A group of United States lawmakers, led by Representative Adam Schiff, have given Alphabet/Google CEO Sundar Pichai an August 1 deadline to acquiesce to a demand for information related to the company’s “AI Overview” feature for Search.
AI Overview is a summary window that presents at the top of Google Search pages when a user makes a query. It usually contains a short text answer, an image if relevant, and supporting links.
According to Schiff’s office, citing numerous reports, AI Overview has a tendency to output incorrect and/or potentially dangerous information.
The complaint, published by Schiff’s office and co-signed by four other members of congress, requests answers to eleven questions about the feature related to the company’s efforts to ensure greater accuracy and safety.
AI Overview
Some of the issues with AI Overview stems from its apparent inability to determine whether a given source is satire or not. “Google’s AI Overviews feature told users to eat at least one rock a day for vitamins and minerals,” wrote Schiff and the other members of congress, adding that the AI was citing satire news site “The Onion.”
In other cases, such as when users began reporting that AI overview claimed Barack Obama was the first Muslim US president, the AI system directly cited debunked conspiracy theories.
Schiff’s letter acknowledged Google’s ongoing efforts to combat misinformation, but stated that more needs to be done:
“We appreciate Google’s efforts to fix these errors by improving the feature through continued testing, taking down responses that violate the company’s content policies, and labeling AI-generated responses as experimental. However, more must be done to adequately ensure that Google can remain a trustworthy and reliable source of information for Americans.”
AI reasoning
As of now, there’s no scientific consensus that these kinds of problems can actually be solved. The gist of the problem is that AI doesn’t “think” or “reason,” despite both terms being commonly used in marketing for AI products.
Generative AI systems imitate human speech but lack the ability to understand context or apply earned experience to a situation.
These problems are long outstanding in the field and were noticed in early large language models such as OpenAI’s GPT-2 and Meta’s now defunct “Galactica” system which was taken down after only three days over the backlash its launch created.
US demands
In order to further understand Google’s efforts, Schiff and his fellow lawmakers have demanded that Pichai and Google respond by August 1 with answers to a list of 11 questions. These include queries about the company’s ranking procedures, automated fact-checking, accuracy, sourcing, transparency, and how it’s working with authorities to ensure safety.