Developing a new food product traditionally requires months of testing. Currently, artificial intelligence (AI) systems are capable of optimizing this stage by analyzing thousands of substances in mere seconds. These algorithms help researchers find new flavors, improve textures, and achieve balanced nutritional profiles, thereby significantly shortening the innovation cycle in the industry.
However, the computer does not 'invent' a recipe out of thin air. The systems are trained on data regarding ingredient composition, chemical makeup, nutritional properties, taste, and texture. Based on this data, they identify relationships and propose combinations that meet specified goals. Research published last year in the journal npj Science of Food indicates that AI can contribute to the creation of products that are simultaneously nutritious, tasty, and sustainable.
Limitations of AI Technology in Taste Assessment
Nevertheless, the technology lacks the ability to 'taste' the dishes it designs. As explained by Manuela Dolinski, president of the Federal Council for Nutrition (CFN), the systems generate evaluations based on patterns identified in large datasets, such as chemical composition, physicochemical properties, and sensory assessments. 'Based on this data, models calculate the probability that a certain formula will possess characteristics related to acceptability, such as sweetness, bitterness, umami, aroma, creaminess, crunch, or juiciness,' stated the CFN president in an interview with Olhar Digital. However, she emphasized that these predictions represent statistical trends and do not fully replicate the human experience of eating. Manuela notes that taste depends on factors such as genetics, age, habits, and cultural context, making human sensory analysis indispensable.
One of the most well-known examples is NotCo, a Chilean company that developed Giuseppe—an AI system designed to create plant-based food. This platform analyzes ingredients at a molecular level and searches for combinations capable of reproducing the characteristics of animal products. According to the company, the system works with a universal fund of over 300 thousand plants, also considering factors such as taste, texture, nutrition, and functionality.
The goal can go beyond simply creating a plant-based version of a known product. The algorithm can also be given specific dietary goals and search for a formula that comes closest to them. This opens up possibilities for products with increased protein or fiber content and reduced amounts of sugar, sodium, or saturated fats.
From composition to taste, AI begins exploring options that would be difficult to check manually. In a study published in June, researchers used generative AI to create burgers with different priorities, including taste, sustainability, and nutrition. The system was also able to statistically reconstruct the characteristics of the Big Mac without having its recipe, by studying patterns derived from human formulas.
Afterward, the team tested the recipes under real consumption conditions. A group of 101 participants took part in a blind tasting at a restaurant and rated the burgers on a seven-point scale. Two versions, developed with a focus on taste, received ratings equal to or exceeding the ratings of the Big Mac on some criteria. One of them received an average taste score of 5.8 compared to 5.4 for the comparable product.
However, when the goal shifted to nutrition, a serious difficulty arose. The burger developed to improve nutritional value demonstrated a better fit with healthy diet components, including more vegetables, whole grains, and plant-based protein, as well as less refined grains, sodium, and saturated fats. But the result was less appealing to participants: the average acceptability score was 3.8 compared to 5.3 for the Big Mac.
This result highlights one of the main problems of the technology. Improving a product nutritionally does not necessarily mean it will be more appealing to the consumer. Taste, texture, aroma, price, and appearance also influence the choice. In this study, the version created with a priority on nutritional quality lost points in these aspects. Thus, AI can find a nutritionally interesting solution, but it still needs to master what makes food desirable.
Manuela warns that the results of an isolated study should not be turned into a general rule, but acknowledges that different formulation goals can compete with each other. Changing parameters to improve product health can affect factors such as moisture and mouthfeel. Furthermore, many consumers are accustomed to high energy density and intense flavors, which causes a nutritionally favorable recipe to deviate from this familiar norm.
Another result attracted attention due to the combination of health and sustainability. The mushroom-based burger, created with an ecological focus, showed an environmental impact index more than ten times lower than that of the Big Mac in the model used by the researchers. The bean-based version achieved almost double the nutritional score compared to the benchmark burger. These are experimental results, not proof that these products are automatically healthier or more sustainable in any situation.
Replacing animal products with plant-based alternatives designed by algorithms also requires caution. 'The benefit depends on the ingredients used, the degree and purpose of processing, the final nutritional composition, the frequency of consumption, the dietary context, and above all, the product being replaced,' states Manuela. She explains that these substitutes may contain high levels of sodium, fats, or additives, confirming the World Health Organization's (WHO) recommendation that vegetarian industrial products should not be considered healthy just because they are plant-based. 'We must not allow technological or plant appeal to serve as proof of health.'
Industry Adopts Technology in Commercial Applications
Industry is already applying this technology commercially. In September 2025, Magnum, a division of Unilever, announced a partnership with NotCo to use artificial intelligence in reformulation and product development. Among the goals were reducing calories, creating new plant-based options, and finding ingredient alternatives. This step demonstrates that AI is beginning to be used to modify existing products.
This pursuit of outwardly lighter labels raises questions about the actual nutritional quality of these new options. Removing critical ingredients such as sodium or saturated fats is a significant improvement, but it is not an absolute certificate of health. The council president notes that even products with claims like 'reduced sodium' or without Anvisa warning labels must be analyzed holistically, checking whether the replacement of one component has been compensated by excessive addition of another.
This raises the question: will the technology help provide nutritious food, or will it lead to the creation of 'embellished' ultra-processed products? The nutritionist explains that AI is a tool whose results depend on the directives provided to the algorithms. 'If the algorithm is focused only on maximizing taste, cost, shelf life, and convenience, it can contribute to the creation of very attractive ultra-processed products, even if they carry individual nutritional claims,' she notes. 'On the other hand, the same technology can be used to reduce critical components, improve nutritional density, replace certain additives, include natural or minimally processed ingredients, and increase fiber content.'
All this clearly shows that algorithms are not yet ready to replace nutritionists, chefs, or food scientists. The recipe suggested by the computer still needs to be produced, tested, and evaluated by humans. For example, in the burger study, the AI instructions indicated ingredients and quantities, but chefs were required to translate this information into recipes that could actually be made and tasted.
Thus, the trend is not about replacing the product developer with AI, but about changing the way this work is done. Instead of testing several dozen options, researchers can use algorithms to explore a much larger space of combinations and select the most promising ones. The final decision still depends on trials, scientific knowledge, and human judgment.
For nutrition science, this change can be particularly relevant. The same technology used to find a substitute for a specific ingredient can theoretically take into account proteins, fiber, micronutrients, calories, sodium, and fats simultaneously. Recent studies already point to AI as a tool for connecting data on food composition, nutritional science, and bodily response.
However, there is a crucial difference between creating a product with a specific nutritional profile and proving that it improves health. A formula may have less sodium or more fiber on the label, but this is not enough to determine its effect on the body. Clinical trials and long-term assessments remain necessary. AI can accelerate the search for candidates, but it does not eliminate the need to prove that what is created actually works.
'Artificial intelligence should not replace professional judgment. It should serve responsible innovation focused not only on industrial efficiency but also on promoting adequate and healthy nutrition, food and food security, and public health protection,' concludes Manuela.
The most fascinating aspect of this transformation is that artificial intelligence is starting to operate in an area that seems inherently human: determining what tastes good. By analyzing thousands of recipes and the relationships between ingredients, algorithms find patterns that would be difficult to notice manually. The next step may be combining this ability with increasingly specific dietary goals, bringing the food industry closer to a kind of personalized food engineering.