AI startups want to crack open recipe book in Big Food’s test kitchens
In the world of big food, artificial intelligence is nothing new.
McCormickThe company, which owns brands like Frank’s RedHot, Cholula and Old Bay, has been using AI in flavor development for nearly a decade, and the company says development timelines have been shortened by an average of 20% to 25% by identifying promising flavor combinations and narrowing down which ideas are worth testing in physical prototypes.
It’s a similar story UnileverA platform where artificial intelligence is deeply embedded in food research and development, where systems can digitally test thousands of recipes in seconds and arrive at actionable concepts with fewer physical trials. For example, Unilever’s Knorr Fast & Flavorful Paste was developed in approximately half the normal time. On the packaging side of things, AI modeled how formulations behave in Hellmann’s Easy-Out squeeze bottle; The company says this saves months of physical laboratory work.
In 2017, a team from Google Brain (now part of DeepMind) Used artificial intelligence to help create the “perfect” chocolate chip cookie recipe.
But while AI is increasingly shaping the way food companies decide what goes on grocery shelves, food companies are quick to emphasize that AI is not taking over the kitchen.
“Human creativity and judgment are leading the way, and AI is a tool that will help us increase our impact,” said Annemarie Elberse, head of ecosystems, digital and data for food R&D at Unilever.
“These tools help inspire the creativity of our taste scientists,” Anju Rao, McCormick’s chief scientific officer, told CNBC. Rao emphasized that artificial intelligence does not replace human expertise, but functions as a tool for co-creation. “Our greatest asset will always be our employees, who bring global perspectives, flavor expertise and human creativity to the table,” he said.
While the growing startup ecosystem is positioning AI as a way to predict and predict sensory outcomes by using large data sets to model how consumers might respond to new food products before they are physically tested, it’s unclear how successful efforts to crack the code in the test kitchen will be. Companies like Zucca, Journey Foods, NielsenIQ and AKA Foods market their platforms as “virtual sensory” or AI-powered systems designed to digitally scan recipes, suggest formulation changes and predict consumer taste before physical prototypes are made.
These companies promise much of what food giants say they’re already doing: creating systems that can reduce the size of traditional flavor panels, reduce the risk of failed launches, and compress product development cycles by identifying promising concepts early in the process. Industry analysts predict the global market for AI in food and beverage will grow roughly From $10 billion in 2025 to over $50 billion by 2030It is supported by increasing investments in data-driven product development, automation and personalization.
But some early food AI pioneers have already moved on. McCormick’s initial AI work was developed in partnership with IBM, which has previously researched AI-focused food projects such as Chef Watson. An IBM spokesman said in a statement that the company was “no longer actively focused on this area.”
Food scientists who have tested these platforms say the technology behind the marketing language is still early and many of the claims are about attracting capital rather than replacing human expertise.
Brian Chau, food scientist and founder of food science and food systems consultancy Chau TimeHe said many AI food startups are still in the data collection phase and trying to gather enough real-world information to make their models meaningfully predictive.
“I think all emerging AI companies, to some extent, overestimate what they can do, which is true of most startups,” Chau said. “They need real industry partners to attract investors, to build data sets, and for any of this to really work at scale.”
Chau said most existing platforms resemble large language models trained on existing recipes, production data and consumer trends, rather than systems that can independently produce viable new products. “When I tested a platform, I found that the output was basically the output you would get from any general AI system,” he said. “Without the private data of real companies, not much added value could be achieved.”
In his view, the technology’s long-term potential depends on whether startups can form partnerships with large food manufacturers willing to share internal formulation data; This is something many companies are reluctant to do due to intellectual property concerns. “Without the big players in the industry feeding real data into these systems, it is very difficult for them to make real predictions,” Chau said. “It’s a numbers game.”
AI is where food science still falls short
From a scientific perspective, researchers say the biggest hurdle is biology, not computing power.
D., professor of sensory and consumer science at the University of California, Davis. Julien Delarue said expectations for AI-powered sensory tools may be inflated due to misunderstandings about what AI can realistically model. “I would say it’s probably a bit of an exaggeration,” Delarue said. “This doesn’t mean that AI isn’t useful; it’s just not something people expect from it.”
While AI can help analyze chemical data and increase efficiency in food development, Delarue said it remains fundamentally limited in trying to predict how humans will perceive complex flavors. “I’m trying to predict what people will perceive from a complex mixture of compounds; the answer is no,” he said.
He explained that one of the main challenges is that human sensory perception is inherently variable. People perceive the same chemical compounds very differently depending on genetics, culture, experience, and even personal history. “There is no such thing as the average consumer,” Delarue said. “Trying to guess what the ‘average’ person might perceive is probably a dead end.”
To overcome this limitation, Delarue says, we would need much more data than we currently have; access to data at an individual level, knowing what each person or group actually perceives. “And that’s a huge task,” he added.
This variability, he said, makes it difficult for any one model (human or machine) to serve as a universal representative of taste.
Even the companies that develop these tools emphasize the centrality of human judgment.
David Sack, founder of AKA Foods, said his company’s platform is designed to organize in-house R&D knowledge, not to replace food scientists or sensory experts. “Food R&D teams rely on a huge amount of valuable information, from past formulations and sensory data to confidential know-how held by individuals,” Sack said. “But it is often fragmented and difficult to systematically reuse.”
Why will humans remain tastemakers?
AKA’s platform helps teams digitally test ideas before embarking on physical trials, allowing scientists to focus on the most promising formulation routes. “It doesn’t replace food scientists or sensory experts,” he said. “Ultimately, humans define the goals, constraints, and success criteria. Sensory experts design and interpret the panels. Scientists decide what to test and what to launch. AI may reduce the number of tests required, but it does not eliminate the need for actual human tasting or validation. When the end consumer is human, humans will always need to be in the loop,” he said.
“Consumers decide whether they like a product based on their own tastes,” said Jason Cohen, founder and CEO of Simulacra Data, which uses artificial intelligence to analyze sensory and consumer data. “We’re still starting with real human sensory data. AI helps us extrapolate insights faster and cheaper.”
Cohen, who is also the founder of Analytical Flavor Systems, which was acquired by NielsenIQ in 2025, said AI is most useful not for altering human perception but for identifying undesirable flavors, narrowing down formulation options and prioritizing which ideas are worth testing.
Chau says large food companies are uniquely positioned to leverage AI-powered tools because they already control large amounts of proprietary formulation, sensory and production data; This is something many smaller brands are still trying to establish.
Delarue thinks the real value of AI in the food industry will be in efficiency, not creativity; It helps researchers analyze data faster, manage complexity, and work under increasing constraints on health, sustainability, and cost. “Designing food today is much more difficult than it used to be,” he said. “You don’t just want to make food that people will enjoy. You need to produce food that is healthy, sustainable and affordable. Artificial intelligence gives us more power to tackle this complexity.”
But when it comes to taste, the reference point is still humans. “Consumers will always be the ones who decide what tastes delicious,” he said. “Not machines.”
