The most transparent AI-training rollout in recent memory belongs to the company now being sued over it.
That is the contradiction to sit with. Zoom buried its terms and rewrote them twice in a week. Slack hid enrollment in a privacy page and made the exit an email address. Adobe reportedly pushed a modal so poorly worded it spent two weeks walking it back for language it never acted on.
Figma did none of that. In June 2024 it published a post called "Meet Figma AI" that said plainly: sharing customer content for training is optional, admins set the preference, Starter and Professional plans are opted in by default, and no training happens before August 15, 2024. Seven weeks of notice. A date. A mechanism, described.
More transparency than Zoom or Slack managed. And it is the one that ended up in the Northern District of California.
That is not an accident of bad luck. It is the whole story. Figma told the truth clearly and got sued anyway, because clarity was never the thing under threat. The thing under threat was trust, and trust does not track what a company disclosed. It tracks what the default was when you arrived.
Trust does not track what a company disclosed. It tracks what the default was when you arrived.
Which brings us to a stage on a Monday afternoon in April 2026, and a question Figma board director Andrew Reed lobbed at CEO Dylan Field about his relationship with Anthropic.
It was a loaded question, and everyone in the room knew it. Three days earlier, Anthropic had shipped Claude Design, a product that seemed to compete with Figma. Three days before that, Mike Krieger—co-lead of the same Labs team that built it—had formally resigned from Figma's board.
Reportedly, a board seat was vacated, then a rival launched from the same team. Field was gracious about Krieger personally: "Mike is a great guy, and he was awesome on the board." For Anthropic itself he reached for a colder instrument. "They were not consistently candid in their communications."
A diplomatic knife. And a confession of the trap Figma is in.
Because Anthropic is not just a competitor. The feud dramatizes something Figma cannot buy its way out of: the companies it depends on for intelligence are the same companies now shipping design tools.
That is the first front. It is loud, personal, and ultimately the less dangerous of the two.
The second front is quieter and structural. Claude Design, Google Stitch, and Vercel's v0 reportedly do not merely make screens faster. They invalidate the workflow Figma was built to host.
For years the sequence was fixed: research, sketch, wireframe, polish, prototype, hand off. Figma sat at the center of it as the flexible canvas where a designer controlled every layer, component, and interaction by hand.
Claude Design proposes a different verb. You do not construct the design; you describe it. The model generates a first version and you refine it through conversation. Stitch will reportedly spin a mobile banking app from the prompt "home screen of banking app for mobile"—no colors specified, no name, running on Gemini 3 Flash, the fast model, not even the powerful one.
The manual canvas is not being beaten on features. It is being routed around entirely.
This matters more than it looks, because the manual canvas is where Figma's data lives. Millions of files, every layer named and nested, every component and interaction logged—arguably the richest structured record of how professionals actually design anything, anywhere. That corpus is arguably Figma's real asset. It is also, arguably, exactly what you need to train an AI that competes with Claude Design.
So Config 2026 was the tell. Facing conversational rivals that threaten the whole workflow, Figma's headline answers were Figma Motion and Shaders—craft features. Real ones, valuable to some. Not all UI designers will ever touch a shader.
They were not, as one writeup put it, what the design community expected to see as an AI answer. When your competitors are reinventing the act of designing and your response is a better gradient, you have answered a question nobody asked.
Here is where the two fronts converge into a single trap.
To fight AI-native rivals, Figma arguably needs to train on customer work. The only training data that could give it an edge over a general-purpose lab is the specialized, structured record of real design labor sitting in those millions of files. But that data belongs to the customers who trusted Figma to be the neutral standard tool—the place your work stays yours.
Figma can monetize that asset only by spending the trust that made it the default. Arguably, there is no version where it uses the files and preserves the promise—the two are the same substance.
The Khan lawsuit did not create this trap. It made the bill visible.
Reportedly filed in November 2025 by startup founder Raza Khan as a proposed class action, the complaint leads not with copyright but with consent—the claim that Figma switched on training over customer files by default after years of promising it wouldn't. Figma disputes it, reportedly saying it trains on general patterns like how to build a weather app, not customer content.
Hold the legal question open; that is where it belongs. The strategic question is not.
Because look at the pattern the AI cases have carved. Policy lands quietly. Enrollment is the default. Somebody outside the company finds it. Then the clarification, the apology, and often the walk-back. Zoom walked it back. Adobe walked it back—both recovered inside a fortnight because reversal was available: a sentence to fix, a toggle to flip.
Slack is the exception worth studying. It rewrote the sentence and kept the mechanism—the same move in different clothes, the email address still the exit.
Figma is a further category still. Nothing was walked back. The tiered default was announced, explained, defended with a stated rationale, and allowed to take effect on schedule. You cannot clarify a decision that was already clear.
And that is the point that should worry anyone watching Figma's position, not just its lawyers. Zoom and Adobe made drafting errors and fixed them. Figma made a choice and stood by it. It looked at opt-in versus opt-out, chose opt-out, and chose it differently by tier—Starter and Professional enrolled by default, the paying enterprise reportedly protected.
Trust is not a feeling. It is a set of defaults you can audit. The only question that matters is whether the switch was off when you got there. For Figma's smallest customers, it was on.
That is the paradox as self-defeating strategy. Every ounce of data Figma extracts to compete against Claude Design comes out of the same account: the standard-tool trust that was its one durable advantage.
Because here is the brutal arithmetic. Against a frontier lab, Figma cannot win on models—Anthropic will out-model it, and can reportedly afford to lose money on design tooling in ways a public design company at a $1.2 billion IPO cannot. Figma's edge was never going to be raw intelligence.
Its edge was being the trusted default. The place teams standardized on precisely because it did not treat their work as raw material. That trust was the moat the labs could not cross with a better model.
And it is the exact thing Figma spent to build the AI that was supposed to defend it.
The two-front war, then, is really one war with a single supply line. To fight the conversational rivals, Figma draws down customer trust. Drawing down customer trust makes the standard tool feel less standard, less safe, more like just another vendor calculating whether anyone will notice.
A lab that can out-model you anyway does not need you to make that mistake. But you made it before anyone forced the choice—seven weeks' notice, published openly, opt-out by tier. The most transparent rollout in the industry, and still the wrong bet, because it spent the only asset a better model can't replicate.
Field called Anthropic not consistently candid. His own company was perfectly candid. That was never the problem.