Key Takeaways:

• By using AI to predict irritation earlier in the process than traditional testing, brands can reduce animal testing and refocus resources. 
• Models trained on human skin outperform animal proxies in both regulatory relevance and real-world accuracy.
• What was once slow and limited can now scale, changing how ingredients are screened industry-wide.

For all its technological advances, beauty safety testing has long relied on slow, expensive systems that are detrimental to animals, that don’t always reflect how human skin actually behaves, or that may cause  unnecessary discomfort to human subjects. As regulatory scrutiny tightens and animal testing bans expand globally, brands face a growing challenge: how to innovate faster while meeting ethical, scientific, and commercial expectations.

A new collaboration between Osmo and the Institute for In Vitro Sciences (IIVS) points to a meaningful shift. By combining large-scale, non-animal skin irritation testing with machine learning trained exclusively on human-relevant data, the two organizations have created a predictive model that could change how cosmetic ingredients are screened, long before they reach formulation.

“This project demonstrated that this type of work can be done in vitro without the use of animals,” said Gertrude-Emilia Costin, Director of Laboratory Services and Study Director at IIVS, to BeautyMatter. IIVS, a nonprofit founded in 1997, has spent nearly three decades advancing non-animal testing methods and working with regulators to validate them. However, the Osmo collaboration pushed the organization far beyond its usual operating scale.

“Prior to this project, a big week for us would be testing about 20 materials,” Amanda Ulrey, President of IIVS, told BeautyMatter, also highlighting that this collaboration quickly turned to a hundred every week.” Over 10 months, IIVS tested roughly 3,000 individual molecules for skin irritation using reconstructed human epidermis models, a method already validated for regulatory use. What was unprecedented was applying it at true high volume.

“When we say high throughput, people usually think of robots or cell-based assays,” Costin explained. “Working with individual tissues at this level, I don’t think anybody envisioned that to be a possibility.” The project required extensive planning, dedicated teams, and continuous weekly execution. “It was intense,” Costin continued, “but we learned a lot, and now we have an SOP if we ever embark on something similar again.”

Why Human Data Matters More Than Ever

For Osmo, the motivation was rooted in a persistent industry bottleneck. “Skin irritation is a thing that actually stops a lot of new ingredients from coming to market,” Richard Whitcomb, Osmo’s CTO and co-founder, said to BeautyMatter. Traditional animal testing, he noted, isn’t just ethically fraught, it is also scientifically limited.

“Animal testing doesn’t always match human data,” added Jacob Sanders, Senior Machine Learning Engineer at Osmo. “You can have false positives and false negatives.” That mismatch is particularly problematic for beauty, where safety claims must hold up across regulatory review and consumer scrutiny. Osmo’s goal was to build a predictive model trained entirely on human-relevant, non-animal data, something Sanders described as critical. “That’s what regulators care about,” he said. “Having a model trained on human data makes it much more relevant.”

Each week, IIVS generated irritation data for approximately 100 new molecules. Sanders then retrained Osmo’s machine-learning model continuously, refining its accuracy over time. “At the beginning, when you’re only in the range of a few hundred molecules, the accuracy kind of bounces around,” Sanders explained. “But when you start to get to about a thousand molecules, you see something really exciting.”

By the end of the study, the model had been trained on roughly 3,000 molecules, and its performance was still improving. “I don’t think it finally leveled off,” Sanders said. Crucially, the AI didn’t just analyse data, it also helped direct the study. Sanders selected which molecules should be tested next based on where the model was least certain, ensuring balanced datasets and more efficient learning.

For IIVS, ensuring data integrity was non-negotiable. “One thing I’m always concerned about is the whole ‘garbage in, garbage out’ scenario,” Ulrey said. “Behind every data point that we have, there are people following the protocol exactly as it should be.” That rigor became the foundation of the AI itself. “It was critical to be able to trust the data,” Sanders confirmed.

The result is not a system that replaces testing outright, but one that dramatically reduces wasted effort. “If you can predict whether a molecule is a skin irritant or not, you can reduce by a lot the number of molecules you need to test,” Sanders said. “That’s less testing overall.”

What This Means for Beauty Brands

For beauty companies navigating tighter timelines and higher development costs, the implications are significant. Earlier visibility into safety risks could mean fewer late-stage failures, faster ingredient approval, and more confident claims around skin compatibility.

“This is a first proof point towards being able to predict whether a molecule is going to be safe for humans,” said Sanders. “It allows us to focus discovery efforts toward areas of chemical space that have a good chance of making it to market.”

Whitcomb sees this as a stepping stone. “If you can do this for other endpoints—biodegradability, aquatic safety—you could eventually evaluate full formulations in silico,” he says. “That’s where this is heading.”

While regulators aren’t yet prepared to trust AI predictions alone, acceptance is growing, especially when models are grounded in validated science. “We work with the EPA [Environmental Protection Agency], the FDA [Food and Drug Administration], and at the OECD [Organisation for Economic Cooperation and Development] level,” Ulrey noted. “For regulators, training and familiarity is key.” IIVS also brings regulators into its labs to observe methods, review datasets, and work through real case studies. “The better the quality of the data and the methods we put in front of them,” Ulrey added, “the better position we’ll be in.”

At its core, the Osmo–IIVS collaboration offers a blueprint for how AI and laboratory science can work together. It does not aim to be a replacement, but an accelerator. “What this project shows is the power of having a combinatorial approach,” Ulrey said. “In vitro wet lab and AI in silico tools together.” For an industry under pressure to innovate responsibly, that combination may prove decisive.