The Potential of AI in the Perfume Industry

Since OpenAI announced ChatGPT, an artificial intelligence system for natural-language generation, in November 2022, the use of AI has spread rapidly across a variety of industries and jobs, and ordinary people can now easily access and use it. In addition to ChatGPT, Google’s Google Bard and Microsoft’s Bing, which uses ChatGPT’s functionality, are now available. Drawing on my own experiences of using them, I would like to explore the potential of artificial intelligence in the perfume industry.

The Strengths and Weaknesses of Natural-Language-Generating AI

To begin with, the AI attracting attention recently is artificial intelligence in the context of natural-language generation; it cannot do everything. I think it is important to understand its strengths and weaknesses and use it appropriately.

Artificial Intelligence and Learning

It is no exaggeration to say that whether users can teach natural-language-generating AI is one of the things people hoping to use it most want to know. To give the conclusion first, the answer is that both “yes” and “no” are correct.

For example, ChatGPT has already been trained using OpenAI’s enormous database, and this is something that OpenAI can do; users cannot make ChatGPT remember particular information. The system is designed not to remember each chat, so it cannot learn new information through conversation. In other words, even if a malicious user tries to teach it misinformation, that user cannot make it learn in a way that affects ChatGPT as a whole. However, because it learns “how to respond appropriately in conversation” from interactions with users, its ability to generate responses tailored to particular needs can improve.

Does that mean it cannot be done? There are things that can be done: using the API to carry out machine learning. For example, if I use a paid API, the ChatGPT I use becomes my own dedicated version, and I can train it through machine learning. This does not affect ChatGPT as a whole; only the ChatGPT accessible through my API becomes able to answer based on what I have taught it. The difficult part is this machine learning: you need to prepare appropriate prompts (instructions) and answers and feed them in. I have summarized the key points of machine learning below for reference.

• Prepare a prompt and its answer as a pair and have it learn them (depending on the method, this can be done using CSV).
• Preparing as many pairs as possible from different angles on a single subject enables ChatGPT to give more accurate, richer, and more natural answers.
• For each answer, a short to medium amount of text is desirable, with a recommended limit of around 100–300 characters. It is known that teaching ten answers of 100 characters each for one prompt, rather than one answer of 1,000 characters, improves answer accuracy, flexibility, and the user experience.

For example, if you want it to learn about a perfume brand, having it learn summaries organized by topic, such as company overview, brand overview, product categories, and history, will allow it to incorporate that information into conversation more naturally. It can also be made to recommend links.

My Impressions of Using Each AI

Here are my impressions of using three systems: ChatGPT, Bing AI, and Google Bard.

• Bing AI and Google Bard are convenient: with a Microsoft account or Google account respectively, you can start using them in Japanese right away. Bing already has an app, making it easy to use on a smartphone and giving it the lowest barrier to entry. Registering for ChatGPT is a little troublesome, and because it is less well known than Microsoft or Google, you may feel a little uneasy. However, as of May 2023, ChatGPT has also begun offering an app in the United States, so this convenience issue looks likely to be resolved before long.
• ChatGPT has the best understanding of context and excels at natural-language generation. This goes without saying for Bing AI, but even Google Bard understands context less well than ChatGPT, and I think ChatGPT makes more human-like conversation possible.
• Google Bard is the best in terms of information. The language model on which ChatGPT was trained currently uses information from before 2021, and moreover, you cannot tell which sources it used to provide that information. Google Bard, by contrast, also uses Google Search to answer, so it provides more up-to-date and comprehensive information. Bing also identifies its sources based on Bing search results, but taking context and what I want to know into account, I feel that Google Bard offers better conversation. However, the paid ChatGPT Plus plan lets you use ChatGPT with an internet connection, so if you are willing to pay, I feel ChatGPT looks like the best choice.
• ChatGPT comes out on top for business use. At present, ChatGPT is the one that has opened its API, allowing free integration with blogs and companies’ own services. There are also a variety of plugins, with more possibilities on the paid plan. However, the free version of ChatGPT is often subject to restrictions and may be unavailable when you want to use it, so I recommend the paid version.

The Potential of AI in the Perfume Industry

Artificial intelligence has actually been used in various ways in the perfume industry for some time. For example, in 2019 Givaudan developed an AI system called Carto for use in the fragrance-creation process. Carto incorporates all the information about past perfumes, analyzes the works created by a perfumer, identifies their originality compared with earlier perfumes, and suggests new combinations, reportedly giving rise to inspiration unlike anything before.
In May 2019, an artificial intelligence system called Philyra, adopted by Symrise and trained by perfumer David Apel, also produced the world’s first AI-created perfumes, released by a Brazilian cosmetics company. At the time, Symrise reportedly prepared three types of perfume: one made by Philyra, a version to which a perfumer had made slight adjustments, and a perfumer’s original creation. In a blind test, a large majority chose Philyra’s perfume.

The examples above involve a different type of artificial intelligence from natural-language generation, but as you can see, AI is already active in the perfume industry. Naturally, using AI requires enormous databases and enormous funding, so for now it is probably limited to major fragrance companies, but that is surely only a matter of time. Including these types of AI, I see the following possibilities for AI in the perfume industry.

• The displacement of less-skilled perfumers and the emergence of new perfumers
  Perfumers memorize a very large number of fragrance materials, understand their combinations, and create perfumes that are described as art. In reality, however, as in other industries, the line between something called art and something considered lowbrow can be understood only by professionals and those familiar with the field. As long as businesses serve the 90% who do not understand that distinction, this is a natural consequence and inevitably arises. For example, when a new type of popular perfume appears, there are companies that create a similar perfume more cheaply. Even major companies take on this work, and it is necessary. Some perfumers become dissatisfied with it and establish their own brands, but business is not so easy, and it is not uncommon for those brands to come to an end quickly.
  The arrival of AI here creates several possibilities: ① perfumers who make nothing but copycat fragrances and lack the skill to create unique perfumes may be replaced by AI; ② perfumers who want to create perfumes freely as they imagine them may be able to make unique fragrances using the money saved through ①; and ③ a new breed of perfumer may emerge who understands AI and can create unprecedented perfumes even with little perfumery experience.
  
• Fewer sales associates and a new style of customer service
  With the arrival of natural-language-generating AI and its greater ability to understand context, I think perfume recommendation tools could move beyond the conventional approach of “making selections to display fragrances you might like.” Instead, people could describe “the perfumes I like or the image I have in mind,” including information they could not fully explain or felt embarrassed to share in person, and AI could understand it and suggest perfumes based on it. Already, if you ask ChatGPT or similar systems, they will make suggestions that at least 90% of perfume sales associates could not provide. They will also quickly supply an answer to any question, leaving aside whether it is true.
  Looking at this trend, I think today’s perfume sales associates may evolve into three types: ① associates who support AI by handling reception, checkout, and administrative work; ② associates who offer human warmth and empathy that no natural-language system can surpass, along with cross-selling and customer-satisfaction-enhancing suggestions that can hold their own against AI; and ③ associates who fully use AI and manage a store while understanding how to use it optimally. For example, simply placing a few tablets and surveillance cameras in a store could mean that people no longer have to handle tasks such as easing congestion, reducing the work involved in payment, and managing inventory. Department stores and major companies still have a long way to go with digitization, so this may be several years away. But at an up-and-coming store such as NOSE Shop, might these measures make it possible to return the gains to the sales associates who contribute, through better pay?