How Google’s Gemini is Pressuring Contractors to Evaluate AI Responses Beyond Their Expertise

But the magic of generative AI is supported by an army of workers, including “prompt engineers” and analysts, who review chatbot outputs to hone their accuracy. This has been a critical process for companies like Google, OpenAI, and others that are trying to perfect AI. But changes to the guidelines for contractors working on Gemini, Google’s core model, have some concerned the risk of inaccuracies will now increase, particularly around sensitive information such as health.

Previously, contractors hired through GlobalLogic, a Hitachi-owned outsourcing firm, were allowed to skip prompts outside their expertise. For instance, those lacking scientific training could avoid evaluating responses to complex cardiology questions. This system was intended to ensure more reliable evaluations by assigning tasks to appropriately skilled individuals.

However, new guidance from Google forbids the skipping of prompts that require special knowledge. Contractors are now supposed to evaluate what they can understand and add a note stating they lack expertise. There are only two exceptions to skipping: when prompts are incomplete or contain harmful content requiring special permission to review.

This policy shift has raised concerns about Gemini’s reliability on technical subjects. Contractors with limited background in areas like rare diseases are now forced to evaluate responses on topics they barely understand. One worker, in internal communication, questioned the change: “Wasn’t skipping designed to improve accuracy by involving someone more qualified?”

Google has not commented on the changes. In an email after the policy came to light, a company spokesperson said that they constantly try to improve the factual precision in Gemini. They said that contractors provide feedback on style and format, among other things-aside from content accuracy-and that their ratings don’t directly change algorithms but overall contractor input does help track system performance.

These new policies might impact the precision of Gemini regarding very important topics. As it is, questions linger whether this system can take away the steam from high-quality AI responses, particularly for domains that are pretty niche. For the company, this is a key turning point in how Google deals with feedback in terms of improving its AI.