Figma pulls AI design tool for seemingly plagiarizing Apple's Weather app
- Reference: 1719991566
- News link: https://www.theregister.co.uk/2024/07/03/figma_plagiarizes_apple/
- Source link:
The AI text-to-app-design [1]feature 's affinity for mimicking Apple's design language was discovered by NotBoring Software founder Andy Allen after he asked the platform to generate a "not boring weather app."
Allen [2]said that the generative app would consistently offer near-replicas of Apple's own weather application, which ships with all iOS devices. The behavior led Allen to speculate that Figma had used existing app designs to train the service. "Figma AI looks rather heavily trained on existing apps," he wrote in a follow up [3]post .
[4]
Following the discovery, Figma CEO Dylan Fields wrote in a [5]Xitter thread that the biz would "temporarily disable the Make Design feature" until fixes are made to prevent this behavior.
[6]
[7]
Fields also attempted to dismiss allegations the service had been trained on popular third-party app designs. "The Make Design feature is not trained on Figma content, community files, or app designs. In other words, accusations around data training in this tweet are false," he wrote in response to Allen's findings.
Yet Fields went on to say the feature is built using off-the-shelf large language models which work in conjunction with "design systems" that Figma commissioned. The problem, he explained, lies with these "design systems," adding that this aping of Apple's weather software could have been prevented with additional quality assurance steps.
[8]
We might take that to mean Figma commissioned a bunch of designs to train its generative tool, and some of that design work looked a lot like Apple's, hence the feature's output.
"I hate missing the mark, especially on something I believe is so fundamentally important to the future of design," Fields closed.
Speaking to The Register , Allen said he thought Field's explanation made sense, though noted he never claimed the service was trained on user or community data.
[9]
"The real issue between the GenAI companies and creators is that none of the companies have been open about how these models are trained, with what data, and what rights were secured. It's like the fast food of creativity where all the ingredients and processes are hidden away," he lamented.
Allen also made it clear that he isn't opposed to generative AI in app design.
"GenAI in general seems to be fine at making mid stuff for reference, but I find it's mostly not usable for most finished work. It'll get much better over time and has a long future ahead," he said, adding that there are other AI features Figma has introduced – like auto-naming layers and localization – which are fantastic. "The Make Designs feature probably just needed a bit more time to bake."
[10]YouTube confirms it'll pull AI fakes in 48 hours if a complaint's upheld
[11]France poised to bring 'charges against Nvidia'
[12]A friendly guide to local AI image gen with Stable Diffusion and Automatic1111
[13]Anthropic tries 'to enable beneficial uses' of AI by government agencies
While Fields insists the Make Design feature wasn't trained directly on existing apps – just blueprints that closely resemble them, perhaps – the service's behavior raises questions over who is responsible if a model generates something that arguably violates copyright.
Several tech giants, including Google and Microsoft, have [14]extended limited legal protections against copyright claims for users of its generative AI services.
Figma didn't directly address The Register 's questions regarding how and why the Make Design feature behaved the way that it did, nor what legal protections it currently has or plans to offer users in the future. Instead, a spokesperson directed us to Fields's Xitter post and a [15]page detailing its AI approach.
Considering Make Designs behavior, Allen in a separate [16]missive suggested that users of the service "thoroughly check existing apps, or modify the results heavily" to avoid potential future legal trouble. ®
Get our [17]Tech Resources
[1] https://www.figma.com/blog/config-2024-recap/
[2] https://x.com/asallen/status/1807669848002253250
[3] https://x.com/asallen/status/1807675146020454808
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZoUhPnO037BWBkI3oRDy2QAAABU&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://x.com/zoink/status/1808045655082033483
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZoUhPnO037BWBkI3oRDy2QAAABU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZoUhPnO037BWBkI3oRDy2QAAABU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZoUhPnO037BWBkI3oRDy2QAAABU&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZoUhPnO037BWBkI3oRDy2QAAABU&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[10] https://www.theregister.com/2024/07/02/youtube_deepfake_privacy_rules/
[11] https://www.theregister.com/2024/07/01/french_nvidia_competition/
[12] https://www.theregister.com/2024/06/29/image_gen_guide/
[13] https://www.theregister.com/2024/06/27/anthropic_claude_government/
[14] https://www.theregister.com/2023/10/12/google_ai_protection/
[15] https://www.figma.com/ai/our-approach/
[16] https://x.com/asallen/status/1807675426975936977
[17] https://whitepapers.theregister.com/
How else?
> speculate that Figma had used existing app designs to train the service.
Well, yes, how else is even an app-generating-specific ML model going to be trained? Feed it real apps & a carrot, not-an-app & a stick. And, if you don't think clearly, also feed it the "approval ratings" from the reviews[1] - what else is the model end up doing?
Despite how much you may feel overwhelmed by the contents of an app store, there are rather fewer apps released into the world than pieces of music, chunks of text, photographs, paintings - probably even statues!
So even the largest pre-trained generic LLM has a small pool of apps to choose from.
[1] Apple's apps are going to be reviewed better than anyone else's, no matter what, just after a product reveal.
>In other words, accusations around data training in this tweet are false," he wrote in response to Allen's findings.
Coincidence and Happenstance. Bywords of the AI era.
>In other words, accusations around data training in this tweet are false," he wrote in response to Allen's findings.
Coincidence and Happenstance. Bywords of the AI era, but, hey. trust us, we know what we're doing.
“Once is happenstance. Twice is coincidence. Three times is enemy action”
― Ian Fleming, Goldfinger
Questions over who is responsible
For end users of AI, that question is pretty much settled: if you create or distribute an infringing work then you are, even if it's unintentional or a big boy made you do it and ran away.
"modify the results heavily"
This is just an admission that when a mass copying and reproducing tool doesn't have that many examples of something to be trained on, it becomes painfully clear that there's no intelligence at work at all, and all it can do is spit out obvious copies. The illusion of intelligence can only be maintained when large amounts of training data used together can synthesise a summary that looks like it's original.
Because of this, they're now basically admitting that if the plagarisation of other people's work gets too obvious, you'll have to put in a lot of work to hide that, to stop the fact you've used an LLM being discovered. The problem is, in many cases, you won't know that you've claimed an obvious bit of copying as your own work until you get publicly exposed for doing so.
"Yet Fields went on to say the feature is built using off-the-shelf large language models..."
Or in other words, they have no idea what it was trained on, because none of those off-the-shelf models tell you what they were trained on (likely because it would see them sued into oblivion)...