- Disclosure and Verification (REQUIRED) ---
articles_with_ai_disclosure_callout: 0
percentage_disclosed: 0
citations_verified_against_primary_sources: false
fact_checking_process_documented: false
hallucinated_citation_incidents_known: 0
ai_detector_in_workflow: false        # Different from plagiarism; see Section 11

# --- YMYL-Specific (REQUIRED if any YMYL content) ---
ymyl_ai_content_volume_per_week: 0
credentialed_reviewer_signoff_documented: false
medical_legal_financial_disclaimer_present: false

# --- Regulatory Posture (REQUIRED) ---
serves_eu_users: false                # AI Act Article 50
serves_california_users: false        # CAITA, AB 853
ftc_endorsement_compliant: false      # If reviews, testimonials present

# --- Current Outcomes (REQUIRED baseline) ---
december_2025_core_update_impact: ""  # "positive", "neutral", "negative", "unknown"
percentage_traffic_change_post_dec_2025: 0
ai_overview_citations_count_current: 0
hcs_audit_score_baseline: 0

Save as ai-workflow-variables.yml. Routing: if ymyl_classification is not non_ymyl and credentialed signoff is missing, route to framework-ymyl.md Section 5.8 first. If december_2025_core_update_impact is negative, route to framework-coreupdates.md recovery in parallel.

3. What AI-Assisted Content Workflow Is

3.1 The Spectrum Framing

AI involvement is not binary; it is a spectrum from zero percent (pure human writing) through one hundred percent (model output published verbatim). Most production sits between. Seven discrete bands:

The editorial baseline should sit in the 10 to 50 percent band. Beyond 50 percent the bylined responsibility test (Section 7.3) gets harder to pass. Beyond 75 percent it cannot be passed honestly.

3.2 The December 2025 Core Update Lessons

The December 2025 core update (December 11 to 29, 2025; third of 2025) was the most consequential AI-content-related update in Google's history. Three measurable lessons:

First, mass-produced AI without expert oversight saw approximately 87 percent negative traffic impact (ALM Corp December 2025 rollout analysis, 4,200 sites). Median decline among affected approximately 40 percent; lower quartile 60 to 80 percent for content farms. Thin affiliate sites lost approximately 71 percent. YMYL sites approximately 67 percent.

Second, E-E-A-T weighting was extended to "practically all competitive searches" including e-commerce reviews, SaaS comparisons, and how-to guides. The credentialed reviewer pattern that was YMYL-only now applies to commercial-intent content.

Third, recovery is 6 to 12 months of demonstrated quality improvement (gsqi.com analysis). The cost of getting AI workflow wrong compounds across multiple update cycles.

3.3 The September 2025 SQRG AI Guidance

Key principles raters now apply (September 2025 SQRG, summarized via Originality.ai 2025 and Search Engine Land September 2025):

• AI involvement is not automatically Lowest. Quality of output matters more than method.
• "If all or nearly all of the main content is auto- or AI-generated (with little or no added value), raters should apply the lowest rating" (direct SQRG language).
• "Scaled Content Abuse" names creating "a lot of content with little effort or originality with no editing or manual curation"; generative AI is the primary example.
• "Filler Content" covers "low-effort, low-relevance content that may visually dominate a page while failing to support its purpose."
• YMYL expanded to elections, civic institutions, and government trust on top of health, finance, legal, and safety.

The SQRG quality bar applies equally to human and AI content. AI just makes it cheap to mass-produce work that fails the bar.

3.4 The 2026 Regulatory Layer

EU AI Act Article 50 (transparency effective August 2026; first draft Code of Practice December 17, 2025; finalization May-June 2026): generative AI outputs must be machine-readable marked. Deployers using AI for text "published with the purpose of informing the public on matters of public interest" must disclose. Penalties up to 15 million euros or 3 percent of global annual turnover (Cooley December 2025; Kirkland and Ellis February 2026).

California AI Transparency Act (SB 942) effective January 1, 2026: covered providers must offer AI detection tools. AB 853 (October 13, 2025) extends to large platforms and distributors. AB 1008 (January 1, 2025) treats AI-generated data as personal information under CCPA (Pillsbury Law CAITA update; Mayer Brown October 2025).

FTC final rule on fake and AI-generated reviews (effective October 21, 2024): prohibits fake or AI-generated reviews, testimonials, celebrity endorsements. Penalties up to 51,744 dollars per violation. AI-generated influencer content requires disclosure when the entity is not a real person (Sidley Data Matters August 2024).

These three layers establish the floor. This workflow clears it by default.

4. Acceptable AI Tasks

AI augments human work without substituting for the substance that demands human expertise, lived experience, or credentialed responsibility.

4.1 Research Assistance

Acceptable when AI output is a starting point and the human verifies every factual claim against primary sources. Patterns: brainstorming search terms; summarizing long documents the human will read; surfacing counter-arguments; suggesting subject-matter experts to interview; identifying gaps for original reporting. Every published claim traces to a primary source the human verified. AI research that becomes published content without primary-source verification fails Section 8.

4.2 Brainstorming and Topic Ideation

Acceptable for topic generation, angle exploration, framing. The decision is the human's. The published topic must satisfy the topical-pillar test in framework-hcs.md Section 6.1. AI cannot generate expertise the entity does not have.

4.3 Outlining

Acceptable when the human writer fills the outline with original content. Outlines are scaffolding; the bar applies to what is built on the scaffolding. Honest test: would a competent editor reading the outline alone be unable to tell whether AI or human produced it? If yes, generic enough to be acceptable. If the outline already contains specific opinions, examples, or claims that the human will then justify after the fact, the outline is doing the editorial work the human should be doing.

4.4 First-Draft Generation with Heavy Human Revision

The most contested band. Acceptable when human revision is substantive: restructuring at least a third for argument or flow; adding original examples, case studies, anecdotes, or data the human knows but the AI did not surface; replacing AI hedge-and-generality phrasing with defensible claims; removing AI filler ("It is important to note", "Furthermore") and manufactured authority ("Studies have shown" without an actual study); verifying every citation, statistic, and quote against the primary source. The unrevised AI first draft is unacceptable even with honest disclosure because HCS, SQRG, and YMYL quality bars apply to the content itself.

4.5 Grammar and Style Editing

Acceptable on completed human drafts. The safest band. Risk: AI editing introduces subtle factual errors when "improving" technical writing. Mitigation: final human read-through with attention to numeric claims and named entities. For YMYL, the credentialed reviewer reads the post-edit version.

4.6 Summarization of the Writer's Own Notes

Acceptable for summarizing notes, transcripts, or source material the writer personally collected. The writer verifies every summary claim against source because AI summarization fabricates: the 2024 JMIR systematic review study found GPT-3.5 hallucinated 39.6 percent of references, GPT-4 28.6 percent, Bard 91.4 percent.

4.7 Transcription

Acceptable and standard (Otter, Whisper, Descript). Risk: mis-hears named entities and numbers. Mitigation: human verification of every named entity and number against the original audio. The byline belongs to the human interviewer, not the tool.

4.8 Translation Drafts

Acceptable when a native-speaker editor revises for accuracy, idiom, cultural nuance. AI translation alone is not acceptable because idiomatic and cultural errors are detectable. The published translation names a human translator or editor.

4.9 Internal Linking, Tagging, and Metadata Drafts

Acceptable for surfacing internal linking opportunities, tag suggestions, category placement, draft title tags, meta descriptions, OG tags. The human editor accepts, rejects, or revises each. Metadata must accurately describe the content per the honest title test in framework-hcs.md Section 5.1.4.

5. Unacceptable AI Tasks

5.1 Bulk Drafting on Topics Without Expertise. Prompting AI to draft articles on topics the publishing entity has no genuine expertise on, in volume, with minimal review, for search ranking. The December 2025 core update target; the September 2025 SQRG scaled-content-abuse pattern. Fix: stop publishing on topics outside the entity's expertise. Topic pillar discipline (framework-hcs.md Section 6.1) precedes workflow discipline.

5.2 Fake Bylines. Assigning a human byline to AI content the named human did not write or substantively revise. September 2025 SQRG explicitly cites this as a Lowest-rating signal. Section 7.3 is the threshold. Fix: name the actual reviewer, or do not byline.

5.3 Fabricated Quotes. AI-generated quotes attributed to real people the AI invented dialogue for. Authors Guild January 2025 hard line. FTC October 2024 fake-reviews rule treats fabricated testimonials as deceptive practice (51,744 dollars per violation). Fix: do not publish quotes not actually said. Paraphrased or composite quotes are labeled.

5.4 Fabricated Statistics. AI-generated statistics not traced to primary sources. Stanford 2024 Large Legal Fictions (Stanford HAI January 2024) tested GPT-4 at 58 percent hallucination on 800,000 legal queries; Stanford 2024 Hallucination-Free (Stanford HAI May 2024) tested purpose-built legal AI at 17-34 percent hallucination. Fix: every statistic has a hyperlink to its primary source.

5.5 Fabricated Citations. AI-generated citations to journals, cases, or studies that do not exist. JMIR 2024 (636 citations in 84 papers): 55 percent fabricated for GPT-3.5, 18 percent for GPT-4; errors in non-fabricated 43 percent and 24 percent. PMC 2024 mental health (176 GPT-4o citations): 19.9 percent fabricated, topic-variant from 6 to 29 percent. Section 8 catches these.

5.6 Mass Production. Publishing more articles per week than review can substantively handle. 40 editor-hours at 1-2 hours per article caps at 20-40 articles per week. Fix: match publication to capacity, or expand capacity, do not relax review.

5.7 AI Passed As Expert-Authored Without Disclosure. Publishing AI under a byline implying expert authorship without disclosing. Disclosure alone is insufficient without credentialed sign-off. Credentialed review alone is insufficient without disclosure. Both required.

5.8 YMYL Without Credentialed Reviewer. AI on health, finance, legal, civic, or safety topics without credentialed human reviewer. Section 10 and framework-ymyl.md Section 5.8. Reviewer name, credentials, license number, verification link appear on the article.

5.9 Bypassing Detection Tools. Running AI through humanizer tools (Undetectable AI, StealthGPT, Phrasly) to evade detection. Inconsistent with Google's quality-not-method framing, FTC deceptive-practices framework, EU AI Act Article 50, Authors Guild standards. When detected, AI-plus-deception is worse than honest disclosure. Fix: disclose and improve.

5.10 Translation Without Native-Speaker Review. AI translation shipped without a native-speaker reviewer. Idiom, cultural, and proper-noun errors survive ranking algorithms but fail readers. Disproportionate ranking declines in the December 2025 core update for non-English markets.

6. The Workflow Stages and Quality Gates

Six stages: Research, Outline, Draft, Edit, Review, Publish. Each has an AI permission level, a human responsibility, and a quality gate that must pass before the next stage begins.

6.1 Stage One, Research

AI permission: assistance only. AI may surface sources, summarize documents the human will read, suggest experts to interview, brainstorm angles. AI is not the only research source.

Human responsibility: identify primary sources, verify claims against them, conduct interviews and original observation, build the source-of-truth manifest the article will cite.

Quality gate: a written list of primary sources exists before drafting. Every claim traces to at least one source. AI-generated "research" not corroborated by primary source is excluded.

<!-- Research stage source manifest, internal to editorial system -->
<section class="research-manifest" data-stage="research">
  <h2>Source manifest for article {{ARTICLE_ID}}</h2>
  <ol class="primary-sources">
    <li>
      <cite>{{SOURCE_TITLE}}</cite>
      <a href="{{PRIMARY_SOURCE_URL}}">{{PRIMARY_SOURCE_URL}}</a>
      <span class="source-type">{{primary | academic | regulatory | interview | observation}}</span>
      <span class="date-accessed">{{YYYY-MM-DD}}</span>
    </li>
  </ol>
</section>

6.2 Stage Two, Outline

AI permission: AI may propose outline structures, section orderings, topics. The human editor selects, modifies, rejects.

Human responsibility: thesis, argument arc, scope decisions, original contribution (the Information Gain check from framework-infogain.md).

Quality gate: the outline includes (a) a named thesis, (b) at least one Information Gain contribution that distinguishes this article from the top SERP results, (c) a coverage promise matching the title per framework-hcs.md Section 5.1.4, (d) citation and disclosure sections.

6.3 Stage Three, Draft

AI permission: variable by section. Well-documented factual sections may use AI first-draft assistance up to the 50 percent band (Section 3.1). Sections expressing professional opinion, lived experience, or original analysis are at the 5 to 10 percent band. YMYL claims are at the 0 to 10 percent band; credentialed reviewer involvement begins here, not later.

Human responsibility: write the parts that require expertise, lived experience, or judgment. Restructure AI-drafted sections to match voice and argument. Replace AI's hedge-and-generality phrasing with defensible claims.

Quality gate: the draft passes the bylined responsibility test (Section 7.3). The named author would honestly defend every claim in front of a peer. If a section fails, it is rewritten or removed.

6.4 Stage Four, Edit

AI permission: high for grammar and punctuation. Moderate for sentence-level clarity. Low for structural changes (editor decisions, not tool decisions).

Human responsibility: line editing for clarity, structural editing for flow, fact-check pass against the Stage One source manifest.

Quality gate: every numeric claim, named entity, quote, and citation is verified against the source manifest. Items that do not trace to a primary source are traced back, softened, or removed. The Section 8 protocol runs here, not at publish.

6.5 Stage Five, Review

AI permission: zero for the substantive review decision. AI may surface issues (consistency, terminology, grammar drift) but the sign-off is human.

Human responsibility: non-YMYL, an editor other than the author approves. YMYL, a credentialed reviewer (medical, legal, financial, or topic-specific credential) verifies factual claims against professional knowledge and signs off. Reviewer name and date are recorded.

Quality gate: written sign-off exists. For YMYL, the credentialed reviewer's name, credentials, and license verification link are ready for the published article per Section 10.

6.6 Stage Six, Publish

AI permission: assistance for metadata drafts (titles, meta descriptions, OG tags) and schema drafts. The human confirms accuracy.

Human responsibility: final read on production (or staging mirroring production), schema validation, disclosure callout placement, internal link verification.

Quality gate: the published article includes the AI use disclosure (Section 9), the author byline with bio link, the reviewer byline for YMYL with credentials and license link, source citations as footnotes or in-text links, valid Article schema with dateModified and author fields, and a content changelog for substantive updates per framework-contentrefresh.md Section 8.

6.7 Stage Workflow Summary

7. Human-in-Loop Threshold

7.1 The Question

How much human modification of AI output makes content "human-authored" for byline, disclosure, and professional responsibility?

7.2 Proposed Thresholds in 2025 to 2026

Authors Guild Human Authored Certification (January 2025): 95 percent human-authored; AI text in a certified book may not exceed 5 percent (de minimis). Strictest publishing-industry bar; recommended floor for pure human-authorship claims.

Google Search Central guidance (February 2023, evolved 2024 and 2025): no percentage. "Appropriate use" of AI is acceptable; the quality bar applies regardless of method. A 10 percent revision adding substantive expertise is acceptable; a 50 percent revision adding only style polish is not.

September 2025 SQRG: no percentage. Substantive human review and verification rate High or Highest; absence rates Low or Lowest.

EU AI Act Article 50 (effective August 2026): addresses disclosure rather than percentage. Any AI-generated or AI-modified text published to inform the public on matters of public interest must be disclosed.

7.3 The Bylined Responsibility Test

The operational test this framework adopts. More useful than a percentage threshold because it applies regardless of how much AI was used.

The test: would the named author honestly defend every claim, example, recommendation, and citation in the article in front of a peer at their professional level, knowing those peers can verify the underlying sources?

If yes, the byline is defensible regardless of AI involvement. Disclosure (Section 9) acknowledges the AI involvement honestly. Both present, article passes. If no, the byline is fraudulent. Substantially revise until the test passes, change the byline to the person who passes, or do not publish.

The test scales across content types. A 30-second editor sign-off on an AI-drafted listicle does not pass for the named editor. A 3-hour substantive revision and verification pass on an AI-drafted technical article passes for the named reviewer even if 60 percent of surface text originated as AI. The test is about substantive contribution, not text overlap.

7.4 The Reviewer Byline Pattern

For AI-assisted content where the drafting "writer" is AI, the byline names the human reviewer who passes the bylined responsibility test, with explicit disclosure:

<header class="article-byline">
  <p class="reviewer">
    Reviewed and verified by <a href="/authors/{{REVIEWER_SLUG}}/" rel="author">{{REVIEWER_NAME}}</a>
  </p>
  <p class="credentials">{{REVIEWER_CREDENTIALS_RELEVANT_TO_TOPIC}}</p>
  <time datetime="{{REVIEW_DATE}}">Reviewed {{REVIEW_DATE_HUMAN}}</time>
  <p class="ai-disclosure-brief">
    This article was drafted with AI assistance and substantively revised,
    fact-checked, and approved by the named reviewer.
    <a href="/ai-policy/">See our AI use policy</a>.
  </p>
</header>

Honest about production method, names a single accountable human, links to the site-wide AI policy. Acceptable across the 25 to 75 percent bands of Section 3.1 when the reviewer's substantive contribution is real.

7.5 What the Threshold Is Not

Not text overlap. A 90 percent text-identical revision that re-orders argument, replaces every numeric claim with verified figures, and adds lived-experience case study is more substantively human-authored than a 50 percent text-different revision paraphrasing AI hedges into different AI-like hedges.

Not detector pass-fail. Originality.ai, GPTZero, and Copyleaks produce false negatives on substantively reviewed AI content and false positives on entirely human content (Section 11). Detector output delegates editorial judgment to a tool whose accuracy on edited AI is below useful bounds.

The threshold is the bylined responsibility test: qualitative human judgment, not automation.

8. Fact-Checking and Citation Verification

8.1 The Hallucination Reality

AI models fabricate citations and statistics at rates not within useful bounds for unverified publication:

• Stanford 2024 Large Legal Fictions (Stanford HAI January 2024, 800,000 legal queries): GPT-4 58 percent, GPT-3.5 69 percent, Llama 2 88 percent.
• Stanford 2024 Hallucination-Free on purpose-built legal AI (Stanford HAI May 2024): Lexis Plus AI and Ask Practical Law over 17 percent; Westlaw AI-Assisted Research over 34 percent.
• JMIR 2024 systematic review (J Med Internet Res 2024, 636 citations in 84 papers): 55 percent fabricated for GPT-3.5, 18 percent for GPT-4; errors in non-fabricated 43 percent and 24 percent.
• PMC mental health 2024 (176 GPT-4o citations): 19.9 percent fabricated; topic variance 6 percent to 29 percent.
• Vectara HHEM-2.3 (2025): Gemini 2.0 Flash approximately 7.6 percent; GPT-4.5 Preview approximately 12 percent.

Every AI-generated citation is verified before publication.

8.2 The Verification Protocol

Step 1, list every citation in the draft: named sources, statistics with stated sources, quotes attributed to real people, references to studies.

Step 2, locate the primary source for each, not a secondary citation. A claim citing "a 2024 Harvard study" requires the actual study by title and author. A claim citing a percentage requires the report and the page.

Step 3, verify the claim against the source. Three fabrication patterns: (a) source exists but cited claim is not in it (most common); (b) source does not exist (easy to detect on search); (c) source exists, claim is in it, but mis-quantified or mis-attributed (most dangerous, partial truth).

Step 4, decide per citation. Verified: hyperlink to the primary source. Unverified: substitute an actual primary source or remove the claim. "Studies have shown" without citation is not acceptable.

Step 5, log the verification. For YMYL, the credentialed reviewer countersigns the log.

8.3 The Tool Stack

Citation existence: direct visit to the cited URL; Google Scholar for academic; PubMed for medical; SEC EDGAR for financial filings; CourtListener for legal; Wayback Machine for deleted or moved sources.

Plagiarism: Copyscape, Originality.ai's plagiarism module, or comparable. These detect text overlap from indexed sources, not citation fabrication.

Statistic verification: the original report or study, not a secondary press release. Industry reports often gate the relevant statistic behind a paywall; editorial budget covers the cost.

8.4 The Citation Format

Every cited claim has a visible source.

<!-- Inline citation with primary source link -->
<p>
  The December 2025 Google core update affected approximately 67 percent of YMYL sites
  <a href="https://almcorp.com/blog/google-december-2025-core-update-complete-guide/" class="citation">
    (ALM Corp, December 2025)
  </a>.
</p>

<!-- Footnote pattern for academic-style citations -->
<p>
  Stanford researchers tested GPT-4 at 58 percent hallucination on legal questions
  <sup><a href="#footnote-1" id="ref-1">1</a></sup>.
</p>
...
<aside class="references">
  <h2>References</h2>
  <ol>
    <li id="footnote-1">
      Dahl, Magesh, Suzgun, Ho. Large Legal Fictions. Stanford HAI, January 2024.
      <a href="https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries">link</a>
      <a href="#ref-1" aria-label="Return to text">return</a>
    </li>
  </ol>
</aside>

Published articles with cited claims but no visible source links fail Section 8.

8.5 The Citation Verification Audit Pass

Quarterly, sample 10 percent of recently published articles. For each: pick one citation at random, click through to the source, verify the claim is supported, record pass or fail. Failure rate above 5 percent means the workflow is leaking unverified citations: increase rigor at Section 6.4 Edit and re-audit. Above 10 percent: pause AI-assisted publication until the protocol is re-established.

9. Transparency and Disclosure

9.1 The Site-Wide AI Use Policy

Every site using AI publishes a policy at a stable URL (recommended /ai-policy/ or /editorial-policy/#ai-use), linked from the footer and the per-article callout. Contents: whether AI is used; tasks AI is used for (Section 4); tasks excluded (Section 5); human review process (Sections 6 and 7); fact-checking (Section 8); YMYL credentialed reviewer requirement (Section 10 and framework-ymyl.md Section 5.8); editorial contact; last update date.

<article class="ai-policy">
  <h1>AI Use Policy</h1>
  <p class="policy-date">Last updated {{POLICY_DATE}}</p>

  <section>
    <h2>Whether We Use AI</h2>
    <p>{{HONEST_STATEMENT}}</p>
  </section>

  <section>
    <h2>What We Use AI For</h2>
    <ul><li>{{TASK_1}}</li><li>{{TASK_2}}</li><li>{{TASK_3}}</li></ul>
  </section>

  <section>
    <h2>What We Do Not Use AI For</h2>
    <ul>
      <li>Drafting articles on topics where we lack expertise</li>
      <li>Generating quotes from real people</li>
      <li>Fabricating statistics or citations</li>
      <li>Publishing without human review</li>
    </ul>
  </section>

  <section>
    <h2>Our Human Review Process</h2>
    <p>{{WORKFLOW_DESCRIPTION}}</p>
    <p>For YMYL content (health, finance, legal, civic, safety), a credentialed
    reviewer signs off before publication.</p>
  </section>

  <section>
    <h2>Our Citation Standard</h2>
    <p>Every factual claim traces to a primary source. Citations are hyperlinked.
    We do not publish unverified AI-generated citations.</p>
  </section>

  <section>
    <h2>Per-Article Disclosure</h2>
    <p>Articles with meaningful AI assistance carry a callout near the byline
    naming the human reviewer.</p>
  </section>

  <section>
    <h2>Editorial Contact</h2>
    <p><a href="mailto:editorial@{{DOMAIN}}">editorial@{{DOMAIN}}</a></p>
  </section>
</article>

9.2 The Per-Article Disclosure Callout

Articles where AI involvement exceeds the 5 to 10 percent grammar polish band carry a disclosure callout near the byline (top of article, not bottom).

<aside class="ai-disclosure" role="note" aria-label="AI use disclosure">
  <h3>About this article</h3>
  <p>{{SPECIFIC_DESCRIPTION_OF_AI_INVOLVEMENT}}</p>
  <p>Reviewed and verified by
    <a href="/authors/{{REVIEWER_SLUG}}/">{{REVIEWER_NAME}}</a>,
    {{REVIEWER_CREDENTIALS}}.
  </p>
  <p>{{REVIEWER_NAME}} verified all factual claims, citations, and recommendations
  in this article. They take professional responsibility for its accuracy.</p>
  <p><a href="/ai-policy/">Read our full AI use policy</a>.</p>
</aside>

Specific descriptions outrank generic ones. "Drafted with AI assistance" is generic. "Research synthesis and first draft generated with Claude (Anthropic) from an outline and source manifest prepared by the named reviewer; reviewer substantively revised, added original analysis from professional experience, and verified all citations against primary sources" is specific. The latter is the standard.

9.3 Required Versus Optional Disclosure

Required: AI involvement above the 25 percent band; all YMYL content with AI at any stage; all content to EU users informing on matters of public interest (EU AI Act Article 50, August 2026); all AI-generated reviews, testimonials, endorsements (FTC October 21, 2024).

Recommended: AI limited to grammar editing or polish. Not regulation-required, but the 12 percent Reuters 2025 baseline of respondents comfortable with fully AI-generated news climbs significantly when humans are explicitly in the loop.

Not required: entirely internal AI use (brainstorming where no surface text survives). Honest test: would a reader be surprised to learn AI was used here?

9.4 The Schema Markup Pattern

Disclosure is also machine-readable via schema.org creativeWorkStatus and reviewedBy:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "{{ARTICLE_TITLE}}",
  "author": {
    "@type": "Person",
    "name": "{{REVIEWER_NAME}}",
    "url": "https://{{DOMAIN}}/authors/{{REVIEWER_SLUG}}/",
    "jobTitle": "{{REVIEWER_TITLE}}"
  },
  "reviewedBy": {
    "@type": "Person",
    "name": "{{REVIEWER_NAME}}",
    "url": "https://{{DOMAIN}}/authors/{{REVIEWER_SLUG}}/"
  },
  "creativeWorkStatus": "AI_ASSISTED_HUMAN_REVIEWED",
  "publisher": {
    "@type": "Organization",
    "name": "{{PUBLISHER_NAME}}",
    "url": "https://{{DOMAIN}}/"
  },
  "datePublished": "{{ISO_DATE}}",
  "dateModified": "{{ISO_DATE}}",
  "mainEntityOfPage": "https://{{DOMAIN}}/{{ARTICLE_PATH}}/"
}
</script>

creativeWorkStatus is the operational disclosure slot. ChatGPT, Perplexity, and Claude have been observed factoring schema-level disclosure into citation decisions per framework-aicitations.md tracking from late 2025.

10. YMYL AI Content

10.1 The Elevated Standard

YMYL (health, finance, legal, civic, safety, major life decisions) is held to elevated standards regardless of AI involvement. When AI is involved the standards apply more strictly because Section 8 hallucination risk carries higher consequences. A hallucinated legal citation in non-YMYL is a Trust violation; in YMYL it is potential civil liability. framework-ymyl.md Section 5.8 is the full requirement; below is operational summary.

10.2 The Credentialed Reviewer Requirement

For YMYL AI-assisted content, a credentialed human reviewer signs off before publication. Credentialed means the reviewer holds the relevant credential:

• Health: licensed clinician (MD, DO, RN, NP, PA), licensed mental health professional (PsyD, PhD, LCSW, LPC), registered dietitian (RD, RDN).
• Finance: CFP, CFA, CPA, ChFC; SEC registration for investment advisors; FINRA registration for securities; state insurance license for insurance content.
• Legal: bar admission in the relevant state for the topic discussed.
• Civic and elections: relevant academic credential, journalism credential with civic beat history, or institutional affiliation with a non-partisan civic organization.

The reviewer's name, credentials, license number where applicable, and license verification link appear on the published article. The credentialed reviewer takes professional responsibility for accuracy; this is the substantive backstop that makes YMYL AI content publishable.

10.3 The Regulatory Layer

YMYL AI intersects regulatory frameworks beyond search ranking. Health: HIPAA for patient data, FDA for devices/pharma, state medical practice acts. Finance: SEC anti-fraud, state insurance, FINRA suitability. Legal: state unauthorized-practice-of-law rules, bar advertising. Civic: September 2025 SQRG YMYL expansion; EU AI Act Article 50 public-interest text disclosure. Every YMYL AI workflow has legal counsel review of the AI use policy. See framework-ymyl.md Section 6 for per-topic mapping.

10.4 The YMYL Disclosure Callout

<aside class="ai-disclosure ymyl-disclosure" role="note">
  <h3>About this article</h3>
  <p>{{SPECIFIC_DESCRIPTION_OF_AI_INVOLVEMENT}}</p>
  <p>Reviewed and verified by
    <a href="/authors/{{REVIEWER_SLUG}}/">{{REVIEWER_NAME}}</a>,
    {{REVIEWER_CREDENTIALS_AND_LICENSE_NUMBER}}.
    <a href="{{LICENSE_VERIFICATION_URL}}" rel="external">Verify license</a>.
  </p>
  <p>{{REVIEWER_NAME}} verified all factual claims and recommendations against
  primary sources and current professional standards. They take professional
  responsibility for the accuracy of this content.</p>

  <p class="ymyl-disclaimer">
    This article is for general information only and is not a substitute for
    {{personalized medical | legal | financial}} advice. Consult a qualified
    {{clinician | attorney | financial advisor}} for advice specific to your situation.
  </p>

  <p><a href="/ai-policy/">Read our full AI use policy</a>.</p>
</aside>

The disclaimer is not a substitute for the credentialed reviewer requirement. Both are required.

10.5 The YMYL Citation Standard

Every citation has a primary source link and a date of access. Health citations include PubMed ID where applicable; finance includes SEC filing identifier; legal includes case citation in standard format with public docket link. The credentialed reviewer countersigns the Stage One citation manifest (Section 6.1).

11. AI Content Detection Tools

11.1 Vendor Claims

Three tools dominate. Originality.ai: 99 percent accuracy with 0.5-2 percent false positive rates per their 2025 benchmarks. GPTZero: 99.39 percent overall with claimed 0 percent false positive rate (GPTZero.me 2025). Copyleaks: 99.12 percent. These are vendor claims.

11.2 Independent Test Results

Scribbr 12-tool comparison (2025): GPTZero correctly identified 52 percent of texts, Originality.ai 76 percent, Copyleaks 66 percent. Diverges 23-47 points from vendor claims because vendor benchmarks use clean AI output against clean human writing; real content includes edited AI drafts, AI-assisted writing, paraphrased content, mixed human-AI text, and writing by non-native English speakers. September 2025 Inteletica found Originality.ai most accurate on realistic mixed inputs, but no tool approached vendor accuracy on real content.

11.3 The Non-Native English Speaker Bias

The most important finding for global-audience publications. Stanford 2023 study (Liang, Yuksekgonul, Mao, Wu, Zou):

• 61.22 percent of TOEFL essays (non-native English) classified as AI-generated.
• Near-perfect correct classification of U.S.-born eighth-grader essays.
• All seven tested detectors unanimously misclassified 19 percent of TOEFL essays as AI.
• 97 percent of TOEFL essays flagged by at least one detector.

Structural bias. Detectors score perplexity; non-native English writing scores lower on lexical richness and syntactic complexity, which detectors mistake for AI predictability. Using detector output as a publishing gate systematically discriminates against non-native English speakers.

11.4 Why Detection Is Not a Quality Signal

Detector output correlates weakly with method of creation and barely at all with quality. Quality is determined by Information Gain (framework-infogain.md), accuracy (Section 8), bylined responsibility (Section 7.3), substantive coverage (framework-hcs.md Section 5.1.4), and depth beyond top SERP results. None of these are measured by AI detectors. September 2025 SQRG aligns: raters evaluate quality, not method.

11.5 Avoidance Versus Disclosure

Humanizer tools (Undetectable AI, StealthGPT, Phrasly, Quillbot paraphraser) are hostile to Section 9 disclosure and inconsistent with Google. Asymmetric risk: humanized AI is currently detectable at lower rates than raw AI, but detectors are improving. When detected, AI-plus-deception is worse than honest disclosure. Fix: disclose consistent with Google, EU AI Act Article 50, FTC, Authors Guild.

11.6 The Permitted Role

Narrow internal diagnostic at Section 6.4 Edit. After substantive human revision, a detector surfaces sections retaining too much AI-like phrasing for further revision. After human first-draft writing, a detector check identifies accidental AI tells. Detector output signals where to revise, not whether to publish. Gates remain Section 6.3 bylined responsibility and Section 6.4 fact-check. Where used, the result is in the internal audit trail, not the public disclosure.

12. Common AI Workflow Mistakes

Ten anti-patterns from 2025-2026 audits.

12.1 Volume Beyond Review Capacity. "50 articles per week" from a 2-editor team (40 hours each) yields 0.5-1 hour per article, below the 1-2 hour substantive minimum. Section 5.6 scaled-content-abuse-by-arithmetic; December 2025 update flagged approximately 87 percent traffic loss. Recovery: match publication to capacity (typically 5-15 per editor per week); re-audit per Section 13.

12.2 Generic AI Disclosure. Every article carries the same boilerplate ("written with AI assistance and reviewed by our editorial team"). Performative, not substantive. Reuters 2025 found only 33 percent believe journalists routinely check AI outputs; generic disclosure does not move that. Recovery: Section 9.2 specific pattern naming tool, stages, reviewer with credentials.

12.3 Fake Bylines. AI content under a byline where the named human did not write or substantively revise. SQRG Lowest signal; civil liability for YMYL; FTC deceptive-practices risk if endorsements. Recovery: substitute the actual reviewer's byline with disclosure, or unpublish.

12.4 Citation Hallucination Survival. Published articles cite studies, cases, or reports that do not exist or do not say what is claimed. Section 8 protocol was not run. Detectable by competitors, journalists, Reddit. Recovery: audit per Section 8.5; pull articles with verified hallucinations; add Section 8.2 as a hard gate before Section 6.6 Publish.

12.5 Detection Avoidance As Strategy. AI run through humanizer tools and published without disclosure. Section 11.5; inconsistent with Google, FTC, EU AI Act, Authors Guild. When detected, AI-plus-deception is worse than honest disclosure. Recovery: stop the humanizer step; disclose per Section 9; improve content per Sections 6 and 7.

12.6 YMYL AI Without Credentialed Reviewer. AI health, finance, legal, or civic content with non-credentialed editorial review. Section 10 elevated standard; civil liability; site-wide Trust violation. Recovery: contract credentialed reviewers; re-review existing; pause YMYL AI until credentialed review is in place.

12.7 The "Update To Fresh" Pattern. Articles run through AI to update stats, dates, references; dateModified updated; re-promoted as fresh, without substantive review. framework-hcs.md Section 9.6 plus AI hallucination risk; updated stats may be fabricated. Recovery: framework-contentrefresh.md Section 7 workflow; verify every updated statistic; roll back dateModified on stat-only updates.

12.8 Fact-Sheet Bulk Output. Programmatic city, service, or product pages where AI fills a template across hundreds. Technically unique, substantively redundant. Doorway-page pattern; SQRG Filler Content classification; March 2024 scaled content abuse update. Recovery: differentiate substantively per location, service, or product; reduce page count to what can be differentiated; apply framework-localseo.md for genuinely local content.

12.9 No Site-Wide AI Policy. AI is used but no public policy describes it. SQRG raters research per framework-sqrg.md Section 4.2.6 and look for editorial and AI policies; EU AI Act Article 50 requires affirmative disclosure. Recovery: publish per Section 9.1; link from footer; reference from per-article callouts.

12.10 AI in Brand Voice Without Voice Control. AI content published without brand-voice revision; reads as generic AI prose. Publication loses distinctive voice; Information Gain contribution per framework-infogain.md gets diluted. Recovery: brand voice document (Phase 2 sibling brandvoice scheduled); revise AI drafts to match voice in Section 6.3 Draft stage, not post-publication.

13. Audit Rubric

Three levels: per-article, site-wide policy, first 90 days. Total score: 56 points.

13.1 Per-Article Audit Criteria

Score: 28 max per article. World-class: 25+/28 with zero Critical fails.

13.2 Site-Wide AI Policy Audit Criteria

Score: 28 max. World-class: 25+/28 with zero Critical fails.

13.3 First-90-Days Audit Criteria

For sites starting AI-assisted production or recovering from a December 2025 core update impact, the first 90 days are audited for trajectory.

Pass/fail per criterion. First 90 days passes when all Critical criteria pass and at least 5 of 7 total pass.

13.4 Overall Audit Score

Combined: 56 points (28 per-article average + 28 site-wide).

Any Critical fail in Section 13.1 (A1, A2, A3, A5, A6, A7, A11, A12) or Section 13.2 (S1, S5, S6, S8, S9, S10, S11) puts the site at significant risk in the next core update regardless of overall score. Critical items are non-negotiable.

14. Maintenance Schedule and Report Templates

14.1 Weekly

Confirm every article this week carries an appropriate disclosure callout if AI was above the 25 percent band. Confirm every YMYL article names a credentialed reviewer. Review one randomly selected article against Section 13.1 per-article criteria.

14.2 Monthly

Citation verification: 10 percent sample, Section 8.5 protocol; investigate failure rate above 5 percent. Disclosure consistency check across the month. Brand voice spot-check on 3 articles.

14.3 Quarterly

Full per-article audit on 10 percent sample against Section 13.1. Full site-wide audit against Section 13.2. AI Use Policy review for accuracy versus current practice. Workflow review confirming Section 6 stage gates operate as documented. Detector tool review confirming its role remains limited to internal diagnostic per Section 11.6.

14.4 Annually

Comprehensive content inventory: total AI-assisted articles, distribution across Section 3.1 bands, percentage disclosed, percentage with credentialed YMYL sign-off. Strategic review on publication volume versus review capacity. Regulatory review covering EU AI Act Article 50, California CAITA and AB 853, FTC rule, and any new disclosure regulations. Detection bias check confirming detectors are not discriminating against non-native English contributors per Section 11.3.

14.5 Post-Core-Update

Within 14 days of every confirmed Google core update: identify pages that lost ranking 30 percent or more for 14 consecutive days; apply Section 13.1 to affected pages; cluster failing Critical criteria; remediate; if cluster falls on AI-assisted content, audit Section 13.2 site-wide; track recovery over 30 to 60 days. If no recovery, escalate to framework-coreupdates.md.

14.6 AI Workflow Implementation Report Template

# AI-Assisted Content Workflow Implementation Report

**Project**: {{BUSINESS_NAME}}
**Implementation Date**: {{TODAY}}

## Summary
- AI Use Policy published: {{YES_OR_NO}} at {{URL}}
- Workflow stage gates installed: {{COUNT}} of 6
- Per-article disclosure pattern installed: {{YES_OR_NO}}
- YMYL credentialed reviewer infrastructure: {{STATUS}}
- Citation verification protocol active: {{YES_OR_NO}}

## Editorial Capacity Baseline
- Editors: {{COUNT}}, Hours/week: {{HOURS}}
- Hours/article (substantive review): {{HOURS}}
- Resulting capacity: {{ARTICLES_PER_WEEK}}
- Current publication rate: {{ARTICLES_PER_WEEK}}
- Status: {{at_or_below | exceeds}}

## Stage Gate Status
{{STATUS_PER_STAGE_1_THROUGH_6}}

## YMYL Credentialed Reviewers
{{LIST_BY_NAME_CREDENTIAL_TOPIC_LICENSE_URL}}

## AI Tasks In/Excluded
{{SECTION_4_USED}} / {{SECTION_5_EXCLUDED}}

## Validation Results
{{SECTION_13_AUDIT_RESULTS}}

## Sign-Off
Implementation lead, Editorial lead, Legal (if YMYL), Date.

14.7 AI Workflow Audit Report Template

# AI-Assisted Content Workflow Audit Report

**Site**: {{BUSINESS_NAME}}
**Audit Date**: {{TODAY}}

## Executive Summary
{{ONE_PARAGRAPH_ASSESSMENT}}
**Overall Score**: {{X}}/56  **Status**: {{BAND}}  **Critical Fails**: {{COUNT}}

## Per-Article Findings (10 percent sample)
{{DETAILS}}

## Site-Wide Findings
{{DETAILS}}

## December 2025 Core Update Impact
- Pre/Post traffic: {{NUMBERS}}, Change: {{PERCENT}}
- Content farm pattern match: {{YES_NO}}, Recovery trajectory: {{ASSESSMENT}}

## Citation Verification Spot-Check
- Sampled: {{COUNT}}, Hallucinated: {{COUNT}}, Failure rate: {{PERCENT}}
- 5 percent threshold: {{PASS_OR_FAIL}}

## YMYL Findings (if applicable)
- Total YMYL articles: {{COUNT}}
- With credentialed reviewer: {{COUNT}}, Missing: {{COUNT_AND_URLS}}

## Critical / High / Medium Findings
{{PRIORITIZED_LIST}}

## Estimated Remediation Effort
{{HOURS_AND_COMPLETION_DATE}}

## Sign-Off
Auditor, Date, Re-audit recommended {{DATE}}.

End of Framework Document

Document version: 1.0
Last updated: 2026-05-14
Maintained by: ThatDeveloperGuy

This framework is the positive playbook. The negative constraints live in the sibling frameworks. The fastest path to good AI-assisted content is the hardest: be honest about the workflow, name a human who takes professional responsibility, verify every citation against primary sources, disclose openly. Sites that adopt this posture clear the December 2025 core update bar, the September 2025 SQRG bar, the EU AI Act Article 50 bar, and the FTC deceptive practices floor simultaneously. Sites that do not are below the floor on at least one dimension.

Companion documents: framework-contentfirst.md, framework-hcs.md, framework-ymyl.md, framework-sqrg.md, framework-eeat.md, framework-infogain.md, framework-schema.md, framework-entitysalience.md, framework-contentrefresh.md, framework-contentaudit.md, framework-aicitations.md, framework-aioverviews.md, framework-searchgpt.md, framework-perplexityspaces.md, framework-trustsignals.md, framework-cross-stack-implementation.md, framework-react.md, framework-tailwind.md, SEO-Search-Appearance.md, SERP-Optimization.md.

Phase 2 siblings scheduled: framework-brandvoice.md (brand voice consistency in AI-assisted production), framework-contentbriefs.md (content brief structure for AI handoff).

From the ThatDevPro Engine Optimization framework library. Studio: ThatDevPro (SDVOSB veteran-owned web + AI engineering). Sister property: ThatDeveloperGuy. Source: https://www.thatdevpro.com/insights/framework-ai-content-workflow/.