Adobe Didn't Steal Your Art. It Just Reminded You It Could.

A designer opens Photoshop on a Tuesday, mid-deadline, and a box blocks the canvas: accept the new terms to continue. Buried in the legalese is a line — 'We may access your content through both automated and manual methods, such as for content review.' Read at speed, in a year when every creative tool seems to be quietly feeding an AI, that line lands like a confession. One screenshot, one tweet, and the fuse is lit. The post crosses 5 million views and 50,000 likes, and within days the consensus across the creative internet is settled: Adobe is using your private files to train its AI.³

Here is the uncomfortable part. None of it was true. Adobe did not change its terms in June to grab content for AI. Firefly was never trained on subscribers' files. The inflammatory clause wasn't even new. And the backlash damaged Adobe anyway — because the story was wrong about the facts and right about the relationship.

The official narrative was that vigilant creators caught Adobe in the act. The real story is the opposite: Adobe got caught not doing the thing everyone accused it of, in language it had been living with for months. A covert AI data-grab, exposed. A self-inflicted communications failure, detonated by a trust deficit Adobe had spent years building elsewhere.

The clause everyone screenshotted was already old

Start with the timeline, because the timeline is the whole case. The machine-learning provisions and the 'automated and manual methods' language were not invented in June 2024. As contemporaneous reporting noted, 'the inclusion of machine learning is not new and has been part of the TOS for years,' and the specific updated wording had been live since February 2024 — Adobe simply hadn't prompted most users to re-accept it until early June.⁶ What detonated in June wasn't a change. It was a notification. Adobe took text that had sat dormant for four months and pushed it into a mandatory accept-to-continue box, in front of millions of creators, with no plain-language explanation attached. The substance was old; the surfacing was new; and the surfacing was the mistake.

The clause itself was boilerplate operational licensing — the kind of 'non-exclusive, royalty-free license to reproduce and display content solely to operate the Services' that every cloud tool needs to so much as render your thumbnail. Necessary, mundane, and catastrophically easy to misread as 'we own your work.' Adobe asked an anxious audience to sign a blank-looking check, and the audience filled in the worst number it could imagine.

Firefly's denial was true — and the truth didn't help

On the central accusation, Adobe had documentation on its side and said so loudly. Its Firefly FAQ states the models were trained on licensed content such as Adobe Stock and public-domain material, not on subscribers' personal files.¹ Its business product page is even blunter: 'We do not train our Firefly generative AI models on any Creative Cloud or Adobe Experience Cloud subscribers' personal content.'² Adobe had even paid bonuses to the Stock contributors whose licensed images did feed Firefly — a paper trail that pointed away from the very theft it was being accused of.¹⁰ By the metric of 'is the accusation factually correct,' Adobe was clean.

It didn't matter, and the reason it didn't matter is the strategic core of this episode. Trust isn't audited; it's felt. When a customer already suspects you of squeezing them, they don't read your FAQ — they read your prompt box, at speed, through the lens of every other grievance they hold. Adobe's denial was accurate. The relationship it was speaking into was not listening.

Why a false story stuck to a true denial

A rumor needs scaffolding to climb, and Adobe had built two beams of it in plain sight. The first was lock-in: a subscription model with no perpetual-license exit, so the natural response to 'I don't trust this company' — leaving — felt prohibitively expensive. People who feel trapped read every new clause as the trap tightening. The second beam arrived almost simultaneously. In the same June, the FTC sued Adobe and two executives, alleging it hid steep early-termination fees — roughly half the remaining contract value — on annual-paid-monthly plans and made cancellation deliberately hard.⁸ Adobe disputes the claims. But for a creator deciding whether Adobe was the kind of company that would quietly take their work, the FTC complaint answered a different question with terrible timing: yes, this is a company comfortable extracting from customers who can't easily walk away.

So the AI story didn't have to be true. It only had to be consistent with the company creators already believed they were dealing with. The data-grab that never happened slotted neatly beside the cancellation fees that allegedly did. That is how a false accusation lands a real wound: it borrows credibility from your actual reputation. Adobe wasn't punished for what it did with Firefly. It was punished for who it had already shown itself to be on the invoice.

Wasn't the rewrite a real fix?

The fair objection is that Adobe responded fast and well. Within days, Belsky — Chief Strategy Officer, not the CEO, a distinction worth noting because the company's most senior leader never became the face of it — co-authored a public response, conceded the wording was confusing, and wrote that 'we should have done this sooner.'⁵ On June 24 Adobe shipped a rewritten Terms of Service with plain-language clarifications about ownership and AI training.⁴ By the textbook, that's a competent crisis response.

But the honest counter cuts deeper. Critics noted that Adobe's reassuring 'plain language' clarifications sat in explanatory gray sections — not inside the legally binding contract text, which still granted Adobe broad operational rights to reproduce and display content; Adobe's own updated terms carried a disclaimer that 'only the terms themselves are legally binding, not these summaries.'¹³ The comfort was in the marketing layer; the power stayed in the enforceable layer. And the training-data story had a second crack: a Bloomberg report indicated roughly 5% of Firefly's images were themselves AI-generated, some made with rival tools, because Adobe Stock accepts labeled AI submissions — quietly undercutting the spotless 'commercially safe' narrative.¹¹ So the rewrite calmed the panic without closing the gap that caused it. It treated a trust problem as a wording problem. The wording was never really the problem.

Adobe spent the summer of 2024 proving it hadn't done the thing it was accused of, and discovered that being innocent is no protection when you've spent years looking guilty of something adjacent. The clause was old. The denial was true. The rewrite was fast. And still the trust leaked, because trust is the one asset that doesn't read your FAQ. The architect of much of Adobe's creator-facing strategy — Belsky had been Chief Product Officer for five years leading all Creative Cloud product development before his CSO role¹⁵ — left for A24 the following spring.⁹ The deeper lesson outlasted the episode: a company that locks customers in had better never give them a reason to imagine the lock is for something worse than convenience. Adobe didn't steal anyone's art. It just reminded an audience that already felt cornered exactly how cornered it was.