Anyone who has tried early-generation vegan cheese or rubbery plant-based burgers knows the open secret of the sustainable protein industry: saving the planet is a tough sell if the food doesn’t taste good.
Consumers might be highly motivated by climate change, but at the dinner table, taste remains king. According to NECTAR’s Taste of the Industry 2026 report, a mere 33% of consumers say they actually “like” dairy-free products—roughly half the satisfaction rate of traditional animal dairy. Meat alternatives face a similarly dismal reception.
Today, Food System Innovations (FSI) is launching a new initiative aimed at closing that gap. Dubbed the Food Intelligence Lab, the interdisciplinary program is stepping out of stealth to build the open-source data infrastructure required to inject AI into the sluggish world of food formulation. Backed by a $2 million grant awarded last fall by the Bezos Earth Fund, the lab’s mission is highly specific: use machine learning to make sustainable protein taste better, faster.
The Wet Lab Bottleneck
Currently, food scientists spend months, if not years, in a trial-and-error loop trying to perfect texture, mouthfeel, flavor, and aftertaste. Startups face limited budgets and siloed, fragmented datasets, meaning that every time a company wants to formulate a new plant-based yogurt, they are effectively starting from scratch.
While AI has radically accelerated R&D in fields like drug discovery and materials science, food science has largely been left behind. The Food Intelligence Lab wants to change that by acting as a centralized, open-source hub.
“AI is already transforming fields like drug and materials discovery, but food still lacks the shared infrastructure needed to fully unlock the potential of AI in this space,” said Anna Thomas, Director of Machine Learning at the Food Intelligence Lab and a computer scientist at Stanford University. “We’re building tools to help food scientists iterate faster and create truly exceptional sustainable protein products.”
The lab is generating large-scale datasets that pair instrumental measurements, like pH, texture profile analysis (TPA), and shear tests, with human sensory data. The ultimate goal? To accurately predict consumer outcomes before a product ever hits a human taste-tester’s tongue.
AI in the Kitchen: The Proxy Foods Case Study
To prove the platform isn’t just theoretical vaporware, FSI has already partnered with AI-powered formulation company Proxy Foods AI to test its mettle.
The two teams co-developed an optimization system called EGBO (Expert-Guided Bayesian Optimization). They set it loose on a notoriously difficult product: plant-based Greek-style yogurt. Leveraging a specialized AI food scientist agent, EGBO achieved a 29% improvement in sensory performance over just 10 formulation iterations. The process took only five days.
The resulting AI-optimized yogurt managed to match its animal-based benchmark on three out of four key metrics: consistency, creaminess, and tanginess. When pitted directly against a professional human food scientist given the exact same time constraints, the AI system won out, achieving a utility score of 0.992 compared to the human’s 0.850.
“Food scientists shouldn’t have to spend months on trial-and-error to get texture right,” noted Panos Kostopoulos, Founder and CEO of Proxy Foods AI. “At Proxy Foods AI… our goal is eventually reaching 90% in silico and reserving the wet lab for the final 10% of validation and optimization.”
Open-Sourcing the Taste Test
Perhaps the most compelling aspect of FSI’s new lab isn’t just what it’s building, but how it’s sharing it. Instead of walling off their algorithms to create a proprietary moat, they are open-sourcing the infrastructure.
Because running human sensory panels is notoriously expensive and time-consuming, the lab has launched TasteBench, a new benchmark and competition available publicly on Kaggle. It challenges researchers to tackle sensory prediction at both the food and molecular levels.
FSI’s own baseline foundation models are already showing promising results. In early tests, their AI proved just as capable of predicting sensory similarity to an animal product as a human panelist (achieving 66.1% pairwise ranking accuracy vs. a human median of 65.0%).
Why It Matters
The stakes for this kind of technological acceleration are incredibly high. Food systems generate roughly 26% of all global greenhouse gas emissions. Livestock alone is responsible for over half of that figure, including massive outputs of highly potent warming gases like methane (44% of global emissions) and nitrous oxide (53%).
Decarbonizing our diets is non-negotiable for climate targets, but consumer behavior won’t shift at scale until alternative proteins achieve price and taste parity.
“We believe AI can be a powerful accelerator for climate and nature solutions when it is paired with the right data, collaboration, and real-world applications,” said Dr. Amen Ra Mashariki, Director of AI at the Bezos Earth Fund. He noted that the Food Intelligence Lab is exactly what their AI Grand Challenge was looking to fund: a bridge between open science, AI, and sustainable food.
For the alternative protein industry, which has seen venture capital cool off over the last few years as consumer adoption plateaued, this kind of infrastructural overhaul is exactly what’s needed. If FSI and Proxy Foods can truly move 90% of food R&D in silico, we might finally see a wave of sustainable proteins that consumers actually want to eat.
