AI Can Generate a Formula—but Can It Develop a Food Product?
Where artificial intelligence accelerates food innovation, where it still falls short, and why human judgment remains essential
Imagine asking an artificial intelligence platform to create a high-protein instant broth.
Within seconds, it may produce an ingredient list, suggested percentages, a nutrition estimate, several flavor variations and alternatives for reducing sodium or simplifying the label. On the screen, the formula appears promising.
Then the first prototype is prepared.
The protein does not disperse evenly. The broth clumps when hot water is added. The texture is grainy, the aroma disappears too quickly and a lingering aftertaste overwhelms the seasoning. One of the recommended ingredients is difficult to source, another increases the cost beyond the target, and the process required to make the formula work is not realistic for the intended manufacturer.
That is the difference between generating a formula and developing a food product.
Artificial intelligence is already being applied throughout food science and engineering. Current uses include process control, food safety, ingredient and product-quality assessment, sensory evaluation, traceability, supply-chain management and product development. Its role will almost certainly continue to expand as companies collect more structured data and develop tools designed specifically for food systems. (Springer)
The question is therefore not whether AI belongs in food innovation.
The more important questions are:
What should AI be trusted to do? What must still be physically tested? And who remains responsible for the final decision?
AI Can Provide an Answer Before We Have Defined the Real Problem
In food product development, the first challenge is not always formulation. It is defining what the product is truly supposed to accomplish.
A founder may ask AI to create:
A healthier snack
A high-protein beverage
A gluten-free coating
A reduced-sugar dessert
A clean-label sauce
A plant-based alternative to a familiar food
Each request sounds reasonable, but none is yet a complete product-development brief.
What does “healthier” mean for the intended consumer? Does it mean more protein, less sodium, fewer added sugars, greater nutrient density or a shorter ingredient statement?
When will the product be eaten? Is it intended to answer hunger, satisfy a craving, replace a meal, support recovery, provide comfort or create a convenient indulgence?
What texture does the consumer expect? What price will the market accept? How will the product be packaged and distributed? Which equipment will be used to manufacture it? How long must it remain stable? What claims are important, and can those claims be properly substantiated?
AI can help organize these questions, but it cannot decide which product problem is worth solving unless people first give it a meaningful objective.
A poorly defined question can produce an impressive-looking answer that is technically interesting but commercially irrelevant.
Where AI Genuinely Adds Value
AI should not be dismissed simply because it cannot replace the entire development process. Used properly, it can make that process faster, more organized and more informed.
During the early discovery stage, AI can help a development team review research, summarize category information, organize consumer comments and compare possible product directions. It can suggest flavor territories, formats, serving occasions and positioning ideas that the team may not have considered.
It can also help developers explore a much larger number of formulation possibilities than would be practical through bench experimentation alone.
A 2026 study provides an important example. Researchers trained a purpose-built generative model using 2,216 human-developed burger recipes containing 146 ingredients. The system generated one million potential burger recipes and allowed the researchers to explore trade-offs among palatability, nutrition and environmental impact. Selected recipes were then prepared and evaluated by 101 participants. Some of the AI-generated burgers performed as well as or better than the comparison burger in flavor or overall liking. (Nature)
This is a meaningful accomplishment, but it is also important to understand what made it possible.
The research did not simply ask a general chatbot to “create a delicious burger.” It used a specialized model, a structured dataset, defined optimization criteria and human sensory validation. The researchers also acknowledged that their model represented ingredients and quantities but did not fully account for processing, cooking methods, physicochemical transformations or water redistribution. They noted that reproducible culinary results still required standardized preparation, culinary expertise and further physical characterization. (Nature)
This is a good illustration of AI’s most valuable role:
AI can widen the field of possibilities, help identify promising candidates and reduce unnecessary trial and error. It does not remove the need to cook, taste, measure and validate.
Ingredient Substitution Is More Than Replacing One Item With Another
Ingredient substitution is another area in which AI may become extremely useful.
A developer may need to replace an ingredient because of cost, availability, allergens, nutrition targets, label expectations, supply-chain risk or manufacturing limitations. AI can help identify possible alternatives and compare them across multiple criteria.
However, replacing an ingredient is rarely a one-to-one exchange.
An ingredient may contribute flavor, aroma, viscosity, moisture control, browning, structure, emulsification, preservation or mouthfeel. It may also carry nutritional, cultural or regulatory implications. A replacement that improves one variable can easily create problems elsewhere.
A recent review of AI-enabled ingredient substitution emphasizes that successful substitution must account for several interacting domains, including flavor and aroma, functional performance, nutrition and cultural acceptability. This is a more realistic model than asking whether ingredient B can simply replace ingredient A. (MDPI)
AI can help model these relationships and reveal alternatives worth investigating. The product developer must still determine whether those alternatives behave correctly in the actual food.
Optimization Is Not the Same as Product Success
AI is especially powerful when it is given a measurable goal.
It can search for formulas with more protein, less sodium, lower ingredient cost, a higher nutritional score or a smaller estimated environmental impact. But the quality of the result depends on the objective it has been given.
The burger study demonstrates this tension particularly well. Its bean-based nutritious burger achieved a Healthy Eating Index score of 63.12, compared with 33.71 for the reference burger. Yet participants gave the nutritious version significantly lower ratings for overall liking, flavor and texture. They were also more likely to describe it as dry, grainy, bland, earthy or soft. (Nature)
The formula performed very well against the nutritional objective, but the eating experience did not perform equally well.
That does not mean the nutritious burger was a failure. It means the next stage of development would need to solve the sensory problems without losing the nutritional gains.
This is where experienced product developers earn their value. Their work is not merely to maximize one number. It is to balance several objectives that may compete with one another:
Nutrition
Flavor
Texture
Cost
Manufacturing feasibility
Shelf life
Ingredient availability
Consumer familiarity
Brand positioning
Regulatory compliance
A product can meet its protein target and still be unpleasant to consume. It can have an attractive ingredient statement and still be too expensive to manufacture. It can perform beautifully in the test kitchen but become unstable during distribution.
AI can calculate trade-offs. Human teams must decide which trade-offs the consumer and the business will accept.
A Recipe, a Formula and a Commercial Product Are Not the Same Thing
These terms are sometimes used interchangeably, but they represent very different levels of development.
A recipe explains what ingredients to combine and how to prepare them.
A formula expresses those ingredients in controlled quantities or percentages so the composition can be reproduced and evaluated.
A commercial product is a much larger system.
It includes ingredient specifications, supplier consistency, order of addition, mixing time, temperature, shear, equipment capacity, packaging, storage, distribution, shelf life, quality controls, labeling, cost and repeatable production.
A formula may state that a dry broth contains a particular amount of protein powder. It does not automatically tell us whether that protein will disperse in hot water, settle during consumption, interfere with flavor release or create an undesirable finish.
A sauce formula may appear balanced at the bench. During commercial production, however, the heating rate, kettle geometry, mixing pattern and holding time may change its viscosity or flavor.
A coating may adhere correctly during a small kitchen trial but fall off after commercial freezing, shipping and reheating.
These are not exceptions to product development. They are product development.
A formula is a proposed design. A commercial product is a validated system.
AI can contribute to that system, but it should not be mistaken for the system itself.
AI Can Model Sensory Response, but It Does Not Experience Eating
AI can describe food in remarkably convincing language. It can discuss acidity, sweetness, umami, aroma, crunch, creaminess and aftertaste.
That does not mean it experiences those sensations.
A language model learns relationships among words, descriptions and patterns in its training data. It may predict what people are likely to say about a combination of ingredients, but it does not physically smell the aroma, feel the texture or notice how the product changes after several bites.
A 2025 case study asked ChatGPT to evaluate 15 hypothetical chocolate-brownie formulations. The system assigned extremely high overall-quality scores—between 8.5 and 9.5 out of 10—to all of them. The formulations included unusual ingredients such as mealworm powder and fish oil, yet the responses remained predominantly positive. The researcher concluded that AI-generated sensory evaluations require comparison and validation against human sensory panels. (MDPI)
This does not mean AI has no place in sensory science.
When AI is combined with appropriately collected instrumental and sensory data, it can detect patterns that may assist trained evaluators. For example, researchers using artificial neural networks with mass-spectrometry fingerprints predicted selected coffee sensory properties with reported accuracies of 87% to 96% in test samples. The authors presented the method as a tool to assist sensory-panel decision-making rather than simply replace it. (PubMed)
A generic language model making a prediction from written ingredient descriptions is not the same as a purpose-built model trained on physical measurements and human sensory data.
AI may help identify patterns within sensory information, but people are still needed to determine whether a product is enjoyable, appropriate and meaningful within its intended context.
A trained evaluator can detect subtleties that are difficult to capture in a simple product description:
An aroma may appear attractive at first but disappear too quickly. A protein note may become more noticeable after swallowing. A snack may deliver an enjoyable first bite but create dryness or flavor fatigue by the end of the serving. A technically accurate flavor may still feel unfamiliar or culturally inauthentic to the intended consumer.
The issue is not whether AI can make a sensory prediction.
The issue is whether that prediction has been calibrated against the food, process and consumer population that actually matter.
Food Does Not Behave Like a Fixed Digital Object
Food systems are particularly challenging for AI because food ingredients are biologically and physically variable.
The same ingredient may behave differently depending on its variety, season, supplier, particle size, moisture level, storage history or processing method. Consumer preferences are also difficult to reduce to one universal standard.
The USDA-supported AI Institute for Next Generation Food Systems identifies several major challenges: food systems are diverse and biologically complex, reliable ground-truth data can be sparse or expensive, much of the available information is privately held, and human preferences affect every stage of the food system. The institute consequently emphasizes knowledge-driven, human-in-the-loop approaches to food AI. (NIFA Reporting Portal)
This matters greatly during scale-up.
A developer standing at the production line may observe that a powder is being added too quickly, the mixing pattern is insufficient, a kettle has uneven heat distribution or the process is too complicated for the available labor.
The written procedure may be technically correct, yet the operators may be unable to execute it consistently during normal production.
AI cannot smell the product in the kettle, notice an unexpected change in viscosity or see that an operator is struggling with an unrealistic process unless those observations are deliberately captured and entered into the system.
That practical knowledge—built through observation, repetition, failure and adjustment—remains one of the most important parts of commercial food development.
The Greatest Risk May Be a Confident Answer
One of the dangers of generative AI is not simply that it may produce incorrect information.
It is that incorrect information can be presented clearly, confidently and professionally.
A recommendation may include exact percentages, processing temperatures, shelf-life estimates, regulatory explanations or scientific-looking references. The level of detail can make the output appear more reliable than the evidence behind it.
The National Institute of Standards and Technology describes this problem as confabulation: generative AI may confidently present erroneous or false content. NIST also warns that users can place excessive trust in automated outputs and recommends that claims about model capabilities be evaluated through empirically validated methods.
This has serious implications for food product development.
An AI-generated answer should not be treated as final evidence for:
Food-safety decisions
Shelf-life determination
Allergen control
Regulatory classification
Ingredient legality
Label claims
Process authority requirements
Commercial production approval
AI can help identify questions and organize information, but conclusions in these areas must be verified through appropriate scientific testing, official sources and qualified professional review.
Confidentiality also deserves attention. Proprietary formulas, processing parameters, supplier pricing, client information and unpublished innovation plans may have significant commercial value. Before entering confidential information into an AI system, a company should understand the platform’s terms, data-retention policies, privacy protections and permitted uses of submitted material. NIST’s generative-AI framework identifies data privacy and intellectual-property protection as important areas of organizational risk management. (NIST Publications)
AI Should Shorten the Path to Testing—not Eliminate Testing
A responsible AI-assisted product-development process could follow seven stages:
People define the problem.The team establishes the consumer, eating occasion, sensory promise, nutritional goals, price, manufacturing requirements and business purpose.
AI helps explore the possibilities.It supports research, concept generation, ingredient screening, comparison of alternatives and identification of information gaps.
The product developer filters the ideas.Culinary, technical, cultural and commercial judgment determine which concepts deserve physical testing.
Bench prototypes test the physical reality.Ingredients are mixed, cooked and evaluated in the actual food matrix.
Sensory and analytical evaluation measure performance.Human feedback is combined with appropriate instrumental measurements rather than relying on either source alone.
Pilot production tests the system.The team evaluates process feasibility, equipment, packaging, shelf life, sourcing, labor, cost and repeatability.
Qualified people approve the final product.Regulatory, food-safety, quality, operations and commercial teams remain responsible for decisions within their areas of expertise.
The results from each stage can then become better data for the next AI-assisted analysis.
This is not a rejection of technology. It is a more disciplined way to use it.
The Future Product Developer Will Need More Judgment, Not Less
AI may eventually reduce the time required to research ingredients, compare options, organize testing data and generate early formulation directions.
That does not make the experienced product developer less valuable.
It may make that person more important.
When hundreds of ideas can be produced in minutes, the challenge is no longer simply generating possibilities. The challenge is knowing which possibilities deserve ingredients, equipment, labor, testing and investment.
Product development requires more than information. It requires judgment.
It requires understanding why a product is being created, recognizing when a technically correct answer will not satisfy the consumer, identifying what must be tested and taking responsibility when the evidence does not support the original idea.
The future of food innovation is not artificial intelligence versus human expertise.
It is artificial intelligence guided, challenged and validated by human expertise.
AI can generate the proposed formula.
People still have to develop the product.
References
Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., & Roberts, K. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. National Institute of Standards and Technology, NIST AI 600-1. DOI: 10.6028/NIST.AI.600-1. (NIST)
Cardoso, V. G. K., Balog, J., Zsellér, V., Karancsi, T., Sabin, G. P., & Hantao, L. W. (2025). Prediction of coffee traits by artificial neural networks and laser-assisted rapid evaporative ionization mass spectrometry. Food Research International, 203, 115773. DOI: 10.1016/j.foodres.2025.115773. (PubMed)
Öz, E., & Öz, F. (2025). Artificial intelligence-enabled ingredient substitution in food systems: A review and conceptual framework for sensory, functional, nutritional, and cultural optimization. Foods, 14(22), 3919. DOI: 10.3390/foods14223919. (MDPI)
Pennells, J., Watkins, P., Bowler, A. L., Watson, N. J., & Knoerzer, K. (2025). Mapping the AI landscape in food science and engineering: A bibliometric analysis enhanced with interactive digital tools and company case studies. Food Engineering Reviews, 17, 465–489. DOI: 10.1007/s12393-025-09413-w. (Springer)
Tac, V., Gardner, C. D., & Kuhl, E. (2026). Generative artificial intelligence creates delicious, sustainable, and nutritious burgers. npj Science of Food, 10, 199. DOI: 10.1038/s41538-026-00953-x. (Nature)
Torrico, D. D. (2025). The potential use of ChatGPT as a sensory evaluator of chocolate brownies: A brief case study. Foods, 14(3), 464. DOI: 10.3390/foods14030464. (MDPI)
U.S. Department of Agriculture, National Institute of Food and Agriculture. (2020–2026). AI Institute: Next Generation Food Systems. Accession No. 1024262, University of California, Davis. (NIFA Reporting Portal)
Comments