Video hosting without AI training. A clause in the contract, not a toggle in the settings.
Most platforms take a licence to your uploads that is wide enough to train models on them, then offer an opt-out buried three menus deep. Here is how that licence works, what Quebec's privacy law does and does not cover, and the sentence in our terms that closes the door.
Your files are not training data, and it is not a setting.
Every hosting platform needs a licence from you to do its job: copy the file to its servers, transcode it if it transcodes, stream it to the people you invite. The whole question is how wide that licence is, and whether it can be widened later without asking.
Ours is narrow on purpose, and the no-training part is spelled out. From the terms of service, in the section on your content and your rights:
The clause, verbatim
No AI training, no datasets. We do not use your content, your files, frames, audio, captions, comments or the metadata that travels with them, to train, fine-tune, evaluate or improve any machine-learning or artificial-intelligence model, ours or a third party’s. We do not build datasets from it, derive works from it, sell it, license it, or analyze it for any purpose other than one you trigger yourself. This is a contractual obligation, not a setting: it cannot be changed by a quiet update to these terms. Any change to this paragraph will be announced to you by email at least 30 days in advance, and you will be able to delete your content before it takes effect. The limited licence described above is the only licence you grant us, and it includes no training right.
Three things make that different from a privacy checkbox. It is a term of the contract you accept at signup, so breaking it is a breach with remedies, not a policy lapse. It binds the company, not the account holder, so there is nothing for you to switch on or keep switched on. And it carries its own notice rule, so the mechanism every backlash on this page shares, a wider licence arriving inside a routine terms update, is closed in advance.
How the wide licence works
“To develop and improve our services” is the standard clause.
The licence in a typical consumer or prosumer platform’s terms is worldwide, royalty-free, sublicensable, and granted for purposes that include operating, developing and improving the service. Nobody hides those words; they simply do not say what they permit.
A licence to use your content to “improve our services” is, read plainly, a licence to use your content as input to whatever the company builds next, and since 2023 what most companies build next is a model. When customers noticed, the pattern repeated almost to the word: an updated terms page, a wave of professionals reading it for the first time, a blog post within days explaining that the company “does not train on customer content”, and in some cases a rewrite of the clause. The clarification is welcome each time. It is also a blog post, and a blog post is not a term of the contract.
Adobe, June 2024
A terms pop-up required users to agree that Adobe 'may access your content through both automated and manual methods', using 'techniques such as machine learning'. Within days Adobe posted that it does not train its generative models on customer content, and by 18 June had rewritten the terms. Its current terms now say so outright: the licence 'does not give us permission to train generative AI models with your or your customers' content'. That sentence, in the contract, is the right ending. It took a backlash to get it.
WeTransfer, July 2025
New terms, announced for 8 August, granted 'a perpetual, worldwide, non-exclusive, royalty-free, transferable, sub-licensable license to use your Content for the purposes of operating, developing, commercializing, and improving the Service or new technologies or services, including to improve performance of machine learning models'. On 15 July WeTransfer deleted the machine-learning language and told the BBC it does not use AI to process shared content.
Zoom, August 2023
Terms in force from 27 July 2023 allowed customer content and service-generated data to be used for machine learning and AI, including training. On 7 August Zoom added that it 'will not use audio, video, or chat Customer Content to train our artificial intelligence models without your consent', and on 11 August rewrote the section.
Dropbox, December 2023
Users found a 'Third-party AI' account setting switched on, sending file content to a partner model when AI features were used. Dropbox's CEO said the data goes out only when a customer actively uses those features and that 'any customer confusion about this is on us'.
Slack, May 2024
The privacy principles read: 'To develop AI/ML models, our systems analyze Customer Data (e.g. messages, content and files) submitted to Slack', with opt-out only by a workspace owner emailing Slack. Within a day the page was rewritten to separate generative features, which now need opt-in, from the recommendation models, which still read customer data unless you opt out.
Vimeo, since May 2024
Vimeo's CEO wrote that Vimeo 'will not allow generative AI models to be trained using videos hosted on our platform without your explicit consent'. A good promise. It lives in a blog post and a help-centre article; the terms of service grant the usual stream, transcode and download rights and do not contain it.
YouTube, ongoing
The terms grant YouTube a worldwide licence to use your content, including derivative works, 'in connection with the Service and YouTube's business'. In December 2024 YouTube added a setting letting creators allow named third parties to train on their videos, off by default; Google's own training on YouTube content continues under the existing licence.
The pattern
Every one of these clauses was legal and in force, and the reversals came because someone read the terms and said so in public. The fix, when it came, was a blog post or a settings toggle. Adobe is the one that went on to put the promise in the contract.
For a post house or an agency, the exposure is not abstract. Unreleased spots, dailies, cuts under embargo and anything covered by a client NDA are uploaded to review platforms every day. A platform whose licence includes “improve our services” has, on paper, permission to use those files in ways the NDA never contemplated. Whether it ever does is a separate question; the point is that you signed the permission.
Law 25 and PIPEDA, in plain English
The law protects the people in your footage, the contract protects the footage.
uncompressed.io is a Quebec company, so Quebec’s private-sector privacy law (Law 25) and Canada’s federal PIPEDA apply to it. It is worth being exact about what they protect, because the claim “the law forbids us from mining your footage” is doing the rounds and it is not true as stated.
Both laws regulate personal information: information about an identifiable person. In a review workflow that covers your account details, the client emails you invite, the viewer names and session data burned into watermarks, and the recognisable faces and voices inside the footage. Under the Act that information can only be collected for a purpose stated in advance (sections 4 and 5), used for that purpose unless the person consents to another (section 12), and communicated outside Quebec only after a privacy impact assessment and under a written agreement (section 17). The Commission d’accès à l’information can impose administrative penalties up to the greater of $10 million or 2% of worldwide turnover (section 90.12); penal fines run from $15,000 to the greater of $25 million or 4% (section 91), doubled on a repeat offence. Since 22 September 2023 a court must award punitive damages of at least $1,000 when an intentional or grossly negligent infringement causes injury (section 93.1).
What the law does not do is treat your master file as personal information. A commercial for a car, a colour-graded feature reel, a drone plate: those are intellectual property, and privacy law is silent on whether a host may run them through a model. The protection there comes from copyright, which you keep in full, and from the licence you grant the platform, which is where the clause above does its work. Put the two together and the honest sentence is this: the law makes us handle the people in your workflow with care and gives them recourse; the contract makes us handle the footage the way you expect and gives you recourse.
Covered by Law 25 and PIPEDA
Account and billing details, invited viewers' emails, watermark identities, session logs, and identifiable people inside the footage.
Covered by the contract
The footage itself: one streaming licence, no training right, no datasets, no derivatives, no sale, 30 days' notice on any change.
Covered by copyright
Ownership. You keep every right you had before you uploaded, and the licence ends when you delete the file.
One more fact worth having in the file for your own compliance team: the European Commission has recognised since 2001 that Canada provides adequate protection for personal data transferred from the EU to organisations subject to PIPEDA, and its January 2024 review kept that decision in place. It is a statement about personal data, not about footage, and we use it only that way.
Where the files live
The precise answer, including the parts that are not flattering.
Masters are stored in object storage in an Eastern North America region, encrypted at rest, and streamed to your viewers from there without transcoding. We do not call that “Canadian infrastructure”: the region covers eastern Canada and the eastern United States, and a page about honest claims should not open with a fuzzy one. The company, the contract and the governing law are Quebec; the bytes sit in a North American region operated by a US-headquartered provider whose agreement licenses customer content only to the extent needed to provide the storage, and whose published policy is that customer content is not used to train its models without consent. That provider offers no Canada-only storage option, which is why nobody using this class of storage can honestly claim one. We do not name our providers publicly, as a matter of security practice; what we state is the region, the encryption, and the contract.
1
licence you grant
store and stream to the viewers you choose
0
training rights
in the terms, for us or anyone else
30 days
notice before that changes
with time to delete first
1 case
where audio leaves for a model
captions, only when you click Generate
The one exception is written down so it cannot surprise anyone. Clicking “Generate captions” on a video sends that video’s audio track to a speech-to-text model and returns subtitles you can edit. Nothing is sent unless you click, no other file is involved, and the audio is not retained for training. Everything else, upload, storage, review, watermarking, delivery, involves no AI service at all.
What the Vault adds
The contract handles what we may do with it; the Vault handles who can see it.
A no-training clause answers one question: what the platform does with your file. The other question, who else can get at it, is engineering. On the Vault tier the master goes dark on the web and is reviewed only in the macOS app, encrypted at rest on the viewer’s machine, with screen capture blocked, a per-session watermark on the frame and an invisible per-person forensic mark that survives re-encoding and a phone pointed at the screen. Access is revocable in one click and every open is logged. The two layers are meant to be read together: one is a promise about us, the other is a wall against everyone else.
Encrypted at rest, app-only playback
No web player, no browser download, no embed. The raw master never reaches a browser.
Per-viewer forensic marks
A Tardos-code fingerprint unique to the recipient, traceable even when up to three copies are combined.
Revocable in one click
Signed links expire in minutes; access can be pulled back mid-review.
Published price
$250 a month, seats $175, no enterprise contract and no sales call to find out what it costs.
Buying advice
Four questions for any platform, before the next upload.
1
Which licence do I grant, exactly?
Find the sentence. If it contains 'improve our services', 'develop new products' or 'machine learning', that is a training licence, whatever the marketing page says.
2
Can it change without notice?
Most terms reserve the right to update at any time with continued use as acceptance. Look for a notice period and a right to delete before a change takes effect.
3
Is the no-training promise in the contract or in a setting?
A toggle protects the accounts that found it. A clause protects every account, including the intern's.
4
Where, precisely, are the files?
Ask for the region, the encryption and the contract terms. 'Secure cloud' is not an answer; 'object storage, Eastern North America, encrypted at rest, licence limited to providing the service' is.
A claim you will hear, and why we do not make it
“We are legally prohibited from using your footage.” No privacy law in Canada says that about a video file. If a vendor tells you the law forbids them from training on your masters, ask them to cite the section. Then ask to see the licence clause in their terms, which is where the answer actually lives.
Questions
Frequently asked
Does uncompressed.io use my videos to train AI?
No, and it is written into the terms of service as a contractual obligation: no training, fine-tuning, evaluation or improvement of any machine-learning model, ours or a third party's; no datasets, no derivative works, no sale, no licensing. The only licence you grant is the one needed to store and stream your files to the viewers you choose. Any change to that paragraph is announced by email 30 days ahead, with time to delete your content first.
Does Quebec's Law 25 make it illegal to train AI on my footage?
Not by itself. Law 25 and PIPEDA regulate personal information, meaning information about an identifiable natural person: your email, the names burned into watermarks, and any face or voice in your footage that identifies someone. They do not regulate your master file as intellectual property. What binds a platform on the footage itself is its contract with you. That is why the clause in our terms matters more than any law, and why a platform whose terms include a training licence is not protected by being 'compliant'.
Where are my files physically stored?
In object storage in an Eastern North America region, encrypted at rest. uncompressed.io is a Quebec company and the terms are governed by Quebec law. We do not describe this as 'Canadian infrastructure', because the region spans eastern Canada and the eastern United States and we will not claim a precision the storage does not have. We do not name our infrastructure providers publicly; the region, the encryption and the contractual position are stated instead.
Do any of my files ever reach a third-party AI service?
Only when you ask for it. Clicking 'Generate captions' on a video sends that video's audio to a speech-to-text model to produce subtitles, and nothing else. Uploading, streaming, review, watermarking and delivery involve no AI service at all.
Can the terms change later, the way other platforms' terms did?
Any change to the no-training paragraph requires 30 days' written notice by email, and you can delete your content before it takes effect. A change that quietly appears in a new version of the terms, the pattern behind every backlash listed on this page, is exactly what that sentence forbids.
What should I ask any other platform before uploading unreleased work?
Four questions: which licence do I grant you, and does it include 'improve our services' or 'machine learning'; can that licence be changed without notice; is the no-training promise a contract term or a settings toggle; and where, precisely, are the files stored. A vendor who answers all four in writing is rare.
Move the library, keep the rights
One licence: store and stream your files to the viewers you choose. No training right, no datasets, no derivatives, no sale. Start free with 1 GB, no credit card.