Exclusive: Google’s Gemini forces contractors to evaluate AI responses outside of their own expertise

Generative AI may sound like magic, but behind the development of these systems is an army of workers at companies like Google, OpenAI, and others known as “quick engineers” and analysts who assess the accuracy of chatbots’ results to improve their AI.

But new internal guidance from Google to contractors working on Gemini, seen by TechCrunch, has raised concerns that Gemini may be more prone to spreading inaccurate information about highly sensitive topics like healthcare to ordinary people.

Contractors working with outsourcing firm GlobalLogic to upgrade the Twins It is owned by Hitachiare regularly asked to rate AI-generated answers on factors such as “correctness”.

These contractors have until recently been able to “skip” certain prompts and thus refuse to evaluate various AI-written responses to those prompts if the query was outside of their domain expertise. For example, a contractor may skip a survey that asks a niche question about cardiology because the contractor lacks the scientific background.

But last week, GlobalLogic announced a change from Google that contractors are no longer allowed to skip such prompts, regardless of their experience.

Internal correspondence seen by TechCrunch shows that the instructions previously read: “If you do not have critical experience (e.g., coding, math) to evaluate this command, please skip this task.”

But now the instructions read: “You should not miss instructions that require domain-specific knowledge.” Instead, contractors are told to “evaluate the parts of the request that you understand” and add a note that they lack domain knowledge.

This has led to direct concerns about Gemini’s accuracy on certain topics, as contractors are sometimes tasked with evaluating highly technical AI answers on issues such as rare diseases they have no knowledge of.

“I thought the purpose of skipping was to improve accuracy by giving it to someone better?” one contractor noted in internal correspondence seen by TechCrunch.

Contractors can now pass on instructions in only two cases: if they “do not fully receive information,” such as a complete request or response, or if there is harmful content that requires special consent forms to be evaluated, the new guidelines show.

Google did not respond to TechCrunch’s requests for comment by press time.