WeTransfer removed the machine-learning clause from its Terms of Service after user backlash (July 2025)

The problem

WeTransfer, the consumer file-sharing service, quietly updated its Terms of Service in early July 2025. Sometime after 2 July it rewrote clause 6.3 — Malwarebytes Labs reported that “Sometime after July 2, it updated clause 6.3 of the document” [source] [Source 3]. The added language granted WeTransfer a “perpetual, worldwide, non-exclusive, royalty-free, transferable, sub-licensable license to use” ([source]) user Content [Source 3]. In mid-July, creatives and other users read the clause as permission to train AI on the files they upload, and the criticism spread across social media.

What was built

This is a Terms-of-Service reversal, not a deployment. In its own blog post WeTransfer quoted the disputed licence, which extended to “improve performance of machine learning models that enhance our content moderation process” [source] — the same clause language The Register also reported [Source 1, Source 2]. The revised terms were set to take effect on 8 August 2025 (a date carried by the press, not the primary — see the verification box) [Source 2].

The outcome

• The clause was rolled back. After the backlash WeTransfer reversed the change. In its first-party blog post the company said it had “removed the mention of machine learning” [source] [Source 1]. The Register reported the company “rolled back changes it had introduced to its Terms of Service after they deeply upset users” ([source]) [Source 2], and Malwarebytes reported that WeTransfer had “rolled back language that allowed it to train machine learning models on any files that its users uploaded” [source] [Source 3].
• WeTransfer’s framing. In the same blog post WeTransfer stated firsthand that it does not use “machine learning or any form of AI to process content shared via WeTransfer” [source] and added: “We apologize for any confusion this may have caused” ([source]) [Source 1]. The Register separately quoted the company saying it does not use “any form of AI to process content shared via WeTransfer” [source] [Source 2]; per Malwarebytes the company said the clause “may have caused confusion for our customers” ([source]) and that it does not “sell content to third parties” [source] [Source 3].

The weakest load-bearing source, named where you meet it: WeTransfer’s own blog post (Source 1) is a Tier-1 first-party primary, but it is the subject narrating its own reversal — an admission-against-interest (it removed the clause and apologised), not an independent audit. It does not stand alone: the rollback itself is independently reported by The Register (Source 2, Tier 2) and Malwarebytes Labs (Source 3, Tier 2), the two independent secondaries around the first-party origin. WeTransfer’s counter-claim — that the machine-learning reference was only ever intended for possible future AI-powered content moderation, not a built or used feature — is recorded here as the company’s position, not asserted as fact.

Why this matters for TIN: an AI-training permission read into a licence clause, then withdrawn under pressure, is exactly the kind of dated, publicly refutable reversal the verified-story model is built to pin down — the record supports the walk-back; nothing here depends on WeTransfer confirming it privately.

Related case files

• Zoom reversed its Terms-of-Service AI-training clause after backlash — the same shape two years earlier: a licence-clause rewrite read as consent to train AI on customer content, withdrawn under public pressure.
• Dropbox walked back its third-party-AI (OpenAI) toggle — another file-storage service forced to clarify, publicly, that user files were not being used to train models.
• Adobe added an AI-training clarification to its Terms of Use after backlash — a creative-tools platform making the opposite move, adding explicit language to quell the same fear the WeTransfer clause raised.
• Slack updated its privacy principles after an ML-training opt-out backlash — a SaaS sibling in the same vein, where the dispute turned on default machine-learning use of customer data.