UL’s leap into the genAI evaluation business raises key questions

UL Solutions, part of the UL enterprise that grew out of Underwriters Laboratories, on Monday jumped into the crowded genAI third-party evaluation service market, joining Stanford University and Microsoft, among many others, but with a more customized approach. The UL team will be asking questions as well as analyzing code. 

Some analysts and others in the AI space have questioned how reliable and precise such an effort would be. Would the workers handling the value-add genAI code know the answers to those questions? Even more cynically they ask whether the workers — or contractors — would answer all questions honestly, or would they be more likely to tell the UL team what they think they want to hear, so that they can get the most favorable rating?

Another issue — and this applies to all of the third-party genAI evaluation efforts — is that the foundation model is off limits. But when evaluating questions about bias, accuracy, and related topics, the foundation model absolutely colors the results. 

Forrester VP/principal analyst Brandon Purcell said the foundation model is a critical element of these kinds of evaluations.

“Even if you get perfect answers to 20,000 questions, an AI is going to have a black box in the middle of it. These foundation models are opaque,” Purcell said, raising questions about the value of the analysis if the foundation model is off-limits.

Another analyst, Paul Smith-Goodson, VP/principal analyst for Moor Insights & Strategy, agreed with the need to factor in the foundation model in any safety or reliability analysis.

“To evaluate layers on top of that, they are also going to have to evaluate how it interacts with the model below it,” Smith-Goodson said. “I think it’s very difficult, very complicated to evaluate the safety aspects of AI because there are so many models. How can they take all of that into consideration? I just don’t know how they are going to do all of that.”

Still some benefits

Despite the potential limitations, Forrester’s Purcell said there are still some benefits to what UL is planning.

“I do think that this is meaningful, but it’s also the best we can do right now, given that there is no transparency,” Purcell said. “This is largely in service of marketing, trying to help companies bridge the trust gap. They really need to bridge that trust gap.”

Assaf Melochna, president of AI vendor Aquant, said good third-party evaluations of genAI code are “increasingly needed, given the hype and explosion of new AI solutions over the past year.”

That said, Melochna stressed that this might be beyond the realistic capabilities of a third party.

“I’ve worked with UL in the past and respect the quality of their work, [but] this rating system risks oversimplifying the complexity of enterprise needs. The UL certification might provide a baseline level of trust, helpful in filtering out weaker solutions, but it shouldn’t replace the in-depth evaluations required to find AI tools that solve specific operational challenges,” Melochna said. “The UL rating should be treated as one piece of the puzzle, not the whole picture. CIOs will need to pair these ratings with their own due diligence to ensure that AI solutions align with business goals.”

src=”https://b2b-contenthub.com/wp-content/uploads/2024/10/UL-Verified-Mark-for-AI-Model-Transparency-example.png” alt=”Sample UL Verified Mark” loading=”lazy” width=”400px”>UL’s new offering “assesses AI model transparency, which is the ability to understand how an AI system makes decisions and produces specific results,” said the company’s news release. “By examining key areas such as data management, model development, security, deployment, and ethical considerations, the benchmark provides a clear and objective rating of an AI system’s transparency and trustworthiness that results in a marketing claim verification.

“A UL Verified Mark for AI Model Transparency may be displayed on products achieving a rating. Systems are awarded a score between 0 and 100 points, with higher scores indicating greater transparency. A score of 50 or less is considered ‘not rated,’ indicating significant transparency issues. Scores between 51 and 60 are rated as Silver, reflecting moderate transparency. Scores between 61 and 70 are rated as Gold, indicating high transparency. Scores between 71 and 80 are rated as Platinum, reflecting very high transparency. Scores of 81 and above are rated as Diamond, indicating exceptional transparency.”

Jason Chan, the UL Solutions VP for data and innovation, said in a CIO interview that the company chose to only evaluate what the enterprise adds on top of the foundation model. 

“Any AI model by design is not transparent. How did it learn all of that stuff? Can you explain how the model works? What was included in the training set of data? How well has it removed data bias?” Chan said. “This is a big undertaking: a couple of months for us to complete the assessment.”

When asked what kind of pricing an enterprise could expect, Chan said it would be based on the enterprise’s needs, the size of the various AI models, and what needs to be evaluated. He declined to offer any guidance on specific pricing.

“How much data is involved and how big is your model? The pricing model is twofold: the survey evaluation and the initial assessment,” he said. “There are also elements of licensing fees so that you can continue to be kept up to date. We need to reassess every 12 months because of data drifting, as the models start to deviate from the original intent.”