How we Represent Machine Learning Products

From spheres to gears to sparkles to spheres again: designers of AI-based brands and features walk a fine line between relatability, conveying functionality and creating a vision of the future

Generative AI features are popping up everywhere behind a sparkles icon ✨. Magic fairy dust! Joy and excitement! ✨✨Vibes ✨✨! Sparkles are adaptable and can be combined with icons representing the actual function of the tool, whether that’s search or image replacement. They’re friendly and shiny new. They also emerge from a long visual history of representing artificial intelligence and automation seeking to balance conveying functionality, making tools relatable and shaping visions for the future. All while hiding the human labor that fuels them.

Spheres, Bubbles, Eyes, Ethereal beings: Humanizing AI Companions

Going back to the 1960s, visions for artificial intelligence oscillate between autonomous beings anticipating our needs and tools integrated in our workflow. AI companions were all the hype in the 2010s, that saw high hopes for autonomous voice assistants. A cornerstone of the mythology of AI is HAL, the artificial intelligence from the 1968 Kubrick movie 2001: A Space Odyssey remains. HAL is an eye, black with a fixed red pupil. It reflects its environment. It only animates at the very end of the movie, its red light diffracting to encompass the entire scene. HAL is also a voice, calm but devoid of variations in tone conveying its un-humanness. 

The sphere, albeit made less threatening, is a core element of AI assistants design. The sphere is frictionless, an elemental shape – consciousness abstracted and unhindered. The 2013 movie Her by Spike Jonze first introduces its artificial intelligence as a twisting sphere, until it blends in, only evoked by color, light, and her name: Samantha. The animated sphere becomes key to mid-2010s voice assistant logos, whether Siri (Apple), Alexa (Amazon), Cortana (Microsoft), Google assistant, Alice (by Russian company Yandex) or Aligenie (Alibaba). Note also the dominance of far less aggressive blue and purple.

Her continues to influence the field. In early 2024, OpenAI replicated the AI’s voice, originally by Scarlett Johansson, without approval. Virtual assistants not only sport names generally considered feminine, they also originally sounded like women. This gendered  history goes back to the first attempts to create interactive programs. Most famous among them is Eliza, developed at MIT between 1964 and 1967. These assistants do the work usually ascribed to women: planning, emotional support, or the dredging work of household logistics. And so they are largely represented that way, subdued and innocuous.

Interestingly, this differs from popular culture’s representations of general AIs, more likely to focus on fears that they would rob us of our livelihoods or seek to control us. Scifi Interfaces’ study of AIs gender in on screen fiction shows a 5:1 male to female ratio. Apple’s early representation of the endgoal of computing (Knowledge Navigator, 1987) similarly presents a highly knowledgeable and nerdy male assistant. Responding to concerns about bias, some companies building virtual assistants have introduced more diverse voices, and Google, late to the game, opted out of giving its voice assistant a first name entirely.

From Chat Robots to Swirling Hexagones, and Abandoning Cogs: Mechanics to Enable Various Projections

In parallel to virtual assistants that are supposed to be versatile helpers, we find narrower AI-powered chatbots. Their representations inherit from other tropes, that of the cute helpful robot. Their function is also easier to represent by speech bubbles or ellipsis. Those representations suggest helpfulness, but in a rather limited fashion, like that of a butler.

In Fast Company, James I. Bowie further argues the branding of generative AI businesses take its cue from the blockchain’s brands use of hexagons, referring to cubes in perspective. They also borrow a page from the earlier playbook of Her and voice assistants, using twisting lines and infinite movement to suggest interactivity. OpenAI’s logo is the clearest example of this swirling hexagons trend. Bowie also compares them to shutters (mechanical eyes) or vortexes. The use of elemental symbols and implied movement strikes a balance between implying innocuity and infinite potential for growth. 

We’re quite far from the humanized branding of voice assistants, or from sparkles. It’s worth noting that Large Language Models are named after animals (llama, alpaca), technical (GPT), or gender neutral names (Claude). Abstract symbols enable projection, which a mid-2010s trend sought to exploit. Inspired by video games’ avatars and generative design, this approach avoided sidestepping human representations that might see them accused of biases. Replika used eggs users could customize. It’s now switched to sims-like representations that I can only describe as icky.

Circa 2-16, when I worked in that industry, and we struggled to explain what machine learning tools even did. A lot of products lied about using machine learning. Key words were Big Data and Deep Learning. Data processing was the metaphor highlighted, not independent thought or novelty. We used cogs as the icon denoting the use of machine learning. A key representation was that of an artificial brain, emphasizing performance and technicality.

Cogs and networks evoked how this tool worked, rather than what it aimed to achieve.

Sparkles: Sprinkle the Magic Dust

Branding generative AI tools means addressing these same challenges of achieving relatability and conveying functionality. But they always were described as radically different from previous AI products, with industry leaders even warning of their potential to overtake humanity. In many cases, they’re also changing how existing tools work rather than adding new features. The sparkle is recognizable at a very small scale, making it an excellent building block for composite icons.   

While cogs can similarly be combined in other icons, they don’t suggest an elevation in the work done by the tool. If anything they suggest a return to earlier forms of automation, or how power is leveraged in a given mechanism. There are obvious benefits to associate these tools with disembodied magic rather than the labor of human cogs creating and annotating training data and the huge environmental costs of servers used to run them. After all, genAI tools need to be playful for casual uses and be palatable to creative workers (whose data was stolen!) in professional tools. Its universal use also suggests a similarity and continuity between these tools, as if they were all the work of the same magical entity.

But I don’t think using sparkle is just a way to convey positive feelings about AI tools. It has another genealogy in Photoshop’s magic wand, allowing the automatic selection of an object, in 1989. In the early 2000s, human-computer interaction researchers explored how it could be used as a pointing device in early VR systems. In interaction design, prior to generative models, sparkles have variously been used to convey excitement and the automation of tasks, from selection, to processing, to image improvement.

Within that history, using sparkles to convey tools using some kind of generative capacity goes beyond conveying positive feelings. Anyone who’s used the magic wand in Photoshop knows its results are uncertain, even with the ability to fine tune parameters.

Returning to Spheres: Punctuation and Movement as Metaphors

Spheres haven’t said their last word. ChatGPT now uses a plain, bold dot signifying writing in progress, always pushed to the end of the sentence being written. It disappears at the end of the task. Claude is represented by an organic blob, somewhere between a spark, an asterisk, an ink drop and a loading wheel. In more specialized industry like self-driving cars, logos inherit from previous metaphors, suggesting movements or places. There’s a whole other essay to be written about the incorporation of Google maps’ blue dot in Waymo’s logo.

We might soon be seeing again different metaphors depending on what a particular tool is aiming to achieve. Claude and ChatGPT are returning to representations associated with writing: punctuation (see also: this post on branding Perplexity AI). Claude goes further, in adopting Copernicus, a humanistic (serif!) typeface. It hints at Claude having the potential to spark a Renaissance period, while being more concrete and approachable. Remain absent those who voluntarily and involuntarily lend their labor to power those tools.

---

This piece is a work in progress.

---

---