Why AI isn’t likely to be replacing human chocolate judges any time soon

---

AI, and the likes of ChatGPT, are now predicted to take over the world and automate (or at least help ‘co-pilot’) everything from translation to legal services, poetry to music composition, writing code to providing psychotherapy. But one area that AI, and even ChatGPT, doesn’t seem likely to take over and automate, at least in the short term, is judging what we enjoy tasting, including craft chocolate. This week we’d like to explore why this is, and also thank all the judges who’ve helped celebrate craft chocolate makers in the recent International Chocolate Awards, Great Taste and Eurobean competitions.

To celebrate these awards we’ve pulled together an award winners box and highlighted a few other awesome award-winning bars too.

And we’ve also tried to explain why this sort of human judgement is going to be one area that AI, at least in the short term, struggles to master or even ‘co-pilot’.

---

Our 2023 award-winners collection boxes:

---

Taste: Art or Science?

For most of our senses and experiences we’ve developed standardised criteria and scientific measurements. Time has minutes, hours, years, etc. Distance has miles (or kilometres). Weight has kilos (or pounds). Sound has decibels (and pitch, rhythm, frequency, tone, etc.). Temperature has Celsius (or Fahrenheit or Kelvin). Colour, since Newton and his prisms, has wavelengths to describe red, yellow, and even stuff we can’t see; and of course the Pantone system.

Food and drink are different. They are about flavour, taste and texture. Measurements and judgements are a lot more subjective. And they are potentially far harder for large language models to absorb and generate, with AI, than images similar to the one above (courtesy of DALL-E), or summaries (such as the one below we asked ChatGPT to produce for this blog), or performing near simultaneous translations, creating jingles and music etc.

Awarding prizes and awards in food and drink is all done by hand (or rather mouth and nose) and by teams of people. The Great Taste Awards is all about lots of people tasting lots of products and calibrating judgements across multiple judges. Magazines and organisations like Decanter, The IWC, IWSC, etc. taste tens of thousands of wines each vintage for their awards. And the Academy of Chocolate and International Chocolate Awards do a similar job for chocolate, each now tasting thousands of bars every year.

No one really has found a way to speed up or automate the process of tasting. This is in large part because we are a long way from having standardised criteria and measurements for ‘taste’ (or more specifically for flavour). We don’t have Newton’s light-refracting prism to help us describe and define aromas and flavours. We don’t have the ingenious concept of the Pantone system to co-ordinate colours across multiple geographies and materials.

Disentangling what we enjoy when we drink and eat is incredibly complicated. To quote Professor Barry Smith:

“…what we ordinarily call ‘taste’ involves input not just from the tongue, but from touch and smell. …The experience often described in unisensory terms as ‘taste’ depends on the multi-modal combining of inputs”.

By this he means that when we delight (or not) in food it is all about our sense of taste, smell, texture, and a few other senses (e.g., spiciness). Untangling these is part of the problem.

Another part of the challenge is that we only have defined measurements for some of these senses. For spiciness we have the Scoville scale. And we do have criteria to measure all the ‘tastes’ (saltiness, sweetness, sourness, bitterness, umami, etc.). Indeed, Itochu in Japan is working on some amazing technologies that are now assessing how different cultures and different ages react to different levels of saltiness and sweetness with an ‘electric tongue’. Texture and mouthfeel are similar, although, as all those startups trying to create artificial “chocolate” and “cocoa butter” can attest, it’s still pretty hard. And as anyone who has tried a mass-produced chocolate bar made with an emulsifier like PGPR will know, the mouthfeel here is VERY different to the delights of cocoa butter.

Flavour (olfaction, our sense of smell and flavours) is particularly challenging. It wasn’t until 1991 that Linda Buck and Richard Axel identified where, and how, the brain’s olfactory process works (they won the Nobel Prize for their work here in 2004). But we are still a long way from having anything like Newton’s prism or a Pantone scale to define colour. Olfaction is hugely complicated. For example; just for starters; we detect flavour in two very different ways; orthonasally (through our nose) and retronasally (through our mouth). And this retronasal capability of savouring flavour with our mouths is a unique human attribute; that is to say, other animals can detect sweetness, sourness, saltiness etc. but they can’t savour flavour.

For those of us who are fans of “real food” (and drinks) there is some good news here: It’s really, really hard to recreate flavours that occur in nature. And indeed tinkering too much with nature, and in particular trying to commoditise and focus on costs and volumes diminishes flavour. For example, there is increasing evidence that “factory farming”, with all its pesticides and fertilizers, molly- coddles fruit and vegetables so that they have less reason to develop flavonoids, anthocyanins, and other key compounds that we now know generate flavour; and so yes, there really is a reason why a tomato from a small farm really does taste better. And this is clearly the case with chocolate.

This is not to say that scientists don’t have some tools to study and research; for example, gas mass spectrometers. But if you read how long and laborious it was to make artificial vanilla, you’ll see how far they are from cracking flavours (the work behind synthesizing artificial vanilla has been compared to “trying to figure out what was inside a mysterious piece of luggage by heaving it off a hotel balcony”). With music, as it’s digitised its components are broken down into pitch, rhythm, frequency, etc., and we can understand how these are combined, we can make recommendations (as do Last.fm, Spotify, etc.) and now even start to create jingles. With food and drink, although we can identify the potential aromas and flavours in a product from its chemical composition, we can’t predict which specific aromas give e.g. vanilla it’s distinctive aroma or why this creates for us in the West notions of sweetness.

And to return to the ‘Great Wave of Savouring Chocolate’ we use in our virtual tastings, it is in the “aftertaste” that these artificial additives fall flat. To use some wine-speak; artificial flavours just don’t have “BLIC”; it doesn’t have balance, length, intensity, or complexity. And artificial flavours just can’t give you this journey. For example, nootkatone was identified in the 1960s as giving grapefruit part of its distinctive flavours and, when distilled, it made a great additive that underpinned Fresca. But nootkatone is not the real thing. It doesn’t have the length or complexity from the myriad of other aromas, volatiles, and flavours in a grapefruit.

The same is true for mass-produced confectionery and chocolate. It doesn’t take you on a journey. Mass-produced chocolate is all about giving you an immediate hit that encourages you to scoff more. It does this via the genius of two discoveries that were articulated and innovated during the 1960s and 1970s by Howard Moskowitz. The first is known as ‘the bliss point’, i.e. the combination of lots of sugar, salts, umami, and fat, that keeps you coming back for more (the so-called “Pringle effect”: Once you pop, you just can’t stop). And the second is “sensory-specific satiety”, that is to say keeping your taste buds stimulated and preventing them from being sated via different textures, tastes, flavours etc. (think how much easier it is to eat 3 scoops of different ice creams versus 3 scoops of the same ice cream). And yes, that’s why most of us can’t resist the last Rolo or Malteser (and I guess Reese’s peanut butter drops if in the US).

---

Join our FINAL ‘deep dive: taste and flavour’ session on September 21st. Explore the biology, chemistry and craft of flavour in our dedicated session before it winds down this year:

---

What AI can and can’t do…

LLMs (large language models), AI (artificial intelligence) and the likes of ChatGPT and OpenAI are truly amazing at everything from playing ‘go’ to creating limericks and lyrics, and from creating imagery to composing music. One of the fundamental building blocks here is access to large pools of data to ‘train’ the algorithms to “identify a cat”, summarise an essay, write a poem, compose some music, translate from one language to another etc.

For food, the data around the bliss point and sensory specific satiety is available via, for example, ingredient labels which detail sugars, salts, fats, acids, etc. And the likes of ASMR videos provide a wealth of data around texture. So be prepared to see some ‘interesting’ new snacks, confectionery, fizzy drinks that are suggested by AI. And they may well manage to turbocharge even more scoffing.

If this does happen, one more prediction: These new products are likely to be in brand extensions for existing products rather than new brands created via AI. One of the more curious aspects of mass chocolate and confectionery histories is why the interwar years saw such an AMAZING burst of new brands: Crunchie (1929), Snickers (1930), Mars Bar (1932), Kit Kat (1935), Milky Bar (1936), M&Ms (1941), etc. Since then, mass-produced chocolate has been more like modern day Hollywood. That is to say, it’s been more about sequels than breakouts; whereas in the 1980s, 1990s, and early 2000s there was an explosion of great new movie concepts which until recently have dominated most cinemas (e.g. Star Wars, Avengers, Top Gun, Harry Potter). But who knows, maybe AI will also help crack the problem of new product launches and customer acquisition too.

BUT cracking ‘flavour’ is very different, and a truly tough nut to crack, even for AI. There isn’t a framework (yet) that fully explains why we detect certain aromas and not others when we savour a great chocolate bar (or glass of wine, cup of tea, etc.). There isn’t a set of definitions or dictionaries like the Pantone colour system for flavour, or dictionary to translate, etc. And it’s only recently that we’ve started to understand how flavour is so individual, and why some people detect specific aromas in foods or drinks. For example, only about 80% of people are blessed with the Fusobacterium nucleatum bacterium in their saliva which enables one to detect ‘gooseberry’ flavours in many New Zealand sauvignon blancs. And that’s why different people will have radically different experiences from the same chocolate. Compounding the difficulty of studying this is the peculiarity that humans are unique in detecting flavour in our mouths, so we can’t study other animals’ behaviour and reactions to deepen our understanding.

Flavour is also ‘cross-modal’ with other senses; especially vision and touch; being critical in helping people identify foods, flavours, etc. Touch dramatically impacts our enjoyment and sensation of any food or drink (come to a virtual or in-person tasting and see how stroking a beard (or some velcro) can impact your enjoyment of any chocolate bar!). And for a fascinating example of the importance of vision, check out the Chef’s Table episode where 20 of the world’s top chefs, when blindfolded, struggled to identify even half of the dishes presented to them (including the likes of mango cut into chunks). Indeed, next week we’ll explore more on the importance of recall and vision, and why ‘verbal overshadowing’ means that asking a witness to describe a criminal immediately after the crime normally REDUCES their ability to identify the criminal in a lineup, and why this is so important for learning to savour and enjoy flavour.

AI aficionados point out that this lack of a framework and definition didn’t stop Deep Mind’s ‘Alpha Zero’ trouncing human champions in the Chinese game of Go even though the formal rules, and all the permutations of plays, weren’t “explained” to Alpha Zero. And this may well be true. But for now, the “raw data problem” means there is no “training data set”. For flavour we also don’t have a definition system for turning words (or numbers) comparable to the Pantone system for colour, or even the Scoville scale for Spiciness, etc. in training data.

We don’t understand how tastes, textures, flavours and our other senses interact at an aggregate or individual level. For example; why does stroking a beard (or velcro) impact some people’s perception of a chocolate bar but not others? (Come to a virtual or in-person tasting to try this out). And arguably artificial flavours further muddy the picture and compound this issue. That is to say, by exposing us to so many artificial flavours (banana, strawberry, chocolate, etc.) in ultra-processed foods we’ve now added another layer of complexity to understanding the myriad of amazing flavours that real strawberries, and of course craft chocolate, can yield.

---

Grab a virtual tasting kit and join our weekly Wednesday virtual experiences to try out some of curious experiences:

---

ChatGPT’s Response

For those interested, after writing this, I asked ChatGPT for its take on when AI would be able to understand, and generate, flavours similar to the way DALL-E generates images. It took ChatGPT just over 3 seconds to produce the below (and this 3 seconds is apparently deliberately slowed down). This 3 second turnaround is more than a little sobering given this blog took me a number of hours to research, construct and write:

“As of my last knowledge update in September 2021, creating an artificial intelligence system that can understand and generate flavors similar to how DALL·E generates images is a complex and challenging task. DALL·E, created by OpenAI, is a text-to-image generation model that can generate images from textual descriptions. Flavor, on the other hand, is a multi-dimensional sensory experience that involves taste, smell, texture, and more. It is a much more complex and subjective domain.

Predicting when AI will be able to create an understanding of flavor similar to DALL·E’s image generation is difficult because it depends on several factors, including technological advancements, research breakthroughs, and the availability of large datasets related to flavor.

To achieve this goal, AI researchers would need to develop models that can:

• Understand textual descriptions of flavors.
• Combine information about taste, smell, texture, and other sensory aspects to generate a comprehensive flavor experience.
• Account for the subjectivity of flavor perception, as different people may have different preferences and perceptions of flavors.

This is a highly interdisciplinary challenge that involves fields such as natural language processing (NLP), computer vision (for analyzing images of food), sensory science, and more. Progress in these fields could contribute to the development of AI systems capable of understanding and generating flavors.

It’s worth noting that there have been some advances in AI-generated recipes and food-related content, but creating a comprehensive understanding of flavor, similar to DALL·E’s image generation, remains a complex and evolving research area. Predicting a specific timeline for such advancements is uncertain and would depend on the pace of technological development and research in relevant fields“.

---

Celebrating the Awards Winners

So anyhow, let’s celebrate all the hard work of our makers, and the tasters, of the Great Taste Awards, Eurobean, and the ICAs. We’ve assembled a box of milk and dark award-winning bars. And as we could only fit four in this box, we’ve included a few more wonderful winners below.

And if you want to know more about the science of taste, flavour please do join a virtual or craft chocolate tasting.

As ever, thanks for your support.

Keep savouring!

Spencer

---

Our 2023 award-winners collection boxes:

---

p.s. If you want to try a truly awesome bar, and one that sadly is a limited edition so probably won’t be coming back any time soon, we’ve a few of Jordi’s AMAZING Bolivian Beni bars that are still bursting with flavour, but as their expiring date is almost upon us we are discounting by 40% to £3.55.

---

p.p.s. As you savour the award-winning bars (or the Jordi’s Bolivia), here are some suggestions on “HOW TO SAVOUR” we’ve pulled together:

“Just as how you are often advised to swirl your wine in a glass, sniff it and then sip it, chocolate has a similar protocol. When you open a bar the classic advice goes along the lines of; admire the shininess of the chocolate, then break off a piece (it should have a clean snap), then sniff it, and finally move on to savouring the bar.

In the process of tasting literally thousands of craft chocolate bars, we believe that there are a few other best practices to make savouring more fun and memorable. These include:

• Try to have a few chocolates on the go at the same time (arguably easier than with wine or coffee). It really highlights the differences.
• Try to share and discuss with friends. It’s more fun, and articulating the notes you detect helps you discover and remember.
• Try to have some ‘crib sheets’; most of the time tastes, flavours and textures are “on the tip of our tongues” but hard for most to articulate. Hence why we always hand out our ‘wave of flavour’ sheets.
• Give it time. Lots of time. Revisit and repeat. Enjoy and savour the journey“.