It was March 2022 and I was in Dubai to visit the pandemic-delayed Expo 2020. I was in the lobby of the Museum of the Future—which had opened just three weeks before I arrived—and in front of me was something I hadn’t heard of before: Algorithmic Perfumery.

The idea was simple. You paid them USD$50 and they would give you a short questionnaire to fill out. You handed it back and then they used your answers as input to “AI [that] supplements your data with existing learning to create scents that are hyper-personal”. The fragrances were created by a machine with 32 distinct scent components (they now have a large-scale machine capable of mixing over 500 ingredients). The machine would send 5 ml (0.17 oz) bottles down an assembly line; you could watch as it created your three personalized fragrances.

I thought it was a cool idea and happily paid up. I watched enthusiastically as the machine created my selections. When they were complete, an attendant packaged them into a presentation box. My own unique fragrances, drawn from what I assumed was at least billions of possible combinations—how exciting!

I opened the box right there at the museum and… they were fine. Just… fine. I didn’t mind any of them, but I didn’t particularly want to wear any of them, either. I took the fragrances back home at the end of the trip, thinking someday I’d give them more of a try, but I just didn’t like them enough to want to wear them. Eventually they fell victim to a spring cleaning, never to be seen again.

What went wrong? To answer that leads to still deeper questions: Is AI capable of creating a signature scent that can hold its own in a world crowded with luxury fragrances? Why do we choose luxury fragrances to begin with?

This isn’t a review of Algorithmic Perfumery. I think it’s fascinating technology and I assume there are people who love the fragrances it has created for them. I didn’t, but that’s irrelevant.

My point is that if there were some groundswell of support for Algorithmic Perfumery, or for similar technology, I would have expected it to have long since appeared in a kiosk below la Coupole in Galeries Lafayette on boulevard Haussmann. Why hasn’t it? I have three ideas.

Perhaps combining scent components into desirable fragrances is a process that is completely reliant upon the experiences, emotions, and senses inherent in being a human being. If this is true, automating the creation of fragrances might be extremely difficult or even impossible with our current level of technology.

I tend to doubt this. We have a long history of people saying things are impossible and then being shown the fool. (See Clarke’s first law for the definitive version of this observation.) Also, whenever I see an argument going down this path, I immediately think of quantum mysticism, especially the concept that somehow our brains are special because of something that relates to quantum physics. I’ve always taken the view that “if meat can do it, silicon can eventually do it”.

For a few years, my closest Nordstrom was in San Antonio, Texas. There was—perhaps there still is—a particular gentleman who worked at their men’s fragrance counter, where they had all the fragrance brands combined into a small area with one salesperson. (Men just don’t get the love at US fragrance counters.) When I say he was a scent whisperer, I mean it. He would take the most fragmentary of input and produce something a man would love. Sometimes it would take two or three iterations for him to converge on a solution. Other times it took just one. I visited once with someone and all he said to the salesperson was “I’m in the Navy and I don’t like most fragrances.” In seconds, the salesperson said he knew just what to try and produced a bottle of Acqua di Giò Profondo Eau de Parfum by Giorgio Armani. It was saline, citric, and fresh, and everything my companion didn’t know he wanted in a scent yet suddenly did, and voilà, he walked out with his first luxury fragrance.

I tell this story because I believe that we’ll see an AI that is a scent whisperer like that salesperson at some point. When? I can’t say. But again, my position is that if meat can do it, silicon can do it.

Perhaps the idea is a good one, and achievable today, but as created by Netherlands-based ScenTronix (the company behind Algorithmic Perfumery), it’s not good enough. It could be that their AI code, whatever form it may take, isn’t sophisticated enough to reliably combine scent components in ways that are compelling to people.

This seems possible to me. One thing that was missing from my experience was any sort of feedback loop. If the process had been…

“We’ve created a set of test strips for you to smell. Here’s an app you can use to tell us your reactions. We’ll use your reactions to create another set of strips, and we’ll keep doing this until we converge on a scent that you find irresistible.”

…I suspect I might have liked the results far more.

I’ll come back to this idea below with an experiment you can try yourself. But that brings us to the idea that I think is closest to the truth…

What if our attraction to a particular fragrance is based in part on its provenance? Think of your favorite luxury fragrance. To you, it smells good on its own, it smells good on (some) people, and it smells good on you. Could it be that part of the attraction, though, is the name on the bottle?

This hypothesis seems highly possible to me. We ascribe value to luxury items based on their price, on their exclusivity, and on their heritage. “A rose by any other name would smell as sweet”, but with all due respect to the Bard, would it really? Would you wear your favorite luxury fragrance if it came from a brand sold in USD$30 gift sets at the local drugstore?¹ Honestly?

If this hypothesis is correct, the Algorithmic Perfumery approach is doomed from the start. Humans like stories. We like to tell them, and we like to hear them. Most of all, we like stories connected to people—stories about people like Thierry Hermès and Louis Vuitton, and about the people who carry on their traditions. Stories about machines aren’t nearly as interesting.²

Here's an experiment you can try right now. Make a list of your favorite fragrances of all time—those you come back to and wear again and again. Provide that list to your large language model (LLM) of choice. Tell the LLM to research your fragrances, identify the scent components within them, then create a profile of the types of fragrances you like. Tell the LLM to research profiles and reviews of other fragrances and identify some that it thinks you’d like, explaining its reasoning behind each of its choices.

Here's the prompt I used when I did this myself nearly a year ago:

I used ChatGPT in its deep research mode, and about 15 minutes later, I had what struck me as an excellent report. I’m going to quote a line from the report verbatim because it read so well:

AI or no AI, that was solid writing. So far, so good.

Enthusiastically, I took the report into a store and started trying. The first fragrance I smelled was the lighter Versace Pour Homme, which ChatGPT described as “an everyday versatile citrus aromatic… a modern classic often recommended as a ‘must-try’ for neroli lovers”. I’m sure that for many people, it indeed is a “modern classic”, but it didn’t do a thing for me.

My next stop was a heavier fragrance, Oud Satin Mood by Maison Francis Kurkdjian. Here’s part of how ChatGPT described it for me: “You absolutely must try Oud Satin Mood. Many fragrance enthusiasts consider this one of Kurkdjian’s masterpieces in the oriental genre, often mentioning it in the same breath as the best rose-ouds on the market.” I smelled it on a test strip and wasn’t sure, so I decided to spray some on my hand… and it was so overpowering that it ended my fragrance investigation for the day. As with Versace Pour Homme, I’m sure Oud Satin Mood is a wonderful fragrance for the right person. Francis Kurkdijian is a highly respected parfumier and many of his works are beloved by fragrance enthusiasts. But this particular fragrance of his truly wasn’t for me.

When I resumed my search a couple of days later, it didn’t go much better. I tried fragrances ChatGPT had recommended from Armani, Kenzo, Montblanc, and Versace without finding anything I liked.

(I want to emphasize that I’m not a fragrance reviewer and these aren’t fragrance reviews. If you’re interested in learning more about the world of fragrances, you could do worse than starting with the review site Fragrantica.)

By no means do I wish to dissuade you from trying yourself the experiment I described. Quite the opposite, in fact. It might work for you. It didn’t work for me—but in a way, that’s the point.

ChatGPT had every advantage. It had detailed descriptions of each of my favorite fragrances. It had its conversational history with me and so knew a tremendous amount about my background and interests. It referred to over 100 reviews of fragrances in its search and was able to draw on their detailed descriptions, both objective lists of scent components and reviewers’ subjective reactions. It wasn’t constructing a scent but instead recommending existing fragrances created by renowned designers. Yet with all that, I didn’t like any of the first half-dozen scents I tried based on its recommendations.

Now think about the approach used by a company like Algorithmic Perfumery. When I tried it in Dubai, what did it know about me? My answers to a few general questions about myself and that was it. If ChatGPT couldn’t get it right, even knowing as much as it did about me, and recommending popular fragrances, what chance did Algorithmic Perfumery have? Little to none.

I don’t want to dissuade you from trying out an AI-driven fragrance creation process. As I’ve noted in this article, I’m not an expert on fragrances—far from it. And just because something didn’t work for me doesn’t mean it won’t work for you.

That said, trying out an LLM-driven fragrance identification process might not be the worst place to start. If you have access to an LLM with a deep research mode, it’s effectively free except for your time. Take my prompt from above, add your favorite fragrances to it, and see what it comes up with.

I should point out that ultimately—and, to be sure, indirectly—my ChatGPT-assisted investigation eventually led me to a brand and combination of fragrances that I liked. More about that another time.

Bonne chance!

What is your signature fragrance? How did you discover it, and what makes it so special to you?

If you’ve tried an AI-generated fragrance, or if you tried the suggestion above to have an LLM recommend existing fragrances to you, what were the results?