There’s no arguing that AI still has quite a few unreliable moments, but one would hope that at least its evaluations would be accurate. However, last week Google allegedly instructed contract workers evaluating Gemini not to skip any prompts, regardless of their expertise, TechCrunch reports based on internal guidance it viewed. Google shared a preview of Gemini 2.0 earlier this month.
Google reportedly instructed GlobalLogic, an outsourcing firm whose contractors evaluate AI-generated output, not to have reviewers skip prompts outside of their expertise. Previously, contractors could choose to skip any prompt that fell far out of their expertise — such as asking a doctor about laws. The guidelines had stated, “If you do not have critical expertise (e.g. coding, math) to rate this prompt, please skip this task.”
Now, contractors have allegedly been instructed, “You should not skip prompts that require specialized domain knowledge” and that they should “rate the parts of the prompt you understand” while adding a note that it’s not an area they have knowledge in. Apparently, the only times contracts can skip now are if a big chunk of the information is missing or if it has harmful content which requires specific consent forms for evaluation.
One contractor aptly responded to the changes stating, “I thought the point of skipping was to increase accuracy by giving it to someone better?”
Shortly after this article was first published, Google provided Engadget with the following statement: “Raters perform a wide range of tasks across many different Google products and platforms. They provide valuable feedback on more than just the content of the answers, but also on the style, format, and other factors. The ratings they provide do not directly impact our algorithms, but when taken in aggregate, are a helpful data point to help us measure how well our systems are working.”
A Google spokesperson also noted that the new language shouldn’t necessarily lead to changes to Gemini’s accuracy, because they’re asking raters to specifically rate the parts of the prompts that they understand. This could be providing feedback for things like formatting issues even if the rater doesn’t have specific expertise in the subject. The company also pointed to this weeks’ release of the FACTS Grounding benchmark that can check LLM responses to make sure “that are not only factually accurate with respect to given inputs, but also sufficiently detailed to provide satisfactory answers to user queries.”
Update, December 19 2024, 11:23AM ET: This story has been updated with a statement from Google and more details about how its ratings system works.