Somewhere in a laboratory that looks nothing like a perfumer's studio, no organ fitted with hundreds of brown bottles, no blotters fanned out on a desk like a paper peacock, no stained leather apron hanging on a hook behind the door, a machine is composing a perfume. The machine smells nothing. It has no nose. It has no opinion on whether vetiver pairs well with grapefruit, no instinct for whether a composition needs more lightness up top or more warmth at the base. It has data. It has roughly four hundred thousand formulas from the past century, digitized and tagged with consumer panel scores, sales figures, regional preferences, and molecular descriptors. It has an algorithm trained to identify statistical correlations between specific ingredient combinations and specific consumer outcomes: purchase intent, perceived quality, emotional association, repurchase likelihood. And it has been asked to produce a formula that will be, by every measurable standard, optimal.
12 min read
It will succeed. The formula it produces will score well in consumer panels. It will test positively across multiple demographics. It will offend no one. It will confuse no one. It will occupy a comfortable, well-populated region of the olfactory space, the kind of territory the industry calls the ‘commercial sweet spot’ and anyone with a working nose calls ‘familiar.’ It will smell, in the judgment of most people who encounter it, perfectly fine.
The question is whether ‘perfectly fine’ is perfumery.
The application of machine learning to perfume development is not speculative. It is happening right now, at industrial scale, in the research divisions of the world's largest fragrance and flavor companies. The technology varies in sophistication: some systems are relatively simple predictive models that suggest ingredient substitutions based on cost and availability; others are deep neural networks trained on decades of proprietary formula data, but the underlying logic is the same in every case. Feed the machine a large corpus of existing formulas paired with consumer response data. Let the machine learn the statistical relationships between molecular composition and human preference. Then ask the machine to generate new formulas that maximize the probability of a desired consumer outcome.
This is, in essence, regression analysis applied to perfumery. It is not, in any meaningful sense, creation.
The distinction matters, and it is worth being precise about why. Regression analysis, the mathematical technique at the heart of most machine learning, finds the line of best fit through a cloud of data points. It identifies the central tendency. It tells you where the average sits. This is enormously useful for many applications. If you want to predict real estate prices, consumer behavior, disease trajectories, or election outcomes, knowing the average tells you a great deal. But perfumery is not a prediction problem. It is, or at least it has historically been, a creative problem. And creative problems are not solved by finding the center. They are solved by finding the margin.
Every perfume that genuinely changed the industry, every composition that, in retrospect, defined an era or opened a new category, did so by deviating from the consensus of its time. The first modern fougère was not what anyone expected from a men's fragrance in 1882. The first great aldehydic floral was not what anyone expected from a women's fragrance in the 1920s. The first fresh masculine built on dihydromyrcenol and hedione was not what anyone expected from a men's fragrance in the 1980s. The first clean-skin musk molecular was not what anyone expected from any perfume in the early 2000s. In every case, the composition succeeded not because it matched existing preferences but because it created new ones. It did not find the center. It moved the center.
An algorithm trained on historical data cannot, by construction, move the center. It can only find it. It can find it with surgical precision, and it can generate formulas that occupy the sweet spot with an efficiency no human perfumer could match. But occupying the sweet spot is not innovation. It is optimization. And the history of perfumery suggests that optimization and innovation are not the same thing, and may in fact be opposed.
There is a counterargument, and it deserves serious consideration. The counterargument runs as follows: human perfumers are also, in a sense, algorithms. They are biological neural networks trained on a corpus of olfactory data, everything they have ever smelled, every formula they have studied, every consumer response they have observed over a career. Their creative process is not, as romantics like to imagine, a flash of inspiration descending from the muse. It is pattern recognition, recombination, and iterative refinement. The perfumer sits down at the organ, selects materials based on experience and intuition, blends a trial formula, evaluates it, adjusts, evaluates again. The process is empirical, not mystical. If a machine can perform the same operations faster and more systematically, what exactly is lost?
What is lost is error.
This sounds paradoxical, so let us be precise. Human perfumers make mistakes. They overdose an ingredient and discover that the overdose creates an effect they had not anticipated and could not have predicted. They accidentally contaminate a trial batch and discover the contaminant adds something interesting. The history of synthetic breakthroughs in perfumery is scattered with such happy collisions. They misread their own notes and combine materials they had not intended to combine, and the result is better than what they had planned. The history of perfumery is littered with these accidents, compositions that owe their character not to deliberate design but to an unplanned collision of materials that a more careful process would have prevented.
An algorithm does not make these mistakes. An algorithm does exactly what it is told to do. It optimizes the objective function. It follows the gradient. It does not wander into unexplored territory by accident, because it does not wander at all. It moves, with mathematical precision, toward the optimum. And the optimum, as defined by consumer panel data, is always the center. The average. The consensus.
The creative potential of error is not a romantic fantasy. It is a well-documented phenomenon in every creative field. The biologist who discovers penicillin because of a contaminated petri dish. The physicist who discovers the cosmic microwave background because of unexplained noise in an antenna. The musician who discovers a new harmonic language because a string broke mid-concert and forced an improvisation. These are not apocryphal stories told to comfort the clumsy. They are documented occurrences of a general principle: creative breakthroughs often come from deviations from the plan, and systems designed to eliminate deviation will also, by construction, eliminate the possibility of the breakthrough.
A second objection to computational perfumery, more philosophical, concerns the nature of preference itself.
Consumer panel data, the data these algorithms are trained on, measures stated preference. It records what people say they like when asked. But stated preference and actual preference are not the same thing. Stated preference is conservative. When asked to choose between the familiar and the unknown, most people, in most contexts, choose the familiar. This is not stupidity. It is a well-documented cognitive bias, the mere-exposure effect, first described by the psychologist Robert Zajonc in a landmark 1968 paper in the Journal of Personality and Social Psychology, and it operates powerfully in olfactory evaluation, where the absence of a shared vocabulary makes it exceptionally hard for consumers to articulate why they like or dislike something. Faced with a genuinely new perfume, one that fits no existing category, that confuses and intrigues in equal measure, a consumer panel will, more often than not, score it low. Not because the perfume is bad, but because the panel has no framework for evaluating it.
An algorithm trained on consumer panel data inherits this conservatism. It learns that novelty is risky and familiarity is safe. It learns that the perfumes people rate highest are the ones that most resemble perfumes they have already rated highly. It learns, in short, the most fundamental lesson of consumer research: people like what they already like. And it optimizes accordingly.
The result is a machine supremely skilled at producing what the industry calls ‘safe bets,’ perfumes that will not fail, that will reach a minimum threshold of commercial viability, that will not surprise, disturb, or challenge anyone who smells them. These perfumes will sell. Some will sell very well. But they will not change the industry, because changing the industry requires producing something consumer panels do not know how to score. The compositions that changed perfumery were, at the moment of their creation, all surprises. They were things no one had asked for, things that scored poorly in preliminary tests, things that succeeded not because the data said they would but because a single person, a perfumer, a creative director, an entrepreneur, believed in them despite the data.
An algorithm cannot believe in anything despite the data. Believing despite the data is the one thing an algorithm is constitutionally incapable of. An algorithm follows the data. That is its virtue and its limitation. And in a field where the most important decisions are the ones that contradict the data, where the entire history of creative advancement is a history of people ignoring the consensus and being right, that limitation is not minor. It is fundamental.
Let me be clear about what I am not arguing. I am not arguing that artificial intelligence has no role in perfumery. It has obvious and valuable applications. It can speed up the reformulation process when a regulatory change forces the withdrawal of a restricted ingredient. It can suggest cost-effective substitutions that preserve a composition's character while reducing its price. It can analyze large sets of consumer feedback data and identify trends a human analyst might miss. It can map the vast multidimensional space of possible ingredient combinations and highlight regions human perfumers have not yet explored, in much the same way gas chromatography once decoded formulas that had previously been locked away as trade secrets. These are useful functions. They save time, cut costs, and expand the perfumer's toolkit. No serious person objects to any of that.
What I am arguing is that these are all optimization functions. They make existing processes more efficient. They do not create. The distinction between optimizing and creating is not semantic. It is the distinction between finding the best route across a known landscape and discovering that the landscape extends beyond its known borders. Machine learning excels at the former. It is structurally incapable of the latter, because the latter requires, by definition, going beyond the data, and machine learning is, by definition, a method for extracting patterns from data.
The perfume industry's enthusiasm for computational tools is understandable. The economics of modern perfumery are brutal. The average development timeline for a commercial fragrance has been compressed from years to months. Briefs are tighter. Budgets are smaller. The cost of failure is higher. In this environment, a tool that can reduce the number of iterations needed to reach an acceptable formula is enormously valuable. But ‘acceptable’ is doing a lot of work in that sentence. An acceptable formula is one that meets the brief, scores adequately in testing, and stays under the cost ceiling. An acceptable formula is not a masterpiece. It is not even, in most cases, particularly interesting. It is adequate. And adequacy, at industrial scale, is the enemy of art.
There is a final consideration, and it may be the most troubling. The more perfume development relies on algorithmic tools trained on consumer data, the more the industry's output will converge toward a statistical average. Every new AI-optimized formula will occupy, by design, the center of the preference distribution. Over time, the center itself will shift, but slowly, because the algorithm's output reinforces the very preferences it was trained on. Consumers repeatedly exposed to AI-optimized perfumes will develop preferences shaped by those perfumes, and those preferences will in turn become the training data for the next generation of algorithms. The result is a feedback loop: the machine produces what people like, people learn to like what the machine produces, and the machine produces more of it.
This is not a hypothetical scenario. It is an accurate description of what has already happened in other creative industries that adopted algorithmic recommendation and generation systems. Music streaming platforms, whose algorithms optimize for engagement, have produced a measurable convergence in the sonic characteristics of popular music: louder, shorter, more repetitive, with the chorus arriving earlier and the dynamic range narrowing. Social media platforms, whose algorithms optimize for attention, have produced a convergence in the visual characteristics of popular content: more saturated, more tightly cropped, more emotionally extreme. The algorithm does not flatten the landscape deliberately. It flattens it as a side effect of optimizing for the average.
Perfumery is not immune to this dynamic. If the industry's development pipeline becomes increasingly dependent on AI tools that optimize for consensus, the inevitable result is a narrowing of the olfactory field. Not a narrowing to a single perfume, the market is too large and too segmented for that, but a narrowing within each segment. Fresh masculines will converge. Sweet femininities will converge. Amber oud fragrances will converge. Every category will become more internally homogeneous, because the algorithm designing each new entry is trained on the same data that produced the existing ones. The field will not narrow to a point. It will narrow to a cluster.
Whether this matters depends on what you think perfumery is for. If it is an industry, a trade producing consumer goods designed to meet market demand, then optimization is the right strategy, and convergence is an acceptable cost. Consumers get what they want. Companies make money. No one complains.
But if perfumery is also an art, a creative discipline whose purpose extends beyond satisfying existing preferences to revealing new possibilities of olfactory experience, then convergence is not a cost. It is a catastrophe. Because art, by any definition worth defending, requires the possibility of surprise. It requires the possibility that the next composition will be something no one has smelled before, something no dataset predicted, something a consumer panel would have rejected because it fit no existing category.
An algorithm cannot produce that. A perfumer can. Not reliably, not consistently, not on schedule, not on budget. But occasionally, unpredictably, against all commercial logic, a human being sitting at an organ surrounded by hundreds of brown bottles will combine materials in a way no machine would have suggested, and the result will be something genuinely new. Something that moves the center rather than occupying it. Something the data said should not work.
These moments are rare. They are becoming rarer. And if the industry is not careful, they will stop happening altogether, not because the technology forbids them, but because the economics no longer leave room for them. The machine will compose. The machine will optimize. The machine will produce perfumes that are perfectly fine, that score well on every panel, and that offend no one.
Whether that counts as perfumery is a question the machine is not equipped to settle. It will have to be a nose that answers.
Seven extraits at 20%, one collection. The Discovery Set brings all seven together in 2 ml.
