The line platforms draw between AI-assisted and AI slop

A policy update lands. The headline says a platform is cracking down on AI slop. You drafted last week’s product page with Claude and made the thumbnail in ChatGPT, so you read it twice, trying to work out whether it describes you. The word in the headline is the same word that describes your workflow, and nothing in the coverage tells you where the boundary actually sits.

It sits somewhere quite specific, and it has sat there consistently for long enough now that you can plan around it. Read the policy pages themselves at YouTube, Google, Meta, Snapchat, Amazon and Etsy and one line runs through all of them, drawn in nearly the same words each time. It is not the line the headlines draw. This guide is for solo operators, freelancers and small teams publishing their own work or their clients’ work on platforms they do not control. It is not a guide to the AI labelling laws, which are a separate rulebook with separate enforcers. And it will not help anyone whose business depends on publishing volume at a rate a person could not review, because ending that is the entire purpose of the rules below.

None of these rules ban AI

Google’s spam policies define scaled content abuse as “when many pages are generated for the primary purpose of manipulating search rankings and not helping users” [2]. One of the listed examples is “Using generative AI tools or other similar tools to generate many pages without adding value for users” [2]. So far it reads like a rule about AI. The qualifier attached to the definition says otherwise: the practice is “typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it’s created” [2]. No matter how it’s created. The tool is explicitly not the test.

Google’s guidance for people making content puts the same rule the other way round. “If you use automation, including AI-generation, to produce content for the primary purpose of manipulating search rankings, that’s a violation of our spam policies” [3]. The violating condition is the purpose. Strip that clause out and the sentence permits automation.

YouTube reads identically. Its channel monetisation policies require content to “be your original creation” and “not be mass-produced, generic, repetitive, or manipulative” [1], and say creators “are rewarded for original and authentic content” [1]. The position on tooling then arrives in one sentence: “If you use automated tools or templates to help create your content, the final product must still demonstrate your creative vision and provide educational or entertainment value” [1].

Etsy goes furthest of the six. Its Creativity Standards list “Seller-prompted AI creations” as a legitimate category of thing you are allowed to sell, meaning “Creations that were generated using AI tools (e.g., AI image generators such as Dall-E) based on a seller’s original prompts” [7]. Sellers “must disclose within their listing description if an item is created with the use of AI” [7]. Generate an image with a model, print it, sell it, say so, and you are inside the rules. Sell the prompts themselves and you are not: “AI prompt bundles” sit on Etsy’s list of examples of items that do not qualify as designed by a seller [7].

The word doing the work is “unoriginal”

YouTube’s rule was called repetitious content until it took its current name, inauthentic content, on 15 July 2025 [1]. In July 2026 YouTube reorganised it into three named categories and was explicit that the rules themselves had not changed [8]. The categories are where the useful specificity lives.

Generic or repetitive content covers work that looks “like it’s made with a template, or that may feel repetitive to viewers after watching several videos in a row from the same channel” [1], and names by example “AI-generated content made with generic or unoriginal templates giving the impression of mass production without adding the creator’s original, authentic insights or perspective” [1]. Unsatisfying or off-putting content covers work that “relies heavily on emotionally manipulative formulas, mimics existing formats or stories to a degree that the videos feel interchangeable, or appears designed to shock or surprise viewers for the sole purpose of getting views” [1]. And channels that use “AI-generated personas to deliver information on sensitive topics” such as health, legal issues, finances or politics, presented as a human expert, “will not be allowed to monetize” [1][8]. Alongside those three sits the older and unchanged reused content policy [1], which covers repurposing “without adding significant original commentary, substantive modifications, or educational or entertainment value” [1].

Meta’s version, published on 14 July 2025, uses the same vocabulary for a different platform. “Unoriginal content reuses or repurposes another creator’s content repeatedly without crediting them, taking advantage of their creativity and hard work” [5]. Resharing, reaction videos and joining a trend all stay allowed, provided you “make it your own through creative editing, voiceover or commentary”, and Meta then sets the bar in a single sentence: “These enhancements must be meaningful—simply stitching together clips or adding your watermark does not qualify as meaningful enhancement” [5]. In the first half of 2025 Meta says it took action on around 500,000 accounts engaged in spammy behaviour or fake engagement, and took down around 10 million profiles impersonating large content producers [5].

Read three rulebooks in a row and the pattern is hard to miss. None of them is describing a technology. All of them are describing a supply problem. Every recommendation system and payout formula on these platforms was designed when producing a passable video, article or listing cost somebody hours. That cost was never the thing anyone valued, but it was doing quiet structural work. It capped supply, and it meant anything arriving in a feed had at least one person’s attention behind it. Generation tools removed the cap without replacing the guarantee.

So the platforms went looking for a replacement, and what they reached for is originality, because originality is the thing the old cost was standing in for. That is why the wording reads the way it does. Phrases like “no matter how it’s created” [2] and “meaningful enhancement” [5] are not vagueness or hedging. They are an attempt to write a rule that survives the next three generations of tools, because any rule that named the tools would be obsolete within a year of being published.

Amazon’s line runs through the first draft

Amazon’s KDP guidelines contain the most operationally precise version of this distinction published anywhere, and it is worth using as a general test even if you never publish a book.

“We define AI-generated content as text, images, or translations created by an AI-based tool. If you used an AI-based tool to create the actual content (whether text, images, or translations), it is considered ‘AI-generated,’ even if you applied substantial edits afterwards” [6]. The other category: “If you created the content yourself, and used AI-based tools to edit, refine, error-check, or otherwise improve that content (whether text or images), then it is considered ‘AI-assisted’ and not ‘AI-generated’” [6]. Amazon requires you to inform it of AI-generated content when you publish a new book or edit and republish an existing one, and states that “You are not required to disclose AI-assisted content” [6].

Notice what the classification is not sensitive to. Effort. You can rewrite every sentence of a generated draft and it remains AI-generated by this definition. The category is fixed at the moment the first version comes into existence, by whatever produced that first version. It is a test of sequence, not of quantity.

That has a direct consequence for how you work, and it is the single most useful thing in this guide. If you write the ugly first version and hand it to a model to tighten, cut and check, the work is yours under every definition on this page. If you prompt for the finished thing and edit downward, it is not, however much you edit. Same tools, same hours, often the same finished quality, and a different classification at the end of it. Most people who feel exposed by these rules are exposed by the order they do things in, not by the software they use.

Disclosure buys you honesty, not distribution

The instinct after reading a policy update is to add a disclosure and consider it handled. Snapchat’s ranking policy exists to correct that instinct, and it is unusually blunt about it: “Our content ranking algorithm rewards authentic, human-made content over wholly AI-generated content created outside of Snapchat, even when AI-generated content has transparency disclosures” [4]. Disclosing does not restore the ranking. Snapchat’s own carve-out is worth reading beside it, because “AI-generated content that was created within Snapchat is eligible for recommendation” [4]. The word authentic is doing commercial work there as well as editorial work.

Every platform here treats disclosure and quality as two separate axes. Google lists disclosure as a self-assessment question to ask yourself, “Is the use of automation, including AI-generation, self-evident to visitors through disclosures or in other ways?” [3], not as a ranking commitment. Amazon and Etsy make disclosure a condition of listing [6][7], which is a permission to publish rather than a claim on distribution.

Label because the marketplace requires it, and because a buyer who discovers the truth later costs you more than a ranking ever will. Do not label expecting it to answer the originality question, because on the platform that addressed this most directly, it does not.

The penalty attaches to the account, not the post

Reading these policies as rules about individual pieces of content is the mistake that turns a survivable problem into an unsurvivable one. Meta says accounts that improperly reuse others’ work repeatedly will “not only lose access to Facebook monetization programs for a period of time, but will also receive reduced distribution on everything they share” [5]. Everything they share, including the work you did properly. Meta also reduces the distribution of duplicate videos so that “original creators can get the visibility that they deserve”, and is testing links on duplicates pointing viewers back to the original [5].

YouTube’s inauthentic content rules sit inside its channel monetisation policies, and the reused content policy “applies to your channel as a whole” [1], so the unit being judged is the channel. Google’s scaled content abuse is defined around “many pages” [2], which is a pattern across a site rather than a fault in one URL. In each case the thing being classified is the property, not the item.

That inverts the obvious response. The fix is not to improve your weakest post. It is to change the ratio, which usually means removing or rewriting a run of thin items rather than adding a good one on top of them. A genuinely useful page sitting on a domain with 400 generated ones inherits the domain’s classification, and no amount of care spent on that one page changes the number underneath it.

Auditing a catalogue you published before you read this

Most of the exposure is not in what you publish next. It is in what is already live, published back when the rules were vaguer and the tools were newer, on accounts you have not looked at since.

Sort it using Amazon’s test [6]. Go through what you have published and mark each item by where the first draft came from. For everything a model originated, ask the question the policies are actually asking: what is in this that only you could have supplied. A named client, a number you measured, a decision you made and can defend, a photograph you took. If the honest answer is nothing, that item is the thing all six rulebooks describe, and it is doing quiet damage to everything published beside it.

Then check the shape of the set rather than the items. YouTube’s language about videos that feel repetitive “after watching several videos in a row from the same channel” [1] and Google’s framing around “many pages” [2] both point at similarity across a catalogue. Twenty pages that are individually acceptable and structurally identical are a worse problem than one bad page, and they are the pattern a classifier finds first.

What still goes wrong

Enforcement runs on classifiers, and classifiers misfire in both directions. A person who writes plainly, works to a consistent format and publishes on a schedule looks, to a pattern-matching system, a great deal like a template. None of these pages promises an appeals timeline or a number you can hit, and the remedies are described in ranges rather than thresholds. Meta’s penalty lasts “for a period of time” [5] and does not say what period.

The rules are written to be read, not measured. Meaningful, significant, generic and authentic are all load-bearing words in the policies quoted here, and not one of them has a stated threshold. You cannot calculate compliance with any of this. You can only make a case you would be comfortable putting in front of a human reviewer, which is a lower bar than it sounds and a harder one to fake than any detection tool.

These rules also have nothing to do with the AI labelling laws, and complying with one tells you nothing about the other. A disclosure that satisfies a regulator does not satisfy a ranking system, as Snapchat says in as many words [4]. And the platforms are not neutral referees of their own standard. Snapchat’s exception for AI content made inside Snapchat [4] is a reminder that authentic, as these documents use it, includes a commercial preference for the platform’s own tools. Expect that preference to grow, and expect the definition of original work to keep being written by the companies that also sell you the software for making it.