Why we can thumb our noses at AI

If you are the sort of person who worries about privacy, the AI doomers probably have you in a headlock right now. OpenAI contractors are reading ChatGPT conversations and Meta’s cutesy AI agent Muse is reported to be skipping past permissions. What you look like and sound like can already be replicated with ease. Anonymity is starting to seem like a quaint idea.

But it turns out there is still one — rather niche — identifier that remains yours alone. The way that you smell is, in technology terms, a data modality into which AI has made relatively little headway.

Does this seem like clutching at straws? Maybe. When a technology writer told me last month that, in all seriousness, they now regarded smell as one of the few bright spots in a world of surveillance, it sounded like the sort of thing someone would say when they’d run out of normal ways to be cheerful. What is there to gain from being undetectable and unreplicable in this way? What sort of aroma-based resistance can you mount?

Yet in the world of olfactory neuroscience there are some who take the idea of not being sniffed out by AI seriously. “I have attributes not digitally coded,” says Barry Smith, director of the Institute of Philosophy at the University of London’s School of Advanced Study. “And I’m happy about that.”

Smith spends a lot of time thinking about senses. Smell tends to be the most neglected — or it was until Covid knocked it out for millions of people. And it was an academic backwater for so long that there is relatively little data available for AI models to ingest.

Keeping this sense beyond the bounds of AI is good not just because it helps us to keep a few shreds of privacy intact but because it maintains a dividing line between real and virtual worlds.

“No matter how good the visuals and sounds — no matter if you can move around a virtual world — we will never be really immersed in a digital world if we cannot smell it,” says Smith. That goes for the notion of uploading ourselves on to the cloud by digitally recording our experiences too. Even if you can see and hear every memory digitally, without taste and smell they won’t be as vivid as the real thing.

If AI could log, identify and recreate scents, there would be obvious commercial gains beyond sci-fi memory devices. A machine that can smell disease could become a medical tool. One that was able to detect rotten produce could be used by food manufacturers or agricultural companies.

But it is proving to be a difficult process because smell is such a physical experience. When we sniff something (or someone) we are inhaling tiny little pieces of them. “The smell which is in me is the fusion of the other person’s body with my body,” was Jean-Paul Sartre’s poetic description. This is either comforting or gross, depending on your mindset. But while large language models can simulate other cognitive faculties, smell is, for now, the outlier.

Whether it will stay this way is debatable. Alex Wiltschko, a neuroscientist and AI researcher who spun his start-up Osmo out of Google Brain, says no. He compares olfactory intelligence (aka OI) to the early days of photography, when different methods of capturing images were being tested. “How do we quantify the ephemeral?” he asks. “We know it is possible, but there are different ways to go about it.”

The barrier standing in the way of a perfectly accurate AI nose is, Wiltschko says, data. For this, Osmo is building data labelling infrastructure — “a Rosetta Stone from the world of molecules”. MIT researchers did something similar with SmellNet, a database of real-world smells.

It is an enormous job. Research by Rockefeller University’s Laboratory of Neurogenetics and Behavior suggests that humans can sift between molecule combinations to detect more than a trillion different scents. Compared to vision, audio and language processing, our understanding of how we perceive those molecules is limited. We can know that a rose contains over 400 distinct odour molecules in various sizes and configurations without knowing exactly how we register its smell.

The last memorable attempt to reproduce real-world smells was Smell-O-Vision, a 1960s cinematic experiment involving puffs of scent that tended to linger and mingle in unpleasant combinations. Making a more sophisticated AI version is still a work in progress. Digital reproduction of the way that we smell remains a philosophical question mark. For now, only the human nose knows.

elaine.moore@ft.com