After criticism from users, Adobe Chief Product Officer Scott Belsky denies using customer projects to train the company's generative AI services
Bloomberg : Tweets: @tcpeter Tweets: Tim Peter / @tcpeter : This does lead to questions about where many generative AI projects (or *any* LLM) will get high quality training data. “The internet” is not necessarily a good answer. Plenty of examples exist where using internet data made AI's dumber https://twitter.com/...
Context & Ripple Effects
Scott Belsky's denial lands mid-criticism: users had flagged that Adobe's services could touch their work, and his response — that customer projects are not used to train Adobe's generative AI — is an attempt to close the question at the product-leadership level rather than through legal text. The exchange also surfaces the harder problem Tim Peter raises: where any LLM or generative model gets high-quality training data, since 'the internet' has repeatedly made models worse.
The denial did not end the scrutiny. Later coverage shows the same tension recurring: Adobe was found to have trained Firefly partly on images generated with rivals' tools like Midjourney that users uploaded to its stock marketplace, then faced a backlash over terms of service allowing automated and manual access to user content, before finally issuing an [[a:867066|explicit clarification that it does not train Firefly on customer content and will never claim ownership of customer work]].
First-order effects
- Adobe's creative users get a direct assurance from its Chief Product Officer that their project files are off-limits for generative AI training — but the assurance is verbal, not yet written into terms they can rely on.
- Belsky's statement puts Adobe's own training-data pipeline under immediate examination, since the company must now reconcile the denial with what actually feeds Firefly.
Second-order effects
- Rivals in creative software gain a positioning wedge: any competitor that contractually guarantees customer content stays out of training data can attack Adobe on trust rather than features.
- Adobe is pushed toward codifying data-use promises in its terms of service — which is exactly where the later backlash and clarification cycle played out.
Third-order effects
- If the pattern holds, training-data permissions become a standard contractual clause across SaaS: vendors that monetize user content for model training face churn pressure, forcing an industry-wide split between platforms that train on customer data and those that explicitly don't.
- The deeper constraint Belsky's denial points to is supply: as clean internet-scale data proves unreliable, proprietary and licensed corpora become the contested resource, and who holds rights to high-quality creative work becomes a structural advantage.
The trend: Generative AI vendors are being forced from informal assurances toward explicit, enforceable guarantees about whether customer content trains their models, as trust over data use turns into a competitive fault line.