Americans who have access to Google’s AI Overviews feature have greatly noted the errors of artificial intelligence. Problems partly admitted by the American company, but also contested. Despite everything, the company announces that it has made numerous changes to the operation of its tool.
“ Cheese can slip off pizza for a number of reasons […] You can also add 1/8 cup of non-toxic glue to the sauce to make it stickier. »
This allegation is false. However, it was affirmed by AI Overviews, the new Google tool supposed to summarize its first page of its search engine results with artificial intelligence. Launched in the USA in mid-May after a year of testing, the tool especially made an impression with its ramblings. Other absurd examples have also been reported, with supporting screenshots.
These errors greatly amuse Americans on social networks; others, however, are concerned about the damage this could have in terms of disinformation. The case, in any case, raises the question of the relevance of the agreement signed at the beginning of the year between Google and Reddit, because the first draws in particular on the second to feed AI overviews.
Two weeks later, Google officially admitted the problem in a blog post published on May 30. The head of research, Liz Reid, spoke: of strange and erroneous insights “, while minimizing the extent of the problem – the person concerned mentioned ” a very large number of fake screenshots. »
Among the contested cases is that of a dog left in the passenger compartment of a car. “ Some of these rigged results were obvious and stupid. […] These [réponses] never appeared. We therefore encourage anyone who comes across these screenshots to do a search themselves to verify “, advised Liz Reid.
Google tries to correct the situation with AI
Despite everything, AI Overviews is not entirely satisfactory and, beyond the montages that may have circulated online, there are indeed responses that are not appropriate. Hence a series of actions taken to limit artificial intelligence more effectively, in order to avoid leaving the road. Not all measures have been detailed, but there have been:
- the development of better detection mechanisms for absurd queries which must not provide an AI summary, and a limitation on the inclusion of satirical and humorous content (to avoid, for example, eating a stone a day);
- updating systems to limit the use of user-generated content in responses that may offer misleading advice;
- Adding trigger restrictions for queries where AI summaries weren’t as useful;
- For topics like news and health, additional improvements have been made (including the decision not to do AI summarization for the most recent news).
In total, Google says it has made about a dozen technical changes to its systems to globally adjust AI Overviews — instead of specifically addressing each erroneous query on a case-by-case basis. According to the American company, this also makes it possible to provide a response to discrepancies which have not yet been reported.
But Liz Reid insists on one point: AI Overviews was not launched rashly.
“ We tested the feature extensively before launch. We put considerable effort into red-teaming, conducted evaluations with samples of typical user queries, and ran tests on some search traffic to evaluate performance “, she said.
However, all of Google’s efforts cannot compare to ” Millions of people are using the feature with many new searches. »And to add further « that on the scale of the web, with billions of requests arriving every day, there are bound to be oddities and errors. »
Follow AI news with Artificielles, Numerama’s dedicated newsletter!
(updated topic to refresh introduction)
Subscribe to Numerama on Google News so you don’t miss any news!