Asking whether your content was written by AI is like asking whether a financial statement was typed or handwritten, instead of asking whether it was audited. The instrument isn't the question; the attestation is. What gets B2B content cited is not provenance but process — sourced claims, and a named person who checked them.
Both artifacts below are free to copy. This page assumes you already decided to use AI; the question it answers is what you put in place around it.
Be careful with that claim, though, because the sloppy version of it is wrong. This is not “AI content is fine.” AI genuinely does correlate with the failure mode — but because volume production skips sourcing, not because a model formed the sentences. The correlation is real, and the causation runs through process. Detection is a proxy for it so noisy that it misdirects the fix: it flags the tool and misses the unsourced number, which is the thing that actually costs you.
Google's own documentation lands in the same place. Its guidance on generative AI content says the technology “can be particularly useful when researching a topic, and to add structure to original content,” while warning that using such tools “to generate many pages without adding value for users” may violate its scaled content abuse policy. And that policy is written about output rather than instrument: it targets “large amounts of unoriginal content that provides little to no value to users, no matter how it's created.” Unoriginal and valueless is the test. Machine-written is not.
What actually changed in August 2026?
Two things, and they land on different people. The EU AI Act's Article 50 transparency obligations became applicable on 2 August 2026, and the European Commission adopted its interpretive guidelines on 20 July 2026 — thirteen days before the rules they explain took effect.
Marking of AI output became a provider obligation. Under Article 50(2), providers of generative AI systems must mark their output in a machine-readable format and make it detectable as AI-generated. The guidelines are deliberately technique-neutral: watermarks, metadata, cryptographic provenance, logging and fingerprints all qualify, alone or in combination. This is not a setting you control. One date to know — a regulation amending the AI Act took effect on 27 July 2026 and gives systems already on the market before 2 August a transitional period until 2 December 2026.
Visible disclosure became a narrow deployer obligation. Under Article 50(4), organizations publishing AI-generated or AI-manipulated content must disclose it in two cases: deepfakes, and text published to inform the public on matters of public interest. The guidelines put ordinary marketing outside that second category by name — their own example of out-of-scope text is “AI-manipulated text that is part of a company's advertisement or product descriptions.” The carve-out has an edge worth knowing: it does not extend to claims about health, consumer safety or sustainability, and AI-manipulated investor material on a listed company's website is given as an in-scope example.
Penalties for the transparency obligations run to €15 million or 3% of worldwide annual turnover, whichever is higher — though for SMEs and startups the Act inverts that to whichever is lower.
One timing detail for anyone sitting on a backlog: content generated before 2 August 2026 does not have to be labelled retroactively, but text generated before that date and published on or after it does.
Keep the two obligations apart, because conflating them produces the two most common wrong conclusions: that you have to badge every blog post, or that none of this touches you.
What counts as human review, and what doesn't?
This is the whole question, so it is worth stating precisely. The disclosure obligation for text carries an exception, and the Commission's guidelines describe it as two cumulative conditions.
Human review or editorial control. Either satisfies the first condition. Human review is a deliberate examination of the substance of the content by one or more natural persons with relevant knowledge and professional judgement in the subject matter — and fact-checking the accuracy of the content is named as a minimum requirement of that review. Editorial control is the same substantive engagement exercised by a responsible editorial entity with the authority to approve, alter or reject the text on substantive grounds, including checking facts and the trustworthiness of sources.
An identifiable person holding editorial responsibility. Not “the marketing team.” A natural or legal person with ultimate legal responsibility for the publication — an individual, an editorial board, or the company itself. The guidelines add a step most teams miss: that person's identity and contact details should be publicly available somewhere easily findable, such as a site's terms or legal information.
What explicitly does not qualify: superficial, solely formal or procedural checks such as spell-checking or grammatical correction, the mere existence of an editorial policy, automated review processes, and cursory approval without substantive engagement.
The test is not how much of the draft a human wrote. It is whether a specific human checked it and could have said no.
One apparent contradiction worth naming, since this page opened by arguing that provenance isn't the variable. The regulator is asking the attestation question too. The exception does not turn on how much of the draft a machine wrote; it turns on whether a qualified person examined the substance and could have rejected it. That is the same question a buyer asks and the same one an engine is trying to resolve when it decides whether to repeat your claim. Different audiences, different consequences, one test — the instrument is not what's being examined. Read from the governance side rather than the compliance side, that convergence is the argument of the enforcement section in our content governance piece.
That distinction is also why most teams' existing process fails the test while feeling like it passes it. A review that catches a clumsy sentence but never re-derives the number in the headline is a copyedit. It is the most common kind of review in B2B marketing, and it is the kind the guidelines exclude by name.
And the exception is voidable after the fact. This is the part nobody has internalized. If an AI system modifies, supplements or reformulates the content after editorial sign-off, the guidelines say the result must be treated as AI-generated — any substantive AI intervention after the review causes the exception to become void. Read that against how content actually ships. The headline, the meta description, the social copy, the summary block: written last, in a hurry, increasingly generated. A piece can pass a genuine review and lose the exception on the way out the door, and the restatement-surface block in the standard below is what stops that.
Why is “make our content less AI-generated” the wrong goal?
Because it is a technical dead end and a strategic mistake, in that order.
The technical part first. Text watermarking does not work by attaching a tag to a document. The published method here is Google DeepMind's SynthID-Text, described in Nature in October 2024: it modifies only the model's sampling procedure, so the mark lives in the pattern of word choices across a passage rather than in anything appended to it. Anthropic has not published its own algorithm — its notice describes an imperceptible watermark woven into the text, and gives effects rather than mechanism. The effects are consistent either way: the mark travels with copy-paste, may survive some editing, and degrades under heavy paraphrase or translation, at which point you have rewritten the text rather than removed a mark from it. Very short passages carry too little signal to detect at all.
So the entire category of tooling that promises to launder AI text is selling you a rewrite. Sometimes an expensive one, in a voice that isn't yours.
The strategic part matters more. Detection and disclosure are answering a question nobody in your funnel is asking. A buyer does not want to know whether a human typed your case study. They want to know whether the number in it is real and whether the customer said what you claim they said. Those are provenance questions in the only sense that pays — provenance of the claim, not of the prose — and they are answerable with a name, a date, and a source.
Nobody is asking whether a human wrote it. They are asking whether anyone stands behind it.
There is a practical footnote worth knowing. A detected mark is a weak signal in both directions. Anthropic's own framing is that it indicates content was processed by Claude and “is not fully conclusive” — translation, proofreading and summarizing all leave the same trace, so a mark does not establish that the ideas originated with a model. An absent mark proves even less: short passages, heavy editing, format conversion and older models all produce clean-looking output. Public detection tooling is not generally available yet either; Anthropic says details are coming in forthcoming technical documentation. Building a process around passing a test that is this noisy, and that you cannot currently run, is not a plan.
Two different risks, and only one of them is real for search.
On the machine side — citation, AI answers, search — we don't see meaningful exposure. If content does the things that make it credible anyway (an original take, real proof, attributed quotes, third-party statistics, primary sources, the answer up front), AI in the drafting process is not what costs you. Google's stated position points the same way: the policy targets unoriginal, low-value output no matter how it was made.
The human side is where the risk actually lives, and it is a different kind of risk. Where a visible “AI-generated” label is required, humans read it, and a label carries associations the content itself doesn't control. That is a perception and communications question, not a ranking one. Which way it breaks is genuinely open: labels could carry a stigma, or they could become common enough to stop meaning anything. We don't know, and neither does anyone telling you they do. Worth planning for, not worth betting a brand on.
Where does this argument stop applying?
It has a boundary, and naming it is what makes the rest of it credible rather than convenient.
Everything above holds for B2B content whose job is to inform, rank, convert, or get cited: informational and commercial pages, SEO content, case studies, documentation, and the material an engine synthesizes when a buyer asks about your category. In that setting the reader's question is whether the claim is true and who stands behind it. The provenance of the prose is not the variable, and optimizing for it means optimizing for an audit nobody is running.
It does not hold for consumer brand work, influencer content, or big-creative. There, human origin is part of what is being bought. An audience that follows a person is buying that person's attention and judgement, and learning a machine produced it is a genuine breach rather than a technicality — the same way a hand-thrown bowl and a machine-pressed one are not the same object even when the shape is identical. If your content's value depends on someone having made it, disclosure isn't a compliance question. It's the product.
Most B2B marketing is not that, which is why this page argues what it argues. But the line is real, and a team running both kinds of work should not carry one policy across it.
The editorial review standard (copy this)
This is the ship gate. One reviewer, never the author, working from the finished artifact rather than the draft.
EDITORIAL REVIEW STANDARD — AI-ASSISTED CONTENT
Run before publication. Reviewer is never the author.
REVIEWER AND AUTHORITY
[ ] Named reviewer recorded, and it is not the person who drafted the piece
[ ] Reviewer has explicit authority to reject, not only to suggest
[ ] Reviewer's name and the review date are recorded somewhere durable
CLAIMS AND NUMBERS
[ ] Every number re-derived cold from its source, not read for plausibility
[ ] Every derived figure (a multiple, a difference, a percentage-point gap)
traced to its own approved source, since arithmetic is a new claim
[ ] Population and time window confirmed for each figure
[ ] Any figure nobody can source is removed, not softened
RESTATEMENT SURFACES — WHERE GOVERNED CLAIMS GO WRONG
[ ] Headline
[ ] TL;DR or summary block
[ ] Stat callouts and pull quotes
[ ] Image captions and alt text
[ ] Meta title and meta description
[ ] Social copy drafted from the piece
ATTRIBUTION AND CONSENT
[ ] Every customer quote traces to a named person who said it
[ ] Approval for this specific use confirmed at review time, from the record
itself — never from a note about the record
[ ] No composite, illustrative, or reconstructed customer anywhere in the piece
[ ] Any figure lifted OUT of a quote and into house prose has its own source
SOURCES
[ ] Each external citation opened and confirmed to say what we claim
[ ] Statistics traced to primary source, not to an article citing it
[ ] No source that cannot be linked
SIGN-OFF
[ ] Reviewer: ______________________ Date: __________
[ ] Decision: approved / returned for changes / rejected
[ ] Nothing in this asset is regenerated by AI after this point — headline,
meta description, summary and social copy included. A substantive AI
edit after sign-off sends the asset back through review.The AI use policy (copy this)
Shorter than you expect, because a policy that nobody can recite is decoration. The load-bearing lines are the four prohibitions and the ownership clause.
AI USE POLICY — MARKETING CONTENT
[Company] · Version 1.0 · Owner: [name, role] · Review: [date]
1. SCOPE
Applies to all external-facing content: web pages, blog posts, case
studies, emails, sequences, decks, social copy, ad copy, and video
scripts. Applies regardless of which tool produced the draft.
2. WHAT AI MAY DO
Draft from an approved brief. Restructure or compress material we
already hold. Summarize an interview or a document. Generate
variations for testing. Translate. Suggest headlines.
3. WHAT AI MAY NOT DO — NO EXCEPTIONS
a. Invent, reconstruct, or paraphrase a customer quote. Quotes are
used verbatim from an on-record source or not at all.
b. Produce a composite or illustrative customer, named or unnamed.
c. Generate, infer, estimate, or recompute a metric. Numbers come
from the source of record.
d. Introduce a factual claim that is not traceable to an approved
source in that source's own words.
4. OWNERSHIP
Every published asset has one named human owner, recorded before
publication. The owner is accountable for the content as if they had
written every word, because for these purposes they did. The person or
function holding editorial responsibility is identifiable publicly,
not only internally.
5. REVIEW
No asset publishes without the Editorial Review Standard completed by
a reviewer who is not the author. The completed review is retained.
Nothing is regenerated by AI after sign-off — headlines, meta
descriptions, summaries and social copy included. A substantive AI
edit after review sends the asset back through review.
6. SOURCE OF RECORD
Approved claims, figures, and customer proof live in one place, in
[system]. Copies elsewhere are not authoritative and are deleted or
redirected when found. Permission to use a quote is checked at
production time, from the record.
7. DISCLOSURE
We disclose AI involvement where law or platform terms require it, and
where a reader would reasonably feel misled without it. We do not use
tools whose purpose is to obscure that AI was involved.
8. WHEN IN DOUBT
Escalate to the policy owner before publishing, not after. A delayed
asset costs a day. A false claim in market costs the claim.Where do approved customer quotes fit into this?
Not where most people assume, and the accurate version is more useful than the assumption.
Approved quotes do not make AI-assisted content stop being AI-assisted. They do not remove marks from the prose around them, and quote density has nothing to do with whether a review obligation applies. If you are hoping a wall of testimonials exempts you from something, it does not.
What a consent record does is supply the half of the standard that is otherwise hard to evidence: a specific, identifiable person accountable for a specific claim. A quote captured on the record, attributed to a named human at a named company, with a dated approval for this use, is auditable proof that someone stands behind the statement. That is exactly what a reviewer needs, what a skeptical buyer wants, and what a synthesizing engine can weigh.
Two rules make it work, and both are easy to get wrong:
- A quote is not a claim. A figure inside an approved quote rides that speaker's approval — it is their attributed observation. Paraphrase it into your own prose and it becomes your market claim, needing its own source and its own backing. This distinction moves in one direction only, and it moves the moment you drop the quotation marks.
- Permission is read live, never cached. Whether a quote is approved for a given use is answered at production time, from the record. When a policy document and the source record disagree about what is approved, the record wins. Documents go stale confidently; records do not.
Both rules cost us real production mistakes to learn, which is why they are stated as flatly as they are.
We ruled the first one in production this summer, and it meant pulling a customer's industry-general observation out of three assets that were already in market. Nobody had invented anything: the line was that customer's honest read of their own industry, said on the record, and we had lifted it out of the quotation marks and into our prose. Inside the quote it was their observation. Outside it, it was our market claim — and it was never our figure to assert.
The second cost us five days. A governing document carried a restriction that had gone stale, and for those five days it graded client-approved material as unapproved. Nothing was at risk except our own throughput, which is the point: the document was confidently wrong while the source record was right the entire time. That is the failure mode a cached permission produces, and it is why the rule reads the way it does.
The longer argument for why claims should originate on the record at all is in content governance when AI writes the first draft.
What should you actually do this week?
In this order, because the sequence is what makes it survivable:
- Name the owner. One person accountable for published content. This takes a minute and it is the single line that makes everything else enforceable.
- Adopt the review standard for new assets only. Do not start with a back catalog audit — you will stall. Forward-only, from the next thing you ship.
- List your numbers. Every figure currently in market, with one source each. This surfaces at least one number nobody can source, and that discovery is the point of the exercise.
- Fix your sequences before your website. Outbound copy drifts fastest, because numbers get retyped from memory there and nobody reviews an email template.
- Then decide about disclosure — with counsel if you sell into Europe, and on your own judgment about what a reader would feel misled by. That decision is easier once the first four are done, and it matters less than any of them.
What the manual version cannot do is stay alive. Interviews pile up unmined, consent answers go stale, and the list of numbers you built in August diverges from what is true by December. Keeping it current is the work, and it is the part worth building infrastructure for.
Where we fit. Everything above is yours to run in a document and a spreadsheet, and running it puts you ahead of most of your market. What we build is the layer underneath: customer proof captured on the record through real interviews, attributed to named people, with consent read live at production time rather than cached in a note. If you want to see what that looks like running, grab thirty minutes with us.
Related: Content Governance When AI Writes the First Draft · Why Verified Proof Is the Currency of AI Citation · A GEO Playbook for B2B SaaS Startups · GTM Engineering for Messaging and Content
This page is not legal advice. Whether a specific obligation applies to your content depends on facts we do not have, and the interpretation of these rules is still developing. Talk to your counsel about your situation.
