The Gaze Interrupted: Dispatch 004

AI and Me

Or, The Probability-Weighted Linguistic Dice

At my last meeting, my mentor, Heather Phillipson, suggested that I clarify my relationship with AI. She challenged me on my position that ‘it’s just a tool’. For many people, a neutral stance isn’t possible with something so apparently revolutionary. I deeply mulled this. What could she mean? Where is the ambiguity?

After the mulling had been done, I realised that for me, this attitude is natural, given my career trajectory, 25+ years in what is now broadly known as UX (although also HCI, and UXR), but it is certainly worth unpacking for clarity in my practice and anyone reading.

It seems everyone is being asked to pick a side on ‘AI’, and I think, like many things, the truth requires more nuance. ‘AI’ means many things. I’m typing away, and occasionally I spell something wrong, or I have a useful suggestion for the next word that I’m about to type. AI fuels Spellcheck. The probability that I meant to type ‘occasionally’ and not ‘occassionally’ (one of my favourite mistakes along with missing the last full stop according to Grammarly - another AI!) is calculated, and I’m offered the most likely correction. AI is embedded, most famously, in spell checks, auto-completes, and recommendation engines such as Netflix and YouTube. No one is asked to pick a side on spell checks (anymore). Pretty soon, maybe even now, LLMs will be similarly accepted and unremarked.

So let’s be precise and talk about LLMs and my work rather than hypotheticals and semantics. What people usually mean by ‘AI’ today is LLMs, but for me, using LLMs actually to generate ideas is the wrong tool for the job—like using an egg whisk to dice carrots. LLMs are obvious and impersonal by design. Ask your day-to-day LLM for a creative idea, and you will often get something linguistically plausible but conceptually sterile. It tries (it’s helpful), bless it, but it’s trained on the whole internet. It’s impersonal.

An example

I asked an LLM for five one-sentence ideas about thyroid eye disease.

Here’s one: “Double Vision Corridor (Installation): A narrow, dimly lit hallway lined with subtly misaligned, intersecting mirrors that force viewers to experience a disorienting, split perspective, capturing the daily struggle and confusion of diplopia (double vision).”

It’s obvious, literal, understandable, yes, but it doesn’t reflect my experiences or those of the patients and clinicians I’ve spoken to. I could keep smashing the refresh key until an idea I liked popped out (artist as editor), but for me, a more guided approach is required. Making a hall of mirrors wouldn’t reflect my personal experience with TED or the variety of experiences of my patients and clinicians. I want to find out and communicate more about the human condition through my own.

I use LLMs as a highly specific toolkit to support my practice:

Writing and admin: I use an LLM to help me write clearly, research, and handle administrative tasks.

Coding agents - LLMs translating natural language to programming languages: I use agents (specialised LLMs) to help me code. They translate natural language into Python to build the back-end and interfaces for my creative systems.

Local AI for privacy: I run three lightweight AI models on my local drive—WhisperX for transcription, Pyannote for diarisation (allocating utterances to speakers), and Presidio for anonymisation. This maintains my privacy and security standards.

In my Haptic Translator app, I use an LLM (Gemini) to classify my utterance clusters and to suggest a material. Why ask this particular tool to suggest materials? Because it’s essentially a linguistic engine. It links concepts by proximity, like a huge multi-dimensional mind map.

In this project, I use LLMs as probability-weighted linguistic dice, but I have final say over whether a material or concept association is valid. If I don’t like it, I nix it. I have creative control.

The Haptic Translator

The app I am building to help me analyse interviews from patients and clinicians. It creates a concept map from scrubbed transcripts (removing all personally identifiable information) and then suggests physical materials associated with the concepts. I can interrogate its choices, choosing to keep some or add my own. This is purely private and local at the moment, but I plan to create a public version which further obfuscates the original text when interviews are complete.

It is possible that what we might call the AI revolution has a different kind of influence over my practice. There is a palpable feeling of distrust in surface appearances and in the mutability of words and images. There is a hope that something deeper is possible, but an uneasy questioning of whether it’s attainable and how. What I find interesting is that questions about the ‘being-hood’ of LLMs can raise questions about our own. And that I think will be the subject of my next dispatch.

Until then!

This research and development project is supported using public funding by the National Lottery through Arts Council England.

Google Sites

Report abuse