Patronus AI launches Gemini-powered tool to detect handcrafted hallucinations for Etsy

Patronus AI is revolutionizing AI evaluation with the launch of its multimodal large language model-as-a-judge technology, specifically designed to assess AI systems that process images and generate text. This innovation addresses critical challenges in detecting hallucinations and reliability issues in multimodal AI applications, with Etsy already implementing the technology to verify caption accuracy across its vast marketplace of handmade and vintage products.

The big picture: Patronus AI has created Judge-Image, an industry-first evaluation tool built on Google’s Gemini model to help developers ensure accuracy in AI-generated image captions and descriptions.

• The technology enables automatic evaluation of image captions based on multiple criteria, including hallucination detection, object recognition, location accuracy, and text analysis.
• Etsy has become an early adopter, using the technology to validate AI-generated captions across its marketplace containing hundreds of millions of handmade and vintage items.

Why Google over OpenAI: Patronus deliberately chose Google‘s Gemini over OpenAI‘s GPT-4V after extensive comparative research.

• “We tended to see that there was a slighter preference toward egocentricity with GPT-4V, whereas we saw that Gemini was less biased in those ways and had more of an equitable approach,” explained Anand Kannappan, cofounder of Patronus AI.
• The research revealed Gemini offered more uniform scoring distribution across different evaluation sources, making it better suited for objective assessment.

Surprising insights: Patronus discovered that conventional wisdom about AI evaluation doesn’t necessarily apply to multimodal systems.

• Unlike text-only evaluations where multi-step reasoning typically improves performance, Kannappan noted it “typically doesn’t actually increase MLLM judge performance” for image-based assessments.

Beyond retail applications: While Etsy represents a flagship e-commerce implementation, Patronus envisions broader applications across multiple industries.

• Marketing teams could utilize the technology for “scalably creating descriptions and captions against new blocks in design, especially marketing design, but also product design,” according to Kannappan.
• Law firms and venture service companies could leverage the technology to extract and summarize information from PDFs and other documents, replacing “legacy technology” currently used for these tasks.