Humans, not muses.

[Georgia Jenkins]

Home / AI / Artificial Intelligence / consent / cultural production / data infrastructures / digital platforms / digital replicas / Georgia Jenkins / image and publicity rights / Instagram / Muse Image / personality right / Humans, not muses.

Humans, not muses.

 Georgia Jenkins Tuesday, July 14, 2026 - AI, Artificial Intelligence, consent, cultural production, data infrastructures, digital platforms, digital replicas, Georgia Jenkins, image and publicity rights, Instagram, Muse Image, personality right

And suddenly everything became a lot more real; last week Meta launched Muse Image alongside Instagram’s ‘reuse + AI’ setting and then shelved the latter feature in a matter of days. Everyday people could have become the subject of AI manipulation, and some supposedly did during this period: once a user tagged a public profile, Muse Image allowed its reuse to generate a new AI image.

Naturally Meta assured its users that its Instagram settings would have allowed them to control the reuse of their content, including when it is used within a generative AI feature. As this was the first image generation model from Meta Superintelligence Labs, this Kat can’t wait (add the necessary level of sarcasm here) to hear about other products down the pipeline. Apparently Muse Video was already in development, but given the vitriol response to Muse Image’s feature, perhaps this is also paused (for now).

At the beginning of last week, as one tech justice advocate described it, ‘It’s hard to see why Mark Zuckerberg thinks facilitating yet more of this creepy image manipulation is a good idea'.

Muse image was much more than meets the eye

Kat-generated data Photo by Tai Bui on Unsplash

Muse Image is described as an agentic model so it uses ‘search and coding tools to improve accuracy, self-refine its own generations, and improve through scaling test-time compute’. While retrieval-augmented generation (RAG) is already used to connect LLMs to live social media feeds among other ‘current information’ sources, this may be the first time a RAG-based feature is used within a prominent social media platform for the purpose of creating user-generated content.

While previously Meta used third-party AI models like Midjourney and Black Forest Labs for its image and video generation features, it seems that during this time, Meta was building its own agentic ecosystem. Back in April, Meta announced Muse Spark, a multimodal AI model that can ‘engage in multistep reasoning and handle complex processes, manage digital workflows, and deploy new features in an enterprise system’. Muse Image is the first Meta product that uses Spark’s features alongside their social graph to generate content.

Under the direction of Alexandr Wang, Meta’s Chief AI Officer, Meta’s open-source Llama models were abandoned in favour of loftier goals. While some speculate that Muse Spark was designed to compete with Anthropic and OpenAI in the AI coding market, others reflect (here and here) that Muse Image will compete for small to medium business corporate social media advertising. Meta had explained that Muse Image ‘will help power imagine generation in Advantage+ creative [subscriptions] – bringing smarter reasoning and iterative refinement’.

In terms of the effectiveness of Muse Image, Meta’s internal benchmarks placed it ‘behind OpenAI’s GPT Image 2 on overall quality, but ahead of Google’s Nano Banana 2 on single and multi-editing image tasks’. For a more accessible assessment, I’d suggest trying to guess which row of ducks belongs to Muse Image, GPT Image 2 or Nano Banana 2 here. While not leading in terms of image-generation quality, the value of Muse Image likely comes from the very ecosystem they are building. Some aptly reflect that, ‘[n]o other image model can do this, because no other company is holding fifteen years of your photographs, your friendships, and your aesthetic history’.

 

Consent a relic of social media platforms past

Despite Muse Image being ear-marked as the latest frontier for social media platform engagement, the broader picture is more than a little concerning. Although users have access to ‘over 30 new AI-powered effects’ for their Instagram Stories or direct messages on WhatsApp, related user-generated data (UGD) likely extracts network effects at scale (e.g. refining recommendation systems). However, RAG amplifies these effects as alongside data scraped from the internet, it accesses UGD to provide up-to-date answers in real time.

UGD on Instagram is more than accessing a user profile photo. It comprises public content (posts, reels, stories, and highlights), engagement (likes, views, saves and comments), relational data that forms a social graph, and time-related data (timestamps, posting frequency and time-of-day patterns). For Muse Image, this data matrix is likely used during RAG to evaluate whether generated outputs reflect the social and aesthetic norms identified within UGD. While at the moment, there seems to be a focus on RAG within the news publishing context, this Kat suspects its incorporation within social media platforms through products like Muse Image deserve more reflection.

In particular, the rollout of Muse Image was controversially automatic. Instagram users largely became aware of the AI reuse feature through other social media accounts and news reporting informing them to opt out. There was no opt in. In addition to authorial consent and harm related to digital replicas (personality-based or commercial licensing of likeness), the feature had potentially long-reaching implications for creativity and cultural production online, particularly when the integral role of social media platforms for communication is considered. A default platform setting to allow reuse of UGD for generative AI tools, meant that public participation online becomes a form of consent. This is a story we know all too well from freemium platform models based on UGD.

 

Comment

It is being reported that the reason that the Muse Image AI reuse feature was shelved may have been due to SAG-AFTRA and the Creative Artists Agency demanding a reversal (here and here). If it were not for industries associated with protecting and monetising likeness, this Kat wonders whether governance mechanisms would have acted as swiftly, especially when contrasted with the Ofcom investigation of Grok which took roughly two weeks to ban the nudity-related deepfake AI feature in the UK. In comparison, the Muse Image AI reuse feature was removed after three days with SAG-AFTRA publicly commenting that the reversal was ‘the responsible thing to do’. This may give more weight to the necessity of collective action to counter the increasing power wielded by actors within AI data infrastructures. Indeed, it is only through standing together that we can address the aggregate value of societal harm left unchecked.

Even if Muse Image is meant to act as a user’s ‘creative partner’ that uses prompts to ‘spark’ ideas, it effectively neutralises an inherently human act, creativity. The added layer of RAG based on UGD complicates this picture as it creates an ecosystem that continuously loops upon itself, entrenching existing inequalities within creative and cultural production online.

A step further, the choice of ‘Muse Image’ as a title is particularly telling. Feminist scholarship has long exposed the imagery that reduces a person to a voiceless object lacking identity and agency. And so, while the AI reuse feature may be shelved, we should consider the effect that these new agentic AI social media ecosystems have on our humanness. We should never be reduced to merely muses, but seen as meaning makers whose communication forms an integral part of creativity. The value of which this Kat explores in her forthcoming book Creative Reuse.

Do you want to reuse the IPKat content? Please refer to our 'Policies' section. If you have any queries or requests for permission, please get in touch with the IPKat team.