Does AI content hurt you there?Does the platform label it?Is fully AI content allowed?
Answers per platform, each linked to that platform's own policy.
Sources last checked September 25, 2026. Platform rules move fast, and the newsletter covers what changes.
38 of 38
X does not ban AI content. Its authenticity policy goes after deception:
Fake personas, unauthorized bots and impersonation are prohibited. AI-made creative is not, so the line is identity, not tooling.
Partly. Substack has a setting under Settings > Privacy: "Tell AI tools not to train their models on your content." Substack says it only applies to tools that respect the signal, and warns that blocking training may reduce how often you show up in AI-generated results.
The setting exists and only works with tools that respect it. Worth considering the trade Substack names: less training exposure, less presence in AI answers.
Only with the rights to that voice, and the terms take something back:
Consent or rights to the voice are required, and likeness laws apply on top. Worth considering keeping written permission filed with the campaign, so the rights travel with the work instead of living in somebody’s inbox.
Only on a paid plan. The free tier is licensed for personal use and nothing else:
A free-tier export in anything that makes money breaches the license, no matter who consented to the voice. Worth checking which plan the audio was actually generated on before it ships, because this is the kind of thing nobody asks about until a client’s legal team does.
Yes, on any plan including the free one. OpenAI puts it in the terms:
The license is the easy part here and OpenAI gives you everything it has. The risk sits elsewhere: whether the output is protectable at all, and whether anything confidential went into the prompt that produced it.
Yes, per Runway’s own help center: you keep ownership of what you upload and generate, and there are no non-commercial restrictions from Runway. That covers monetized YouTube, Reels and TikTok, social posts, festival entries, and product advertising.
Runway grants commercial rights. Worth checking your plan’s training terms if the footage shows unreleased work.
No. OpenAI has discontinued it, and there is a deadline that matters if you made anything there:
If anything you made in Sora is still in a live campaign, export it at sora.chatgpt.com/sunset and re-host the file yourself. OpenAI says Sora data is permanently deleted after any final export window.
Its terms grant Midjourney a perpetual, worldwide, irrevocable, royalty-free license to reproduce, adapt, display, and distribute the content you put in and the assets you make. That is a broad grant. Read it before you put a client’s unreleased product into a prompt.
The license you grant is in the terms. Worth considering: keep unreleased products and anything under NDA out of the tool entirely.
Adobe’s AI ethics page states: "We do not and have never trained Adobe Firefly on customer content." Adobe says it trains on licensed content such as Adobe Stock and public domain material, and does not scrape the web.
Adobe states it does not train Firefly on customer content. Worth considering quoting that wording directly when procurement or legal asks.
It can, and it is a setting you control:
This is a setting, not a rule. Worth considering opening your team account’s privacy controls and choosing deliberately instead of leaving it as is.
OpenAI’s terms say it may use content to provide, maintain, develop, and improve its services. Consumer accounts have a setting to opt out of model training, and business and API terms are written differently. Check the plan you are actually on before pasting client material.
Your plan’s terms decide this. Worth considering as team practice: check the plan before pasting client material, and turn training off where the setting exists.
Anthropic’s consumer terms say: "Subject to your compliance with our Terms, we assign to you all of our right, title, and interest, if any, in Outputs." As with every tool, being assigned those rights is not the same as the text being protected by copyright.
Anthropic assigns you the rights. Worth remembering that being assigned rights does not stop a competitor from generating something similar.
Yes, and Google says so in plain words:
Assume a person may read it. Worth considering the three-year reviewer copy before an unreleased product, a client list, or anything under NDA goes into a prompt, because that copy outlives the delete button.
Yes, automatically, when the whole image came from Firefly:
Adobe attaches this automatically. Worth considering flagging it before you hand a Firefly file to a client as a photo.
Mostly, with one catch worth knowing before a client hand-off:
The library-content exception is in the terms. Worth checking whether library assets are in the design before it becomes a logo or a paid ad.
Not under Canva’s rules. Its AI terms prohibit removing, altering, or disabling provenance or metadata tags on AI-generated content, C2PA included, and prohibit passing AI output off as human-made.
Canva’s terms prohibit it. Beyond the rule, it buys you nothing, since platforms run their own detection.
Yes, twice over, and neither mark is visible:
Nothing here requires action from you. Worth considering: assume every ChatGPT image is traceable, and make your disclosure match what a platform label would say anyway.
OpenAI’s terms say you keep your input and own the output, and OpenAI assigns you its rights in that output. Two catches: output is not always unique, so another user can get something similar, and owning a file is not the same as it being protected by copyright.
The terms settle ownership. Worth considering for brand work: add human contribution before something becomes an asset you may need to defend.
On consumer plans (Free, Pro, Max), Anthropic says your chats are used to improve models only when:
The setting is yours to choose. Worth considering: leave training off for client work, and use commercial plans when the data is not yours to share.
Soon, and you will not be able to see it:
No action required, but the ground has moved: text Claude drafts heavily can now be flagged as likely Claude, while text it only polishes mostly cannot. Worth considering that difference before calling a piece “human-written.”
Google applies SynthID, an invisible watermark embedded in the content itself, to media from its generative models. It is designed to survive some edits, and Google publishes a detector for it. Google’s generative AI terms cover what you may do with the output.
No action required. Worth considering: expect Gemini output to carry a detectable mark, and plan your labeling around that rather than hoping nobody checks.
Usually yes, with exceptions worth knowing before a client campaign:
The revenue threshold is a rule: over $1M a year, you need Pro or Mega to own what you make. Plan tiers are worth sorting before the work starts, not after.
Not by itself. Google has said its focus is the quality of content, not how it was produced, and that using AI to help make helpful content is fine. What gets penalized is content made mainly to game rankings.
Google’s spam policies are the rule; the tool is not. Worth considering judging a page on whether it deserves to exist at all.
Google does not put an "AI" badge on ordinary search results. Its guidance points to disclosure where it helps readers, such as saying who or what produced a page, rather than a required label.
No label to plan around. Worth considering saying who made a page where that helps a reader trust it.
Yes, with a limit:
Scaled content abuse is the line Google enforces. Worth considering that hundreds of thin pages is the fastest way to cross it.
In some places, yes, and Google may add one for you:
Required where the law requires it, and Google says plainly that its setting "doesn’t guarantee compliance with specific regulations." Worth considering labeling by where the ad runs, not by which tool made it.
Yes, when it looks real. YouTube asks creators to disclose realistic altered or synthetic content when uploading. Ordinary production help, like beauty filters or lighting fixes, does not need it.
YouTube requires the disclosure when content is realistic and synthetic, so toggle it at upload. Filters and lighting fixes do not need it. An AI voice that makes a real person say something they did not say does.
Yes. Photorealistic AI content gets a label right under the player (an overlay on Shorts). Animated or lightly altered content gets it in the expanded description. Since May 2026 YouTube also applies the label itself when its systems detect realistic AI you did not disclose.
YouTube can add the label itself. Worth considering disclosing first so the label lands where you expect.
Yes when it looks real, and TikTok often applies the label before you do:
Realistic AIGC has to carry the label. Worth knowing that the metadata does this on its own for files from tools that sign their output, so the toggle is usually confirming a label TikTok was going to apply anyway.
Allowed, with limits. It has to be labeled when realistic, and the community guidelines still bar synthetic media that misleads about real people or events.
Allowed, with the labeling rule attached. Content that fakes a real person or event is what gets pulled.
Yes. Meta applies AI labels to images it detects as AI-generated across Facebook, Instagram, and Threads, using industry provenance signals attached to the file.
Meta applies this automatically. Worth considering warning clients in advance, since edited photos can pick up the tag too.
Meta has not published a rule that AI content is downranked for being AI. The label is about provenance. Reach still comes down to the usual signals and the content rules everyone follows.
No published penalty for AI itself. Worth considering judging the creative on its own terms instead of blaming the label.
Yes, and there is one place LinkedIn does require a disclosure:
No rule against AI-assisted posts, and no general AI label. Worth knowing the one hard line: an AI clip or image of a real person needs the disclosure in the post itself, which catches the synthetic-spokesperson and fake-testimonial ideas before you hit publish.
Allowed, labeled, and in some categories users can ask to see less of it:
Nothing is banned, but Pinterest is the one platform with a live dial for users to see less of this, and it points at beauty, art, fashion and home decor. Worth weighing before an all-AI campaign in exactly those categories, and worth knowing the classifiers do not need a watermark to label you.
Yes, in the listing, and AI work also changes which category you are selling in:
Disclose in the description and list it as designed, not made. Etsy has been rewriting its seller policies through 2026, so worth re-reading the current version before a big launch.
Yes, to Amazon, every time you publish or republish:
Tell KDP at upload, and again on every republish. Worth knowing the liability for the content, copyright included, stays with you regardless of which tool wrote it.
Yes, and Steam shows a version of it to buyers:
Write the disclosure for players, since it lands on the store page. Valve’s own announcement dates from January 2024, so worth checking the survey wording at submission time.
Yes, and there is no opt-out setting. The only way out is not posting publicly:
Treat anything a brand posts publicly on Reddit as AI training data. Worth keeping product details and anything unreleased out of public threads, because no setting will pull it back.
Every answer we hold about a single platform, on one page.
Runway
Adobe Firefly
Canva