Key takeaways
- The third-party training setting is off by default and lives in Studio under Settings → Channel → Advanced settings. It is a channel-level decision applied to eligible public videos, not a per-video choice.
- It governs outside AI companies only. Google told TechCrunch at launch that it would carry on training its own models on some YouTube content under its existing agreement with creators, and CNBC reported in June 2025 that Gemini and Veo 3 had been trained on a subset of the library.
- Saying yes publishes a flag. YouTube exposes a per-video trainability endpoint that returns all, none or a list of permitted parties, with changes taking up to seven days to appear.
- There is no documented payment for opting in. The money in AI training data is in a separate market for unused footage, which Bloomberg put at roughly $1 to $4 per minute in January 2025.
- You can only grant what you control. Guests, licensed music, stock clips and client work all sit inside videos whose training rights are not entirely yours to give away.
- None of this protects you from impersonation, and none of it mentions thumbnails. Those are separate systems with separate forms.
Somewhere in your channel settings there is a checkbox that decides whether a list of the largest AI companies in the world is allowed to train on your back catalogue. It has been there since December 2024. Most creators have never opened the page, which is a reasonable response to a switch that offers no money, no analytics and no obvious consequence either way.
It is still worth ten minutes of your attention, for a reason that has little to do with the checkbox itself. Working out what it does forces you to find out what was already happening without it — and the gap between the two is the actual story. The setting is narrow, honest about its own limits, and surrounded by practices it does not touch.
This is what the control does, what it provably does not cover, what the record shows about training that happened before anyone was asked, where the paid version of this transaction lives, and how to decide for your own channel. Where a number appears below, it comes from a named source; where nobody has published one, there is no number.
Where the setting is and what it says
On a standard channel, the path is Studio → Settings → Channel → Advanced settings. There you will find an option to let third-party AI companies use your YouTube content to train models they are developing. Tick it and you are asked a second question: all third-party companies, or only ones you select from a list.
YouTube's documentation on third-party training is unusually plain about the mechanics. The setting is off by default. It can be changed at any time. It applies to eligible videos uploaded to your channel, where eligibility means the video is public, complies with YouTube's Terms of Service and Community Guidelines, and has the permission of its rights holders. Multi-channel operations run the same decision from the Settings page in Studio Content Manager, where only the Administrator role can turn it on — one more reason to know exactly who holds which role on your channel, which we covered in the guide to YouTube channel permissions.
Two things the interface does not make obvious. There is no per-video control: the choice is made once, at channel level, and inherited by everything eligible. And the "all third-party companies" option is genuinely open-ended — it extends to companies that are not on the list, including ones that do not exist yet. YouTube also says companies building AI models can apply to be added to the list, so the roster is a moving target. Whatever was accurate when any article was written, including this one, read the list in your own Studio before you tick anything.
When TechCrunch covered the launch on 16 December 2024, the list ran to eighteen companies. The seventeen it named were AI21 Labs, Adobe, Amazon, Anthropic, Apple, ByteDance, Cohere, IBM, Meta, Microsoft, Nvidia, OpenAI, Perplexity, Pika Labs, Runway, Stability AI and xAI. That roster is a useful snapshot of who was asking in late 2024, not a current inventory.
The sentence that defines the whole feature
YouTube's own framing contains the limit: the setting controls third-party training. Google is not a third party to YouTube.
At launch, Google told TechCrunch that it would continue to train its own AI models on some YouTube content in accordance with its existing agreement with creators, and that the new setting did not otherwise change the Terms of Service. TechCrunch also reported that YouTube was unable to say whether the setting had any retroactive effect on third-party training that had already happened.
In June 2025, CNBC reported that Google was using YouTube videos to train Gemini and Veo 3, drawing on a catalogue it described as 20 billion videos, and that Google said it used only a subset and honoured specific agreements with creators and media companies. The part of that reporting creators reacted to was not the training but the awareness gap: CNBC said none of the leading creators and intellectual property professionals it spoke to had been told by YouTube that their content could be used this way. YouTube's position was that it had disclosed the practice previously.
So the honest description of the switch is this: it is a control over everyone except the company that owns the platform. That is not a scandal, and it is not hidden — it is written into the help page. It is simply the first thing to understand, because it determines what a decision here can and cannot achieve. Opting out of third-party training does not take your work out of AI training in general. It takes it out of a specific, named, voluntary programme.
The permission you may not own
Eligibility requires that the rights holders of the video permit third-party training. On a solo talking-head upload, that is you. On an interview it includes your guest. On anything carrying licensed music, stock footage, game capture under a publisher's terms, or work made for a client, the training rights are split across parties who never saw your settings page. The checkbox does not ask you to confirm this. It assumes it.
Saying yes publishes a public flag
The least discussed part of this system is the most technically interesting. Your answer is not a private preference held inside Studio — it is readable by anyone.
Google documents a YouTube Video Third-Party Trainability API: a single get method, called with a video ID, which returns a permitted list. The value is either all, none, or the specific parties you allowed. The help documentation adds that changes to your setting can take up to seven days to be reflected in that publicly accessible interface.
That design tells you what the feature is for. It is a machine-readable licence signal, built so that a company wanting to assemble a training set can ask, video by video, whether it has permission — and so that the absence of permission is documented rather than ambiguous. The seven-day lag matters in one direction only: turning the setting off is not instant from the outside world's point of view, so this is not a switch to flip on, test and flip back.
It also means the choice is not invisible. A per-video flag reading all is a public statement about your channel's posture, available to anyone who looks — researchers, journalists, competitors, and any future buyer of your channel.
What was already happening before the checkbox existed
A consent control introduced in December 2024 arrived after several years of training that did not involve consent at all, and the record of that is public.
In July 2024, Proof News published an investigation, co-published with Wired, into a dataset called YouTube Subtitles — a slice of The Pile, the large text compilation assembled by the nonprofit EleutherAI. Proof found subtitles from 173,536 videos drawn from more than 48,000 channels, and identified companies that had described using The Pile in their own research papers and posts, among them Anthropic, Nvidia, Apple and Salesforce. More than 12,000 of those videos had been removed from YouTube since the dataset was created in 2020, and their transcripts were still in it. Anthropic's spokesperson described the material as a very small subset of YouTube subtitles; companies that responded denied wrongdoing.
Three months earlier, in April 2024, YouTube's chief executive Neal Mohan had told Bloomberg that if OpenAI had trained Sora on YouTube videos it would be a clear violation of the platform's terms of service, while saying he did not know whether it had. A YouTube spokesperson later confirmed to reporters that the terms prohibit unauthorised scraping or downloading of YouTube content.
Put those together and the shape of the problem is clear. Scraping was already against the rules; the rules were not obviously stopping it; and the party with standing to complain about a terms-of-service breach is YouTube, not you. Which is exactly why the next section exists.
The courts are still deciding what the default was
Creators have taken the question into court, and the claims have shifted in a way worth noticing.
In August 2024, a YouTube creator named David Millette filed a proposed class action against OpenAI in the Northern District of California, alleging that creators' videos had been transcribed to train ChatGPT; reporting on the complaint described more than a million hours of video obtained by automated means. A separate Millette suit followed against Nvidia. Notably, neither alleged copyright infringement — they were pleaded as unjust enrichment and unfair competition, which tells you how uncertain the copyright theory looked at the time.
The later filings are more aggressive. On 26 November 2025, the entities behind the channels H3H3 Productions, Mr.ShortGame Golf and Golfholics filed a putative copyright class action against Nvidia in the Northern District of California, accusing it of circumventing technological protection measures to scrape videos for AI training. In April 2026, Law360 reported proposed class actions in Seattle and California federal courts in which YouTube creators made similar anti-circumvention claims under the Digital Millennium Copyright Act against Amazon, OpenAI and Apple.
The anti-circumvention framing is the interesting move: rather than arguing about whether training is fair use, it argues about how the material was obtained. None of these cases has produced a ruling that settles the question. What they establish, for a creator deciding what to do this week, is that the legal default is genuinely unsettled — and that a documented no on your channel is a cleaner starting position than silence if any of it ever matters to you.
Who can train on your videos, and what you can actually do about it
| Party | What the Studio setting does | What you can do |
|---|---|---|
| Google, for its own models | Nothing — the setting covers third parties | Nothing within Studio. Google cites an existing agreement with creators |
| Listed third-party companies | Grants or withholds permission per company | Select individually; change at any time, with up to a seven-day lag publicly |
| Unlisted third parties, via the "all" option | Grants permission to any company, listed or not | Leave it unticked if an open-ended grant is not what you mean |
| Companies that scrape without asking | Nothing — they were already in breach of the terms | Rely on YouTube's enforcement, or litigation; the past cases are creator-led |
| Buyers of licensed footage | Unrelated — a separate commercial transaction | Negotiate directly or through a footage marketplace |
Read down the last column and the practical weight of the decision becomes clear. The row you control cleanly is the second one. The first and fourth rows are where most of the anxiety lives, and the checkbox does not reach either.
What you get for saying yes
As far as any published source shows: nothing. No payment, no revenue share, no early access, no analytics, no distribution benefit. Coverage of the launch noted the absence of compensation plans, and nothing since has established one. YouTube's documentation is also careful to say that a company you authorise may not end up using your content at all, and that because third-party companies are separate from YouTube and Google, it ultimately cannot control what they do with it.
That is an unusual deal to be offered: a grant of rights, for free, to a counterparty whose conduct the intermediary disclaims responsibility for. There are still reasons a creator might take it. If you believe better models are good for your own work, contributing is coherent. If your content is instructional and you would rather assistants answer questions accurately from your material than inaccurately from someone else's, that is a position — and it rhymes with the reasoning behind making a channel legible to AI-driven discovery, which we worked through in the post on YouTube AI search visibility. If you have a specific relationship with a specific company on the list, selecting only that company is a way to formalise it.
What is not a reason is the hope of being paid later. Nothing in the setting creates a claim. It is a licence grant, not a contract with you as a party.
The paid door is a different door
There is real money in AI training footage, and it has not been flowing through this checkbox.
Bloomberg reported in January 2025 that AI companies were paying roughly $1 to $4 per minute for footage, with most unused material created for platforms such as YouTube, Instagram and TikTok selling at the lower end, between $1 and $2 per minute, and prices rising for 4K, drone and 3D-animated content. The intermediaries in that market deal in unpublished material — the B-roll, the alternate takes, the hours that never made the cut. Troveo, one of the brokers in that reporting, said it had paid creators over $5 million, a figure from the company itself, and that its contracts barred AI firms from making digital copies of creators or recreating their scenes.
Note what is being sold there, because it is the opposite of what the Studio setting gives away. The licensing market pays for exclusivity and unpublished supply, under contract, with terms about likeness. The checkbox offers published work, non-exclusively, to a list of companies, with no terms you negotiated. If the money matters to you, the footage market is the conversation to have; the Studio setting is not a negotiation at all.
Before you sign anything with a footage broker
The per-minute rates above are what reporting says buyers have paid, not a price you are owed. Any deal is worth reading for three things in particular: whether it is exclusive, whether it permits synthetic recreation of you or your location, and whether it covers material in which other people hold rights. A contract that bars digital copies of the creator is a feature, not boilerplate.
What opting in does not do to your channel
Three fears come up, and all three are answered by what the documentation covers rather than by speculation.
It does not change your monetisation or your reach. Nothing in YouTube's documentation ties the training setting to the Partner Programme, to ad revenue, or to recommendation. If anyone tells you that opting in boosts distribution or that opting out suppresses it, ask them for the source; there isn't one.
It does not protect you from impersonation. A model trained on your videos and a scam advert wearing your face are different problems with different remedies. The tool for the second one is likeness detection, which is opt-in, identity-verified and entirely separate from this setting — and which, as the guide to YouTube likeness detection sets out, finds matches rather than removing them.
And it is silent on thumbnails as such. The documentation and the API both talk about videos, by video ID; neither addresses a thumbnail as a standalone image, and a thumbnail does not stay inside YouTube — it is in search results, in the Data API, and in every aggregator and embed that has ever cached your channel. Whatever the setting says, the copy of that image sitting on somebody else’s server is outside this system. If someone takes a thumbnail outright, your route is a copyright complaint, which is the subject of the post on YouTube thumbnail copyright.
How to decide, by channel type
| Channel | Case for allowing it | Case against |
|---|---|---|
| Face-led personality or vlog | Little, unless tied to a named relationship | Your face, voice and mannerisms are the asset; synthetic substitutes compete directly with you |
| Instructional or reference library | Your explanations reaching people through assistants, attributed or not | Assistants answering the question is the reason someone does not need your video |
| Faceless, narration over stock | Low personal exposure; the format is already commoditised | You frequently do not hold the rights to grant, which makes eligibility doubtful |
| Client or agency work | Essentially none | The rights are the client's; granting them is not yours to do |
| Music, gameplay or licensed footage | None | Labels, publishers and platform holders hold overlapping rights |
If a single rule helps: the more of your value sits in a person, and the more of your library contains other people's rights, the less there is to gain from ticking the box. The strongest case for opting in is narrow and deliberate — one company, for a stated reason, on a library you wholly own.
The weakest decision of all is the accidental one. A channel with several managers, an agency with Content Manager access, and nobody who has read the page is the configuration most likely to end up granting an open-ended licence by mistake.
The ten-minute audit
- Open Studio → Settings → Channel → Advanced settings and look at the third-party training section. Record what it currently says, rather than assuming it is off because you never changed it.
- If you run Content Manager, check the Settings page there too, and confirm who holds the Administrator role.
- Decide between three postures: off, specific companies, or all. Write down the reason, so the next person with access does not undo it on a hunch.
- If you allow anything, give it a week before expecting the public flag to match, and check a video ID through the trainability endpoint if you want to verify it.
- While you are in Studio, look at content detection for likeness detection enrolment and at your AI disclosure settings for altered or synthetic content. Different systems, same afternoon.
- If you take guests, add a line to your release covering AI training rights. It costs nothing now and saves an awkward email later.
Where this leaves the work
The reason to spend ten minutes on a checkbox with no money attached is not the checkbox. It is that going to look at it is the only way to find out what your channel's current posture actually is — and most channels discover that the answer was set by inattention rather than by a decision.
What follows from the decision is smaller than the surrounding noise suggests. Models already trained on scraped material are not untrained by your setting. Google's own use sits outside it. The one thing you genuinely control is whether a named list of companies has your documented permission going forward, and whether that permission is open-ended. Those are worth getting right on purpose.
Then go back to the part of the job that still decides whether anybody sees the video. The packaging is yours, it is not a policy question, and it is the only lever in this entire discussion with a measurable effect on views this week. If you want the thumbnail side of the AI question — what is allowed, what has to be disclosed, and where generated imagery helps rather than hurts — the post on whether AI thumbnails are allowed on YouTube covers it, and the roundup of Made on YouTube 2026 covers the tools YouTube is building in the same direction. When you want to make the thumbnail rather than read about it, Thumblore generates options in your channel's style, and you keep control of which one ships.
Frequently asked questions
Is the setting on or off by default? Off. YouTube's documentation states the third-party training setting is turned off by default, so a channel nobody has touched is not granting permission through this control.
Can I stop Google training on my videos? Not through this setting. It covers third-party companies; Google has said it continues to train its own models on some YouTube content under its existing agreement with creators.
Do I get paid for opting in? No published source shows any compensation attached to the setting. Paid arrangements exist in a separate footage licensing market, usually for unpublished material and under contract.
Can I allow one company and refuse the rest? Yes. After ticking the option you choose between selected companies and all third-party companies, and you can change the selection at any time.
Does it apply retroactively? YouTube was unable to say at launch whether the setting had any retroactive effect on training that had already taken place. Treat it as a decision about the future.