April 3, 2023

Local Memo: How Google Fights Fake Content in Local

Damian Rollison

Local Memo: How Google Fights Fake Content in Local

In this week’s update, learn about Google’s fight against fake content in local; AI-generated Google reviews; Italy’s ban of ChatGPT; Twitter’s now public ranking algorithm; Google’s restructuring to emphasize Bard development; and generative AI reaching the public in Gmail and Docs.

How Google Fights Fake Content in Local

In a new report, Google outlines the techniques the company utilized in 2022 to battle fake content posted to Google Maps and Google Business Profiles. Some stats and facts that stand out from the report:

Google Maps has one billion monthly users (presumably a global figure).
Google confirms that, in 2022, it launched a significant update to the machine learning algorithm that fights spam content. The announcement points specifically to improved capability in identifying suspicious activity across many Business Profiles at once.
Google uses machine learning to combat a trend whereby scammers post photos to profiles overlaid with the scammer’s phone number, in an attempt to lure customers away from the legitimate business.
Google removed 20% more fake reviews in 2022 than in the previous year, for a total of 115 million reviews, most of them removed before they reached publication.
Google also removed 200 million photos and 7 million videos for being low quality or for content violations.
The company blocked 20 million fake profiles in 2022, 8 million more than the previous year.

Users Spot Fake AI-Generated Google Reviews

But alas, Google’s battle against fake content cannot claim 100% success. Case in point: users are now spotting fake reviews in the field that were clearly written by ChatGPT or another AI chatbot. Strangely, the reviews make no attempt to hide this fact, nor do they sound much like an authentic review. Example: “I'm sorry to hear that you had a bad experience with Right Way Garage Doors. As an AI language model, I can't provide any further help without more information about the details of your experience. If you could provide me with more information, I would be happy to help you with any concerns or issues you have.” As you can see, this example, similar to others, looks more like an attempted reply from the business than like the feedback of a consumer. It’s unclear what may have motivated the spammers to post reviews like these, and also unknown whether less transparently fake AI-generated reviews may be appearing elsewhere.

Italy Bans ChatGPT

Italy has announced a ban that blocks the use of ChatGPT, citing privacy concerns. The move follows shortly after an open letter signed by U.S. tech leaders, including Elon Musk, called for a six-month suspension of advanced AI development due to the “profound risks” it poses “to society and humanity,” and asking for governments to step in and issue moratoria if necessary. Italian authorities are reportedly investigating whether ChatGPT abides by GDPR privacy regulations. Google’s rival AI tool, Bard, remains available in Italy but is restricted to users over 18. The invocation of GDPR suggests that other European countries could follow suit in outlawing ChatGPT, which has already been banned in North Korea, China, Iran, and Russia.

Twitter Shares Its Feed Ranking Algorithm

Fulfilling earlier promises from Elon Musk, Twitter has published its feed ranking algorithm to GitHub, making many — though not all — details of its platform’s ranking methodology available to the public. The code specifically ranks content in the For You version of a user’s feed. The version that was shared publicly was supposed to reveal only the factors considered in ranking decisions but not the weights applied to each factor, which were to remain private according to statements from Twitter. But developers quickly discovered that those weights were, in fact, present in the published code. They reveal that likes and retweets outweigh all other factors by a significant margin, followed distantly by Twitter Blue status and other factors. There are also factors that make a post less likely to appear in feeds, such as tweets containing no text or only a URL. In a blog post, Twitter also revealed that it dynamically defines and tracks large discussion groups called SimClusters, now a key factor in determining which out-of-network content a user sees. The For You feed now contains about half in-network and half out-of-network content.

Twitter algo 101
Boosts - Likes 30x - Retweets 20x - Twitter Blue 2-4x - Trusted circle 3x - Images/videos 2x - Replies 1x Negatives - URL only - No text - Mute - Block - Unfollow - Report pic.twitter.com/mrCuGXB2gJ — Peter Yang (@petergyang) April 1, 2023

Google Restructuring to Emphasize Bard Development

Google has reorganized the Assistant division of the company to focus on Bard development, suggesting that Bard may soon be integrated into tools that currently utilize Assistant as a voice interface, such as Nest, the Pixel phone, smartwatches, and Android Auto. The news comes amidst reports that Bard will soon incorporate Google’s Pathways Language Model (PaLM) in addition to its foundational model, Language Model for Dialogue Applications or LaMDA. PaLM will reportedly bring greater competence in math and logic. Coding capabilities are also reportedly coming soon to Bard.

Google Rolling Out AI in Gmail and Docs

In another sign of the swiftness of AI developments, Google has already begun public testing of the very recently announced generative AI integrations for Gmail and Google Docs. A “small group” of trusted testers includes consumers as well as enterprise and professional users, all adults in the U.S. I’ve recently shared a preview of the Gmail AI interface, which will include prompts to generate and revise text. For Docs, there will be a “Help me write” button that allows users to enter a prompt for composition. Users can judge the result with a thumbs up or thumbs down, view an alternative passage, refine, and insert the result into their draft where it can be further edited. Google will collect feedback from test users prior to any further public release.