The Human Editorial Pass: A Publish-or-Hold Workflow for AI-Assisted Drafts
“A human reviewed it” is not an editorial process. It might mean someone fixed two commas, glanced at a green detector score, and clicked publish.
A real human pass has named owners, visible evidence, and gates that can stop the work. This workflow is designed for a small publishing team, but one person can use it by changing hats between passes. The important part is separation: source validation, substantive editing, voice editing, and final approval are different jobs.
What This Workflow Is For
Use it when a draft was generated, outlined, summarized, translated, or substantially rewritten with AI assistance. Also use it for human drafts produced so quickly that factual and editorial review may have collapsed into one hurried pass.
This process does not use an AI detector as a gate. Detector scores do not prove authorship, and they do not tell you whether a page is accurate or useful. See why AI detector scores are not proof if a percentage is currently driving your review.
The workflow instead follows a simpler principle: the publisher is accountable for every sentence it publishes. The Associated Press states that AI does not replace the journalist’s central role in gathering, evaluating, and ordering facts, and that AP journalists remain responsible for accuracy and fairness. Google similarly asks publishers using generative AI to focus on accuracy, quality, relevance, and helpful context about creation.
Assign the Roles Before Editing
Do not let “everyone” own quality. That means nobody can stop the page.
| Role | Owns | Must not silently do |
|---|---|---|
| Commissioning editor | Reader, purpose, scope, deadline | Expand the topic after reporting begins without resetting scope |
| Author or reporter | Original work, draft, source packet, revision | Treat generated output as a source |
| Source validator | Factual claims, links, quotations, dates | Rewrite disputed claims to make them sound safer without resolving them |
| Voice editor | Clarity, rhythm, specificity, publication voice | Change the meaning of a verified claim |
| Approver | Risk, disclosure, final publish-or-hold call | Approve with unresolved blockers because the calendar is tight |
On a two-person team, one person can be author and voice editor while the other validates sources and approves. On a solo project, take a real break between roles and keep separate checklists. Distance is not independence, but it reduces the temptation to remember what you meant instead of reading what you wrote.
Gate 0: Commission the Article in Six Lines
Before anyone prompts or drafts, create a commissioning card:
- Reader: Who has this problem?
- Job: What should they be able to do afterward?
- Contribution: What will this page add that existing pages do not?
- Evidence: What original work and primary sources are required?
- Boundaries: What will the article explicitly not claim?
- Owner: Who can decide to hold publication?
Example:
For editors managing AI-assisted blog drafts, provide a repeatable publish-or-hold workflow. Add a role table, source log, and approval matrix. Use current Google and AP guidance. Do not claim that following the workflow guarantees rankings or eliminates errors. The managing editor owns the final call.
If the “contribution” line is only “a comprehensive guide,” stop. That describes a format, not value. Read the complete guide to AI slop and identify what this new page will actually do differently.
Gate 1: Require a Source Packet
The author hands off evidence with the draft. No packet, no edit.
The minimum packet contains:
- A list of every primary or authoritative source used.
- Notes showing which claim each source supports.
- Original interview, test, observation, or analysis notes.
- Versions or prompts that materially shaped the output, when relevant.
- Known uncertainties and facts that could not be confirmed.
- Permission or license notes for third-party media.
Use a simple source log:
| Claim ID | Draft claim | Source | Source type | Checked on | Result |
|---|---|---|---|---|---|
| C1 | Google defines scaled content abuse by purpose and value, not only by production method | Google spam policies | Primary policy | 2026-07-11 | Supported |
| C2 | Human review is part of AP’s experimental AI-assisted workflows | AP standards update | Primary organizational guidance | 2026-07-11 | Supported |
| C3 | Our readers prefer short intros | Internal analytics | First-party data | 2026-07-11 | Needs date range and sample size |
An AI answer, search result snippet, or autogenerated bibliography is not a source. It is a list of leads. The validator must open each cited page and confirm that the claim survives contact with it.
Gate 2: Validate Claims Without Editing the Voice
The source validator works from the source log and marks factual claims in the draft. For each material claim, choose one status:
| Status | Meaning | Action |
|---|---|---|
| Supported | The cited source directly supports the wording and scope | Keep; improve citation placement if needed |
| Overstated | The source supports a narrower claim | Narrow the claim and notify the author |
| Stale | The claim may have changed since the source was published | Find a current source or add a precise date |
| Misattributed | The source repeats another source or does not own the claim | Trace the claim to its origin |
| Unverifiable | No adequate evidence was found | Remove, report as uncertain, or hold the article |
| Original judgment | A conclusion, recommendation, or interpretation—not an external fact | Attribute it to the author and show the reasoning |
Check the dangerous details first: names, numbers, dates, quotations, policies, safety advice, causal language, superlatives, and claims about what a platform “rewards” or “penalizes.”
Google’s people-first content guidance asks whether a page provides original information or analysis, sources claims clearly, demonstrates expertise, and avoids easily verified factual errors. Treat those questions as an editorial floor, not an SEO trick.
The source-validation stop conditions
Hold the draft if any of these remain:
- A central claim depends on a source that does not support it.
- A quotation cannot be traced to an original recording, transcript, or document.
- A consequential claim has no qualified reviewer or authoritative basis.
- The article summarizes other summaries without adding original value.
- The author cannot explain where a key claim came from.
Do not “soften” an invented claim into something vague and leave it in. Unsupported fog is still unsupported.
Gate 3: Run the Substance Edit
Once the factual skeleton is stable, the substantive editor asks whether the article fulfills the commission.
Work in this order:
- Put the reader’s decision or procedure before the background lecture.
- Remove sections that do not advance the commissioned job.
- Turn abstract advice into steps, thresholds, examples, or decision rules.
- Surface meaningful trade-offs instead of producing balanced-looking lists.
- Add the author’s original observations where they change the recommendation.
- Make uncertainty explicit; do not fill gaps with confident transitions.
This is where many generated drafts fail. They are organized but not argued. Every heading looks plausible, yet no section changes what the reader can do. The 15-minute pre-publication audit gives you a weighted way to identify that problem before it reaches final approval.
Gate 4: Edit for Voice Without Adding Fake Humanity
Now the voice editor reads the article aloud and marks sentences that could appear on any competitor’s page.
Replace generic language with one of four things:
- A concrete noun: name the actual document, tool, person, or constraint.
- A decision: say what the publication recommends.
- A reason: show why that choice follows from the evidence.
- A limit: say when the advice fails.
Do not add fake personal anecdotes, random slang, intentional typos, or theatrical outrage to make a draft “sound human.” Voice comes from attention and judgment. If you want a line-editing checklist for repetitive syntax, padded transitions, and vague abstractions, use how to humanize AI writing.
Run these four tests:
| Test | Question | Failure signal |
|---|---|---|
| Substitution | Could another brand or topic replace ours unchanged? | Generic claims and borrowed voice |
| Read-aloud | Does every paragraph march at the same tempo? | Uniform sentence length and predictable transitions |
| Ownership | Can the author defend the recommendation? | Hedging, passive voice, or “experts say” |
| Compression | Can 30% disappear without losing meaning? | Repetition and conclusion-shaped filler |
Gate 5: Review Risk and Disclosure
Before approval, identify the highest foreseeable cost of being wrong. Increase review in proportion to that cost.
- Low risk: opinion, entertainment, or reversible how-to advice. Standard editorial review may be enough.
- Medium risk: purchasing, professional, or reputation-sensitive advice. Require current primary sources and a named approver.
- High risk: health, legal, financial, safety, elections, or accusations about people. Require appropriate subject-matter review and organization-specific policy; this lightweight workflow is not sufficient by itself.
Then decide whether readers need context about AI assistance. Google’s generative-AI guidance recommends sharing creation context when it helps users. Disclosure should answer a reader’s likely question, not hide behind “AI may have been used.”
Useful disclosures are specific: “An AI tool produced the initial transcript; the author checked it against the recording,” or “The illustration was generated for this article and reviewed by the editor.” If the tool had no material effect on what the reader sees, a disclosure may add noise. Establish the rule before publication and apply it consistently.
Gate 6: Use the Approval Table
The approver reads the commission, source log, unresolved notes, and final draft. They do not approve from the polished page alone.
| Gate | Pass evidence | Hold trigger | Owner |
|---|---|---|---|
| Purpose | Reader and job are explicit; every section serves them | Article answers a different or vague question | Commissioning editor |
| Original value | Draft includes reporting, testing, analysis, or a useful original procedure | Page only recombines existing summaries | Author + substantive editor |
| Sources | Material claims map to opened, current sources | Central claim is unsupported, stale, or misattributed | Source validator |
| Voice | Prose is specific, readable, and owned | Generic filler or fake anecdotes remain | Voice editor |
| Risk | Review depth matches potential harm | Required expert or policy review is missing | Approver |
| Transparency | Byline, method, and relevant AI use are clear | Readers could be materially misled about creation | Approver |
| Mechanics | Links, image alt text, metadata, and rendering are checked | Broken evidence trail or inaccessible media | Publisher |
The final decision must be one of three states:
- Publish: all gates pass; only non-substantive corrections remain.
- Revise: the contribution is sound, but named repairs remain. Assign each repair to an owner.
- Hold: missing evidence, expertise, permission, or accountability makes publication unsafe or pointless.
“Publish now, fix later” is not a fourth state.
A Small-Team Version You Can Run Today
If this process sounds heavy, use a 30-minute two-person version:
- Five minutes: Author hands over the commissioning card and source log.
- Ten minutes: Editor opens sources and checks the five most consequential claims.
- Five minutes: Editor marks generic paragraphs and missing decisions.
- Five minutes: Author repairs only those blockers.
- Five minutes: Editor completes the approval table and records publish, revise, or hold.
For low-risk articles, this creates named owners and recorded gates that an undefined “human in the loop” does not provide. For higher-risk material, add the expertise the topic requires rather than compressing the same checklist.
Primary Guidance Behind the Workflow
- Google Search: Creating helpful, reliable, people-first content — originality, sourcing, expertise, page quality, and who/how/why questions.
- Google Search: Guidance on using generative AI content — accuracy, quality, relevance, and creation context.
- Google Search spam policies — scaled content abuse is defined by manipulative purpose and lack of value, however content is produced.
- Associated Press: Standards around generative AI — journalists remain accountable for accuracy and fairness.
- Associated Press: Updates to generative AI standards — experimental AI-assisted translations and summaries begin with human journalism and receive human editing and vetting.
Accountability Is the Product
The point of a human editorial pass is not to sprinkle personality on machine prose. It is to create a traceable chain of responsibility: someone commissioned the work, someone gathered evidence, someone checked it, someone shaped the voice, and someone had the authority to say no.
That chain is slower than clicking publish. It is also the difference between a page that merely exists and one a reader can trust. When the work is done, take a break in The Rebel Arcade—a set of small, human-made experiences that celebrate attention, judgment, and imperfection.