Discussion
  • Content SEO

Where does AI assistance stop in your editorial process: outline, first draft, fact-check, or never in publishable copy? What set that line?

Published 1 reply

Discussion starter: a content SEO question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.

Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.

1 reply

  • admin
    Admin
    Original poster

    Draw the line by what AI originates, not by production stage. The rule that holds up is this: AI can transform material a person supplied and can check, but it does not originate facts, quotes, figures, first-hand experience or expert judgement, and it is not the final check on any of them. On that rule, AI fits in research support, structuring, line editing and packaging (titles, summaries, alt text). It can write first drafts from the writer’s own notes for low-claim formats. It stays out of the fact-check of record.

    “Never in publishable copy” is a defensible line for some content: investigations, reviews based on hands-on testing, clinical or financial guidance. As a blanket rule it is hard to enforce, because AI now sits inside word processors, grammar tools and CMS editors, so the policy has to say which uses count.

    Four things set the line: the cost of verifying fluent text, accountability for every published claim, Google’s documented position on scale and value, and, for publishers in the EU, the way the AI Act treats human editorial control.

    What Google’s documentation supports, and what it leaves to you

    Google’s guidance on generative AI content says it “can be particularly useful when researching a topic, and to add structure to original content”, and warns that using it “to generate many pages without adding value for users may violate Google’s spam policy on scaled content abuse”. The helpful content guidance asks whether the use of automation, including AI generation, is “self-evident to visitors through disclosures or in other ways”, and says that using automation to produce content “for the primary purpose of manipulating search rankings” breaks its spam policies. The rater guidelines call generative AI a helpful tool for content creation that, like any tool, “can also be misused”.

    Two things follow. Google’s own examples of good use (research, and structure for original content) are assisting roles, not originating ones. And none of these documents sets a stage-by-stage rule or says AI-assisted copy ranks better or worse; Google’s 2023 guidance described its focus as the quality of content rather than how it is produced. Where the line sits is an editorial decision. The documents constrain scale and value, not stage.

    The line by task, not by stage

    Task Where AI helps Where it stops Human check
    Research Summarising sources you have opened; suggesting angles and gaps Citing a source nobody on the team has read Every source in the copy is opened and read by a person
    Outline Structuring the writer’s notes, interviews and test results Building the outline from what already ranks, which pulls the piece towards copycat structure and no new information Editor confirms what the piece adds
    First draft Low-claim formats from supplied material: product copy from verified specs, event listings, summaries of your own reporting Claim-dense, opinion, experience or YMYL pieces, where drafting is the thinking Writer rewrites and owns every claim
    Editing Grammar, house style, length, readability Changing claim strength (“may” becoming “does”) or meaning Compare edited sentences that carry claims against the original
    Fact-check Listing every checkable claim (names, numbers, dates, quotes, causal claims); flagging text that contradicts a table Declaring a claim true A person verifies each claim against a primary source
    Quotes, statistics, citations Transcribing interviews you recorded Supplying or rewording a quote or figure from memory Copied from the recording or source
    Titles, descriptions, alt text, schema Variants for a person to choose from; first-pass alt text Publishing unreviewed output at scale Spot-checks against the page
    Translation A draft for a native-speaking editor Publishing without native review Native editor checks terms, units and local rules

    Packaging is not a free pass. Google’s generative AI guidance applies its accuracy and quality focus to titles, meta descriptions, structured data and alt text as well as body copy, and Merchant Center requires AI-generated product titles and descriptions to be specified separately and labelled as AI-generated, with IPTC DigitalSourceType metadata on AI-generated images. An ecommerce team generating feed copy has a labelling requirement on that side even if its editorial pages carry no disclosure.

    The fact-check trap:

    Automating the final check is tempting, and it is the wrong job for AI. A model asked “is this correct?” returns a plausible answer, not a verified one, and it cannot see your notes, your recording or a paywalled source unless you supply them. Use it to build the checklist of claims, not to tick it.

    Why the first draft is the contested stage

    A fluent AI draft changes the editor’s job from writing to verifying. For a claim-light piece, that is a saving. For a claim-dense piece, it can be a loss: nobody sourced each claim while writing, so every claim has to be traced after the fact, and errors that sound right survive a skim. Drafts also pull towards the consensus of what is already published, which works against the thing that makes a page worth ranking: new information from reporting, testing or expertise.

    Hypothetical example:

    A 1,500-word explainer on a tax change contains 40 checkable claims. Written from the reporter’s notes, each claim is sourced as it is written, and the fact-check confirms the trail. Drafted by AI from a prompt, the same piece arrives with 40 claims of unknown origin. The fact-checker has to find a source for each, and anything that cannot be traced gets cut or rewritten. The draft took minutes; the verification took longer than writing the piece would have.

    The external constraints that set the line

    The EU AI Act makes human review the dividing line

    From 2 August 2026, Article 50 of the EU AI Act requires deployers who publish AI-generated or manipulated text to inform the public on matters of public interest to label it. The exception is content that has undergone human review or editorial control, where a person holds editorial responsibility for the publication. The European Commission’s FAQ (updated July 2026) describes human review as a deliberate examination of the substance by people with relevant knowledge and professional judgement, and says superficial checks such as spell-checking or grammar correction do not qualify. The FAQ’s grace period, to 2 December 2026, extends only to marking and detection requirements for systems placed on the market before 2 August 2026.

    For an EU publisher, then, a substantive edit by someone with the relevant knowledge, with a person or organisation holding editorial responsibility, is where the labelling duty stops. The scope is “matters of public interest” (politics, health, consumer safety and similar), not every page, and this is a summary, not legal advice.

    Accountability for claims

    Every published claim needs someone who can stand behind it, whether the exposure is defamation, consumer law or professional rules in health and finance. That is why quotes, figures and claims about named people or companies are the hard edge of the line. If nobody on the team can say where a claim came from, it does not run.

    Google’s scale constraint

    Google’s spam policies define scaled content abuse as many pages generated primarily to manipulate rankings rather than help users, however they are produced. Risk rises with volume and falls with added value, so a team using AI to publish more pages needs a stricter line than a team using it to publish the same pages faster.

    Setting the line by content risk

    Low risk

    Product copy from verified specs, listings, internal summaries, metadata variants. AI drafts allowed; a person checks against the source data and spot-checks at scale.

    Medium risk

    How-to guides, explainers, evergreen updates. AI for outlines from notes, editing and claim lists; people write the sections that carry the argument; human fact-check.

    High risk

    News reporting, investigations, reviews based on testing, health, legal and financial guidance. AI limited to transcription, research support, editing and claim extraction. No AI-originated prose in publishable copy.

    Human only

    Quotes, statistics, citations, first-hand experience, expert judgement and the final sign-off, at every risk level.

    Measure where your line should sit

    1. Design: pick one content type and produce paired pieces, some from AI first drafts built on the writer’s notes and some written from scratch, with the same editors and the same fact-check protocol. If you can, keep the fact-checker unaware of which is which.
    2. Primary metrics: errors caught at fact-check per 1,000 words (by severity), errors that reach publication (corrections), and total editor hours per piece.
    3. Secondary metrics: time to publish, engagement, and search performance as a guardrail.
    4. Size: editorial metrics build up per piece, so size the test by count rather than weeks. Twenty to thirty pieces per arm gives a first read, not a significance test. Search performance on a sample that small is too noisy to decide anything, which is why it is only a guardrail.
    5. Decision rule: move the line only if published error rates are equal or lower and total editor time falls. If AI drafts raise the error density found at fact-check, the draft is shifting work into verification, and for that content type AI belongs upstream (research) and downstream (editing and packaging).

    The policy I would write

    If this were my editorial team, the written policy would fit on one page. AI may transform supplied material and assist with research, structure, editing and packaging. It may draft low-claim formats from verified inputs. It may not originate facts, quotes, figures, experience or judgement. A named person verifies every claim against a primary source and owns the piece. Then I would revisit the line each quarter with the error and time data, and move it only when those numbers justify it.

    Need help with this on your own site?

    Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.

    See SEO services

Add a reply

Replies are open to approved members. Log in or apply to join.