Google trains Gemini with an army of underpaid humans and exposed to shocking content

Google Gemini

Long -awaited, Gemini (Google Bard), the conversational agent fueled by the AI ​​developed by Google, is deployed in the form of a free online service and a mobile application for Android and iPhone devices

• License:
  Free license
• Author :
  Google
• Operating systems:
  Online service, Android, iOS iPhone / iPad
• Category :
  IA

Their role? Intellate the model and correct your mistakes. Except that working conditions are very difficult and these people are exposed to sometimes shocking content. If IA companies present their models as autonomous and knowledge wells that replace humans, the truth is quite different.

Humans exposed to shocking content

Technology depends massively on the hard work of real people. AI is not magic: it is a pyramidal human work system. “These assessors IA occupy the essential invisible intermediate level”declares Adio Dinika, of the distributed Ai Research Institute, at Guardian.

These evaluators moderate chatbots to ensure that the responses are accurate but also protect users from inappropriate content. Rachel Sawyer worked as a general practitioner for Google last year. She explains that she was “Shocked” to learn that it was necessary to deal with “disturbing contents”.

“Not only had I was not warned and I never signed a consent form during integration, but neither the title of the post nor the description mentioned the moderation of content.” explains Rachel Sawyer.

These assessors have no expertise and check accuracy in fields such as architecture, astrophysics or even the medical field. According to an evaluator, she was “Press” To work faster while forming the model on very sensitive medical subjects such as chemotherapy for bladder cancer.

“I imagined a person sitting in her car, learning that she had bladder cancer and looking for what I was editing.” explains this person.

Google disputes that these assessors help make the models smarter. According to a spokesperson for the Mountain View firm, these “Quality assessors” allow “External feedback on products, which help measure the operation of systems, but do not directly impact algorithms or models”.

Ridiculous wages for the work done

It is the Japanese subcontractor Globalogic which employs thousands of Americans to train Gemini. The hourly salary is very low, since we are talking about $ 16 to 21, a bite of bread compared to the astronomical salary of researchers in AI.

“These are expert people who do an excellent writing work, paid below their values, to create an AI model that the world does not in my opinion”says an ia assessor.

Many assessors are undergoing large pressures in the face of tight deadlines and directives that are constantly changing. “We didn’t know where it was led, how it was used, for what purpose”explains another Globalogic contractual.

So yes, there is a massive work behind a technology sold as a revolutionary which puts an end to human labor. However, despite these efforts, the AI ​​continues to hallucinate constantly, like Gemini.

Numériques settles in Beaugrenelle Paris for The most tech days : product demonstrations, use or purchase advice, exchanges with our journalists … Discover the full program.