Creating a new recipe is a sensory experience for a chef. Far from being just an outcome of a procedural kitchen protocol, delicious dishes are profound culinary statements.
Starting from a vast number of ingredients and infinite ways of combining and processing them to arriving at a lovely meal is a process of gastronomic discovery. This act of creation requires a combination of genius, expertise, and intuition. Even master chefs with multiple culinary innovations to their credit may struggle to explain the underlying process behind an innovation.
Until recently, culinary experiences were exclusively placed in the realm of subjective sensations. The complex interplay of taste, odour, visual, auditory, and tactile sensory mechanisms make cooking exceedingly complex, beyond the reach of quantitative sciences. However, the success of AI-driven approaches for imitating creative processes, such as poetry, painting, and music, has heightened expectations for their use in the culinary space.
Cooking is a unique cultural phenomenon at the core of human creative endeavours. Recipes, flavour, nutrition, and health form the pieces of the food jigsaw. In the last decade, computational gastronomy has evolved as a formidable data science aided by structured compilations of culinary databases, domain-specific algorithms, and applications.
For instance, RecipeDB, a structured compilation of recipes from global cuisines and underlying algorithms that can identify culinary entities, is ripe for exploration. This organized repository can help estimate the nutritional value of dishes, among other things.
Similarly, FlavorDB represents the taste (gustatory) and smell (olfactory) correlates of food ingredients, capturing the molecular and evolutionary reasons behind their use in cooking. Data from studies on the health effects of food and estimated carbon footprints of recipes highlight important health and environmental impacts, crucial for public health and sustainability.
Attempts have been made to develop new recipe generation algorithms based on text generation strategies and large language models (LLM). Ratatouille is powered by over 118,000 structured recipes compiled from around the world, with culinary nuances obtained from state-of-the-art named entity recognition algorithms. However, unlike LLM-generated poetry, painting, or music, validating computer-generated recipes is arduous and resource-intensive.
Aligned with Alan Turing's historic essay questioning the ability of a machine to think like humans, the 'Turing Test for Chefs' attempts to test the ability of a computer to generate human-like recipes. This framework randomly presents to an expert chef a recipe from a stack of traditional recipes or those generated by a computer. Based on expertise and intuition, the chef is then required to assess the recipe as either fake or genuine on a scale of zero to five. Zero stands for a fake recipe beyond doubt, and five represents an authentic recipe.
Final-year students at the Institute of Hotel Management in Delhi, were engaged to test the Ratatouille algorithm's ability to generate recipes that could fool a chef. The results showed reasonable success (with an F1-score of 69.88%) for the LLM-based models in creating new recipes.
For example, the Ratatouille-created Thai shoyu burrito recipe was so convincing that even a chef thought it was genuine. The algorithm provided detailed instructions on ingredients and quantities, pre-processing steps, cooking steps, and the appropriate utensils and methods, as well as the time, style of preparation, and serving suggestions.
However, the model can be made better by understanding the complex syntax and semantics of the recipe. The proof of the pudding is in the eating when it comes to culinary creativity. The real test of success will be when recipes can trick chefs into thinking they were made by humans. Such recipes would need to be cooked and judged by experts to see if they taste right. Even though this judgment can be subjective, it sets the stage for using AI in gastronomy.
Novel recipe-generation strategies have the potential to create tasty and nutritious dishes with public health consequences. When combined with appropriate data, these can also yield personalised, cost-effective, and environmentally sustainable recipes. However, taste will remain a crucial factor in the acceptance of AI-generated recipes. Therefore, the skill, intuition, and creativity of chefs will be indispensable to make sense of the output of AI. It would be interesting to see whether an AI-generated recipe could win the MasterChef competition.
While the increasing availability of culinary data and the complex, multi-sensorial experience make gastronomy an excellent territory for AI-assisted discovery, only humans will be able to tell a good dish from a bad one. While AI will act as a catalyst of human ingenuity, it won't be a substitute.
