Your AI can see you now. It still doesn’t know your eyes are blue

You’re at dinner. Across the table, someone’s wearing unusual glasses - sleeker than normal frames, with barely perceptible sensors embedded in the temples. They seem engaged in conversation, but something feels slightly off. They hold your gaze a fraction too long. They pause mid-sentence, eyes unfocused, as if listening to something you can’t hear. When they look at you, their attention lingers in a way that makes you uncertain whether they’re seeing you or seeing through you.

Their glasses contain an AI assistant. While you’re talking, the system processes you continuously: facial landmarks mapped, clothing catalogued, body language tracked. They could ask it anything. “What’s their mood?” “How old are they?” “What brand are they wearing?” The AI would answer instantly, drawing from visual data it’s already captured and processed.

Something is happening in the space between you and those sensors. Something that sits uneasily with our existing categories of being seen, being photographed, being observed. What’s emerging occupies a territory we’re only beginning to understand, let alone design for - a domain where processing happens before meaning-making.

What vision actually is

In March 2000, Mike May experienced something that had happened to fewer than twenty people in recorded history. At 46, after 43 years of blindness, bandages were removed from his right eye following experimental stem cell surgery. Light flooded in. Within seconds, he could name colours perfectly: that’s blue, that’s red, that’s yellow.

Then he looked at his wife and couldn’t recognise her face.

May had been blind since age three, when a jar of chemicals exploded. He’d lived a remarkable life in darkness: champion downhill skier racing at 65mph, CIA analyst, tech entrepreneur. The surgery gave him back his eyes yet vision remained elusive.

Researchers studied May intensively for fifteen years. The pattern was consistent: he could see motion perfectly but couldn’t identify a still image of a cube. He could navigate crowds but couldn’t tell shadows from solid objects. He recognised colours flawlessly but couldn’t recognise faces. His eye worked perfectly - ophthalmologists confirmed 20/40 vision. The problem lay in processing. His visual cortex had repurposed itself during those 43 years of blindness. When signals returned, the machinery for making sense of them had been dismantled.

May’s case reveals what we tend to forget: vision requires active construction. The brain builds elaborate processing systems through years of visual experience. Miss that developmental window and you can have perfect eyes yet remain effectively blind. Understanding what “seeing” actually means prepares us to understand what happens when machines encounter visual data.

The cultural construction of sight

Languages develop colour terms in a recognisable sequence. Research across hundreds of languages shows the same pattern: first come terms for dark and light. Then red. Then either green or yellow. Then blue. Always in that order.

Blue arrives remarkably late. Homer’s epics describe a “wine-dark sea” and bronze sky. The word kyanos appears but meant dark rather than blue. Ancient Greek had no word for blue - not because Greeks couldn’t perceive it, but because they had no practical reason to name it separately.

The pattern holds globally. Japanese didn’t distinguish blue from green linguistically until the Heian period (794-1185 CE). Even today, Japanese speakers say aoshingou - blue traffic light - for what’s clearly green. Russians have separate words for light blue (goluboy) and dark blue (siniy) and perceive them as different colours the way English speakers distinguish red from pink.

Why does red arrive first? Blood, ochre, fire - red signals danger and life. More practically, red dyes were easy to make. Madder root, cochineal insects, iron oxide ochre - all readily available. Blue was different.

Blue barely exists in nature. Creating indigo dye required sophisticated chemistry: harvest woad at the right time, ferment the leaves, extract precursors, oxidise them, reduce the compound back to soluble form. Because of this difficulty, indigo was called “blue gold”. When better sources were discovered in India during the 16th century, European governments tried to ban imports to protect local woad industries.

You name what matters. Red ochre appears in cave paintings from 40,000 years ago. Blue appears virtually nowhere in ancient art until Egyptians developed synthetic pigments around 2500 BCE. The sky has always been blue, but for most of human history, blue was unnecessary to name. Perception stayed constant. Language followed need and capability.

This pattern - where culture, technology and communication needs shape how we carve up perceptual space - matters when we think about what AI vision creates.

A new perceptual domain

In 2008, Radiohead released a music video unlike anything made before. “House of Cards” contained no cameras, no lenses, no traditional filming. Instead, director Aaron Koblin used rotating LIDAR scanners and structured light sensors. Lasers swept across lead singer Thom Yorke’s face, measuring distances, mapping surfaces.

The result is haunting. Yorke’s face appears as a point cloud - thousands of white dots floating in black space. As he moves, the dots shift, maintaining topology whilst revealing nothing about colour, texture or expression. It captures everything about spatial relationships and nothing about experience. Koblin saw something others hadn’t: there’s aesthetic and informational richness in the space between sensor and subject.

The LIDAR doesn’t capture what a camera sees or what Yorke experiences. It captures geometry - the shape of the space between. AI vision systems create versions of this intermediate territory. They generate what we might call a look space - a domain where processing happens before meaning-making, where different viewers extract different patterns from the same visual field.

Consider its characteristics:

Speed: Systems like YOLO (You Only Look Once) process images in real time, identifying objects, people and actions within milliseconds. Point it at a traffic scene and it distinguishes vehicles, pedestrians, traffic lights, road markings faster than human attention can shift focus. It can identify a specific face in a crowd of thousands while a human is still scanning the first row.

Scale: Satellite imaging systems process visual data across decades and continents. Climate scientists track rainforest changes, ice coverage shifts, ocean temperature variations at scales impossible for human monitoring. The look space persists and accumulates in ways human memory cannot.

Persistence: The field doesn’t fade. Human vision is constructed moment by moment, biased by attention, shaped by expectation. This domain simply exists - patterns detected, relationships mapped, data logged. It doesn’t forget, doesn’t get distracted, doesn’t privilege foreground over background.

Geometry: Like the LIDAR in that Radiohead video, modern vision systems capture spatial relationships independently of meaning. Medical imaging AI detects microscopic patterns in tissue that look identical to human observers but indicate process changes. Manufacturing systems spot product variations invisible to quality inspectors but correlated with equipment drift.

We’ve created something genuinely new: a domain that exists between sensing and understanding, queryable in ways previously impossible. Two people looking at the same scene see different things based on their experience, attention and culture. An AI system looking at that same scene sees patterns matching trained parameters. Each extracts different information from the same visual field. The look space is where that extraction happens.

Designing in the new space

The “glasshole” phenomenon - the social backlash against Google Glass wearers in 2013 - was early evidence of discomfort with this domain before we had language for what bothered us. People objected to being recorded, but that wasn’t quite it. Plenty of contexts involve cameras without triggering the same response. Something about the combination - always-on, processable, queryable, worn casually - created unease.

The technology has evolved significantly since then. Processing power has leapt forward in bounds. Dedicated processors now handle vision, security and privacy as separate, optimised streams. It’s an arms race: each capability increase triggers new privacy concerns, which drive new protective features, which enable new applications. AR glasses from various manufacturers now look ordinary while creating look spaces far more sophisticated than Google Glass managed.

Apple, characteristically late to this territory, has prioritised development of their AR offerings. Given their history of arriving after competitors but revealing deeper thinking about interaction design and privacy frameworks, their approach will be worth watching. How do you design for a look space that respects both capability and constraint?

You’re at that dinner table again. Someone asks their glasses, “What’s everyone wearing?” The system has already processed clothing brands, styles, price indicators. Or: “Who seems uncomfortable?” The AI analyses micro-expressions, posture, gaze patterns. The technology works. The question is what we build in this space.

Medical imaging offers a different view of the same look space. Radiologists examining mammograms bring years of training and contextual judgement. They also bring fatigue, cognitive biases, attention limits. An AI system examining the same images processes pixel values, identifies statistical anomalies, flags patterns correlated with malignancy. It matches patterns to trained parameters with tireless consistency whilst experiencing nothing of cancer’s meaning or the patient’s vulnerability.

They work differently because they’re doing different things. The AI flags patterns in the look space that human attention might miss. The human applies contextual understanding the system cannot access. Together, outcomes improve. Understanding what this look space enables means building systems that use it intelligently rather than trying to replicate human vision in machines.

Climate monitoring, manufacturing quality control, accessibility tools, autonomous vehicles - we’re building infrastructure in this domain. Each application makes choices about what patterns matter, how the look space gets queried, what happens with the processed information. The AR glasses scenario and the medical imaging scenario use similar technical capabilities - pattern detection, classification, analysis. Yet the applications create entirely different experiences and raise different questions. What we design in this look space determines value and risk.

Principles for a new domain

Cultures developed colour vocabularies based on what mattered and what they could make. Red came first because red dyes were accessible and red signalled important things - blood, fire, ripeness. Blue arrived late because blue was difficult to create and less immediately useful. Need shaped language, which shaped how we carved up perceptual space.

We’re at a similar moment with this look space. We’re developing frameworks, vocabularies, principles for working in a domain that didn’t previously exist. This requires more than rules about what systems can or cannot do. We need paradigms - ways of thinking about what this space is and how to work within it.

Some principles are emerging:

Specificity over capability: Rather than asking what the system can see, ask what specific patterns matter for this application. Medical imaging systems and social media systems might use similar pattern detection, but the applications diverge entirely. Clarity about purpose matters more than limits on capability.

Transparency about the domain: When systems process visual information, acknowledging what’s happening in that intermediate space becomes crucial. Rather than saying “the AI sees you” (anthropomorphic, misleading), say “this system processes facial landmarks for X purpose” (specific, accurate).

Designing for difference: The look space has properties unlike human perception. We can design for what this field enables - persistence, scale, speed, geometric precision - rather than trying to replicate how humans see. Artists like Aaron Koblin exploring LIDAR’s aesthetic possibilities and companies like Apple investigating how we interact with augmented look spaces are working with these properties rather than against them. The opportunities come from understanding what becomes possible when you can query visual space at this speed and scale, when patterns persist longer than human memory, when geometry can be captured independently of meaning.

Understanding these differences opens possibilities. Mike May could name blue instantly but couldn’t recognise his wife’s face. AI systems process pixel values perfectly but experience nothing. Both reveal that “seeing” means different things in different contexts.

What we’re building

The systems we’re creating establish this look space as infrastructure. It sits between raw sensing and human interpretation - fast, queryable, persistent, geometric. It enables things impossible before: identifying faces in crowds, tracking climate changes across decades, spotting microscopic tissue anomalies, helping blind users understand their environment.

New applications emerge as people understand what this domain enables. OpenAI’s Sora is creating generative video from text prompts, giving millions of people tools to create visual content they couldn’t make before. Some dismiss the results as “AI slop”. Others, like John Gruber at Daring Fireball, find themselves fascinated by how it’s enabling creation of new and unexpected visual ideas. Both responses make sense. We’re learning what works in this space and what doesn’t, what delights and what disappoints. The look space makes both possible.

What happens in this domain depends entirely on what we choose to build there. Earlier cultures developed colour vocabularies based on what mattered - what they could make, what they needed to communicate, what served their purposes. We’re doing something similar now. The look space exists. What we imagine and build there will shape how this domain develops.

Some questions worth asking:

What patterns matter? Rather than building systems that try to see everything, what specific patterns serve the purpose at hand? Medical imaging needs different pattern detection than accessibility tools, which need different approaches than climate monitoring.

What does persistence enable? The look space doesn’t fade like human memory. Satellite data accumulates over decades. Quality control systems track subtle drifts invisible to human observers. Manufacturing patterns emerge over time spans longer than human attention. What becomes possible when visual patterns persist and accumulate?

How do we design for processing without perceiving? Systems in the look space experience nothing whilst processing everything. This fundamental difference opens opportunities - patterns humans miss become visible, scales humans can’t monitor become trackable, speeds humans can’t achieve become routine. How do we build systems that leverage these differences intelligently?

What emerges at speed and scale? When you can query visual space at millisecond speeds across continental scales with geometric precision, what becomes possible? Climate patterns, disease progressions, manufacturing processes, urban development, accessibility tools - each application discovers new possibilities in the look space.

The look space is here. Processing power continues to evolve. Dedicated hardware refines what can be extracted from visual fields. Privacy frameworks develop in response to new capabilities. Applications multiply as people discover what this domain enables.

What we design there - the principles we establish, the applications we build, the questions we ask about what matters and what doesn’t - we’re just beginning to understand. Your AI can see you now. Something is happening in the space between you and that sensor. A domain with its own properties, its own possibilities, its own requirements for thoughtful design.

What we imagine and build in that look space will shape how this territory develops. The questions we ask now determine what becomes possible later. We’re naming what matters, carving up new perceptual territory, deciding what to build in a domain that didn’t exist before.

That’s where the real work begins.