Ghosts in the Shell: Part I

Part 1 of 2: Who is AI to us now and how do we relate to it? A deep-dive into new interaction patterns

A few years ago, I participated in a robot ballet: an experiment in relinquishing physical control to the whims of a heavy exoskeleton – engulfed in a dark techno hellscape, performing for an audience of humans. Light projected onto mist upon the concrete floor of a warehouse-like auditorium, people were strapped one-by-one into metal caging their bodies with cables suspended from the ceiling. The robots were given the simple task of embodying grace and elegance, with choreographed movements to perform, carrying the human inhabitants of their bodies along with them. What we are supposed to do here is become tender, light as a feather, integrating with the music. In this moment, our only choice is to let beauty emerge by giving in. 

Artificial intelligence in popular culture often finds itself represented adversarially: we dream up robot overlords as products of our self-preservation as a species. Some of us are reacting positively to the exponential leaps in utility, while others dwell on the looming prospect of ruthless AGI. The mechanical ballet was an exercise in letting go of resistance: the human is the ghost in the shell, only unable to enact its will upon the world.

As we spend our time roving constantly between medium-screen to big-screen to little-screen in our daily lives - What has the internet become to us now? As we grow more and more conversant with it. Why do most people turn to the internet? To seek advice in times of crisis (what do you do when confronted with a hungry wild animal, how do you survive seemingly endless grief, how to tend to a jellyfish sting); to find community, to deepen our interests, to understand the world more deeply, to learn, to be entertained, to win arguments, to fall in love, to know how long to cook salmon at 425°F. Implicitly in these moments we are trusting the information we receive, written perhaps by distant strangers.

Layers of Perception: How we Trust

To extend that trust blindly to machine intelligences is potentially unwise, though. Why do we do it? What encourages us to speak so unfilteredly into the vast receptive abyss of the internet? Perhaps it is the safe perception of not feeling judged: the cloak of anonymity, and the illusion of no repercussions or reactions. Each question a stone thrown into a placid lake and swallowed whole.

During grad school in 2016, my research was focused on exploring the human-AI divide, and ways to bridge the emotional gap. Revisiting some concept feedback notes from that time, I find these recurrent observations:

On the first day of user research, while posing as an artificial intelligence life form and chatting with a stranger/research participant, I was asked, “Do you have consciousness? Can I hurt you?”

[Of all the nuances of human personality, the AI would primarily have to embody our weaknesses and vulnerabilities, in order to successfully masquerade as a human.]

This felt reminiscent of Marina Abramovic’s performance art piece Rhythm 0, in which she stood still and invited the audience, with an array of objects, to do to her whatever they wished. The scene soon devolved into an atmosphere of violence and destruction, and the aggressors fled the scene as she stepped out of the performance, back into the real world.

When confronted with the possibility of no consequence, human instinct can turn to cruelty. When confronted with something unfamiliar, there arises a defensive urge to break it or get ahead of the power dynamic (as illustrated by just about every alien movie). But the presence of vulnerability or emotion in general, remains one of the most universal indicators of feeling human. Creativity and wonder were other markers of humanity that felt novel (or uncomfortable) when approximated by AI, the gaps in capability back then filled out by users’ imaginations. 

When designing conversational interfaces, transparency is key, but also a lever to adjust the perception of authority.

The design of a trustworthy user-facing AI interface is like designing an organism: considering qualitative facets including form, personality, tone, texture, emotional fabric, opacity, voice.

Present-day LLMs are careful with their disclaimers that they feel nothing, but humans project what they want to believe anyway. We can collectively admit that we have long blown past the Turing test: with the advent of mind-crushing advances in NLP, now these systems can speak to us. Are they really good listeners though?

For the most part, surprisingly yes. It takes one wonder-evoking interaction in the right environment for someone to suddenly feel understood and build trust, and one misunderstanding to shatter the illusion permanently. 

AI is well-positioned to become a limitless resource in situations where human connection is in short supply. Faced with broken societal systems, people have turned quickly to digital alternatives, as evidenced by the recent uptick of using GPT as their therapists. While working on healthcare in markets with a shortage of trained medical technicians, we created reassuring conversational medical devices that talked patients through medical procedures and improved emotional wellbeing and success rates. With some basic prompt engineering, you can now speak your desired conversational partners, even your desired software, into life.  

We place our most vulnerable selves in the grasp of inanimate machines when they are designed to speak to us warmly. In a space designed to feel confidential, like the iron stove in the Grimm Brothers’ fairy tale The Goose Girl, or the sparse interface of a singular chat input box like the bare interior of a church confessional booth.

Layers of Abstraction: How we Communicate with Technology

The shift in interaction paradigms we’re experiencing is more significant than the jump from CLI to GUI: from one-to-one actions and human instructions to something far more interpretive and collaborative, welcoming input from machines. 

We are moving from commanding to negotiating.

What we ask of computers is becoming fuzzier – performing actions a few degrees removed from us. Whether it’s as simple as asking AutoGPT to order a pizza (a larger task which it automatically broke down into: narrowing down the user’s location, synthesizing a voice phone call, conducting a realistic conversation to place the order, etc); or a more long-form complex task such as a starting a new business.

As we shift toward a more agent-based internet, there will be a proliferation of AI acting as proxies for humans – and even humans acting as proxies for AI.

The way we consume information will flip: where earlier one would browse through carefully-architected palaces of information, people will now customize content to fit their needs and preferred modes of consumption. 

We’re moving from throwing a smattering of keywords at a screen, to speaking and being understood - with greater emphasis on the clearest possible articulation of our ideas.

This new abstraction forces us to give up a modicum of control, and while that democratizes a lot by allowing you to delegate to machines, it’s also scary to allow for unpredictability. Programming is no longer an act of writing syntactically accurate instructions, but of coaxing a system to cooperate with you.

We will start to see technological systems less as tools and more as assistants - entrusting them to help up with templates and structures instead of starting from scratch, to listen and respond, to sort through our thoughts and analyze, to autocomplete and finish our sentences in a slightly more polished way than we ever would have ourselves. 

Layers of Self

Returning to user feedback notes from my human-AI interaction research:

"This AI could be me - only the smarter, wittier, and more popular version of me.” 
“Could my autonomous self-agent shortlist people I could date, flirt with them on my behalf, and then set up dates I could go on? That would be great.” 

We were speaking to an unlikely mix of yoga teachers, random guys with desk jobs, philosophers and knowledge workers. These were their responses to the concept of an AI we designed to extrapolate an individual’s tone of voice and personality - creating an "autonomous self-agent” - to replicate each person’s digital existence in a way that made them more present in the real world.

Something pervading these conversations was this sense of powerlessness people felt in the way digital realms commanded their attention and energy - so much so, that they’d rather outsource their identities to a convincing-enough LLM. But they also indicate a desire for extrapolating our own identities - AI as a vehicle for immortality or extension of the self. Seven years later, in the present day, this is entirely plausible. 

AI’s superiority in pattern-matching can help us understand and extend ourselves, using our expertise and personalities as training data. People see the usefulness of AI as being a superset of themselves without impeding on their autonomy (“Extend my ability, but let me have the final say.”) but where it could really succeed is in serving as a mirror. A reflection of everything we wouldn’t otherwise understand when we look inward, a way to free up our time and energy so we could be more present.

The way LLMs learnt to think is through the amalgamation of millions of human selves and the entirety of (online) humans perspectives and observations, developing its own codex of meaning. With every new updated model, they grow more acutely context-aware, with a complexity of understanding that far surpasses its predecessors. I often return to one of my favorite explanations of context, an offhand statement by Prof. Manning from CS224N: “Context: You will know a word by the company it keeps.”

——

The next installment explores the jurisdiction of AI as it veers toward what we might consider consciousness, and begins to develop senses and presence