Machines can smell. That may not be a big step for mankind, but in the development of Artificial Intelligences (AIs) it was one of the final frontiers to sensory perception that approximates the nature of mammals. They could hear, see and feel. Now they can differentiate between flavors, track down gas and rotten meat, find out that there is a smell in the taxi or that the new car smells wonderful. If the digital conquest of humanity is now enough for you, then you are on the right track, because smell and taste are the two senses that reach directly into the emotional center without the filters of experience and cultural imprinting. This makes Covid such a difficult matter psychologically because the loss of smell and taste is not life-threatening, but the psychological effect is enormous. The loss of the sense of smell – with or without Covid – leads to severe depression and mental disorders. A colleague’s tweet illustrates this quite well in the opposite direction: “Yes, sex is great – but have you ever slowly regained your sense of taste after a corona infection?”
There are certainly still enough boundaries between man and machine. There are some things that AI will never learn. Feelings, understanding and creativity, for example. The impression that machines are competing with their creators is nevertheless strengthened, especially because the term for the technology is so unfortunate. It was thought up by the computer scientist John McCarthy 65 years ago for the third-party funding application for a computer conference, because it sounded much nicer than the actual term “complex information processing”. But AI has little to do with intelligence, but above all with automation. And it caused as much damage as it did blessing in the last industrial revolution. If you would still prefer to order your own plaice with a team of oxen, you don’t have to read any further.
Therefore, the digital conquest of human abilities is initially only a (admittedly instinctively strong) perceived problem. Because the better the sensors on the machines work, the more they can do. Most importantly, they can make more precise decisions. The better the data you collect, the better the results. And since AI is not about intelligence, but about the automation of increasingly complex decision-making processes, smell and taste remained an unattainable stage of development for a long time. Apart from the fact that the science of smells is still surprisingly underdeveloped, even if the beginnings go back to antiquity. Lucretius was one of the first to deal with it when he also formulated a “theory of smell” in “On the Nature of Things” in 55 BC.
This is mainly due to the amount of data that the nose and palate process – incessantly. When the first basic research appeared four years ago in the weekly journal of the American Society for the Advancement of Science, the comparison formula was very simple.
Mixing perfumes is a dream destination
If you direct a beam of light with a wavelength of 510 nanometers at an eye, the person sees the color green – as easily as it is unambiguous. However, there are around 400 odor receptors in the human body that are used to differentiate between smells. Rockefeller University in New York was preparing a test in which 476 odorous substances were to be classified using 19 terms such as “fishy”, “sweet” or “burnt”. At the same time, the test subjects were asked to assess how intense or pleasant a smell is. The end result was a data set with over a million readings.
That was the basis for the “Olfaction Prediction Challenge”, a competition for olfactory prognoses, which the IBM research center announced. 22 science teams from all over the world applied. The results were a next step. From now on one could assign certain molecular structures to certain experiences. Molecules that were similar to those of the vanilla bean, for example, create a smell that is reminiscent of a bakery. Sulfur elements in molecules, on the other hand, act like a garlic note.
There are enough commercial applications. Mixing perfumes is a dream goal. If you take the classic Chanel No 5, for example, which was the first perfume with a composition of different fragrances in 1921, you quickly understand that the time for development with an olfactory AI would be enormously shortened. Chanel No 5 consists of 31 different fragrances. The top note is a so-called accord of three aldehydes: C10 occurs naturally in orange oil, C11 in coriander leaves, C12 in bitter oranges. Then the nose perceives the middle note of jasmine, rose, lily of the valley, iris and ylang-ylang and finally the base note of sweet grass, sandalwood and cedarwood, vanilla, patchouli, civet and musk.
31 different fragrances, the top note alone is a so-called accord of three aldehydes: the classic Chanel No. 5.
(Photo: BENOIT TESSIER / REUTERS)
Such a number of fragrances would still overwhelm KIs. But before the perfume, there is the food, household goods, auto and machine industries that are interested in such applications. The start-up company Aryballe from Grenoble is primarily aimed at companies that maintain vehicle fleets that AI can use to find out which of their vehicles should be taken out of service for cleaning. Have you ever driven a sharecar that smells as if the previous user had eaten garlic sausage and a pack of cigarettes?
The benchmark is not humans, but dogs
According to Aryballe, the food industry would primarily benefit from the fact that quality control could be automated. An application that should also be of interest to manufacturers of kitchen appliances when it is ready for the market. Why shouldn’t the refrigerator determine which foods are still safe to eat and which are not? And then there is the consumer goods industry. It can be anything. The literal scent brand of a car, a detergent or a shower gel is one of the most complex development projects there is.
Even if research into human olfactory perception and the development of artificial olfactory sensations are still in their infancy, the path is already clear: the more criteria, predicates, molecular structures and active substances that are entered into the data sets, the more diverse and precise the applications will be. The benchmark is not the human being, but the dog. Dogs can track down diseases, drugs and fugitives. That already indicates where the possibilities and dangers lie.
Artificial intelligences that use biochemical sensors to detect diseases, toxins or gases can save lives. Sniffer AIs, on the other hand, can open completely new gates to the darkness for surveillance. Everyone has an olfactory key that identifies them. In the next steps, odor sensors could identify fear and nervousness.
But one thing remains from research now. With smelling and tasting, the training of the machine senses will come to an end. Computer science and sensor technology are still a long way ahead of robotics. And one thing shouldn’t be forgotten: even a machine that can identify smells better than a human or even a dog will never experience them. Whether the scent of Chanel No 5, fresh bread or a bouquet of violets – for an AI it is only biochemical data.