The generative AI can resemble magic, but behind the development of these systems are employees of employees in companies like Google, Openai and others, called “fast engineers” and analysts, which evaluate the accuracy of the results Chatbots to improve their AI.
But a new internal directive transmitted from Google to entrepreneurs working on Gemini, seen by Techcrunch, has aroused concerns that Gemini could be more inclined to make inaccurate information on very sensitive subjects, such as health care, ordinary people.
To improve Gemini, entrepreneurs working with Globalogic, an outsourcing company Hitachi propertyare systematically invited to assess the responses generated by AI according to factors such as “veracity”.
These entrepreneurs were until recently able to “jump” certain guests, and therefore choose to assess various responses written by AI to these guests, if the prompt was outside their domain expertise. For example, an entrepreneur could skip a prompt who asked a niche question about cardiology because the entrepreneur had no scientific training.
But last week, Globalogic announced a change of Google that entrepreneurs are no longer allowed to ignore these guests, whatever their own expertise.
The internal correspondence seen by Techcrunch shows that before, the directives said: “If you do not have critical expertise (for example, coding, mathematics) to assess this prompt, please ignore this task.”
But now the directives are read: “You should not skip prompts that require specialized knowledge in the field.” Instead, entrepreneurs are invited to “assess the parts of the prompt that you understand” and include a note that they have no knowledge of the field.
This has led to direct concerns concerning the precision of Gemini on certain subjects, because entrepreneurs are sometimes responsible for assessing the very technical answers on AI on questions such as rare diseases in which they have no training.
“I thought that the goal was to jump for precision by giving them better to someone?” An entrepreneur noted in internal correspondence, seen by Techcrunch.
Entrepreneurs cannot now skip prompts in two cases: if it is “completely missing information” such as the full prompt or the answer, or if they contain harmful content which requires special consent forms to assess, according to the new guidelines.
Google did not respond to requests for comments from Techcrunch by the time of the press. After the publication of this story, Google, which did not dispute our reports, told Techcrunch that the company “constantly worked to improve factual precision in Gemini”.
“The evaluators carry out a wide range of tasks on many different Google products and platforms,” said Google spokesperson Shira McNamara. “They do not only examine the answers for the content, they also provide precious comments on style, format and other factors. The notes they provide have no impact directly on our algorithms, but when taken in total, are a useful point of data to help us measure the operation of our systems. »»
Updated with Google post-publishing comments.
You can send advice safely to this journalist on signal at +1 628-282-2811.