Carl T. Bergstrom
I wanted to consolidate a few thoughts on google, misinformation, large language models, enshittification, and the fate of the web as we know it.
It started when Carl Zimmer shared this remarkable example of Google being fooled by machine-generated bullshit online.
Carl T. Bergstrom
But the clock is ticking because google is so big and so powerful that its own internal enshittification process has massive spillover effects on the entire WWW.
For the first time in a quarter-century I can imagine the death of the web as we know it, into a wasteland of GPT-6 content—the output of a model trained on mostly on the output of GPT-5—trawled only by chatbots pretending to be human readers to scrape up what ever remnants of long-term contract ad dollars are left to be had.
/fin
Carl T. Bergstrom
Meanwhile the cost of creating content with the superficial appearance of well-considered text has fallen off a cliff, thanks to ChatGPT.
So filling the web with quality material is less rewarding than ever, and filling it with shit is orders of magnitude cheaper than ever.
Maybe someone unseats google before the web becomes a wasteland of ChatGPT-authored, search engine-optimized babble.
Carl T. Bergstrom
If google wasn't big enough to move the market, this would just be a story about google's downfall and vulnerability to competition.
But if google succeeds in keeping people on-site instead of visiting links, the incentives for people to create high-quality content fall through the floor. Whether you are a commercial provider funded by ads or an academic trying to disseminate something more than soundbite "featured snippets", google's efforts undercut your motives to create quality content.
Carl T. Bergstrom
And in doing so, Google becomes a *bullshitter* in the exact sense we wrote about in our book Calling Bullshit.
They have designed a system that generates "language, statistical figures, and data graphics intended to persuade by impressing...a reader or listener with a blatant disregard for truth."
Carl T. Bergstrom
So just as Google has strong incentives to keep users onsite using shitty large language models that hallucinate, it has strong incentives to keep users onsite by making even relatively unreliable guesses about what constitutes a definitive answer to a search query.
Carl T. Bergstrom
(A bigger motivation is the same thing that keeps Google on top of the search game. They get all of the queries and thus have all the training data. Bing can't ever catch up. If Microsoft's GPT4+Bing went unanswered, Microsoft would get all of the training data for improving LLM-based search.)
Carl T. Bergstrom
This is where enshittification comes in. The search engine business is a curious internet business. The better you are at it, the less time people spend on your website—and the less ad revenue, among other things, you receive.
Google knows thisl. If they can give you the definitive result that you are looking for without you leaving the page, they win.
It also, IMO, explains part of the motivation for Google in deploying an utterly second-rate large language model—Bard—in production code.
Carl T. Bergstrom
Designing algorithms to return definitive results and then getting duped by conspiracy theorists is embarrassing; doing the same and getting pwned by the hapless stupidity of large language models is malpractice.
IMO this is not simply a matter of Google's algorithms being no good. What's going on here is that Google is playing a losing game by trying to extract true information from the web algorithmically. Why would Google choose to play a game it could only lose?
Carl T. Bergstrom
One issue is Google's degree of culpability. Google is a search engine, not an encyclopedia. Ask a stupid question? Get a stupid answer. And seriously who is going to ask are there countries in Africa that start with K?
My view is that Google's culpability was radically transformed when it went from simply returning web pages as search results to presenting results from web pages as definitive answers, via its "featured snippets", "knowledge panels", and "people also ask" feature.
Ruth Mottram
@olav @ct_bergstrom kagi search is miles better than either bing or Google.
Plus no ads or tracking. But it's a paid subscription.
olav
@ct_bergstrom the clock is especially ticking because the only standout (at least for search) is Bing, which of course went Chat-GPT first. Amazon is more evil than the two combined and their search doesn't even work on their own website, despite 25 years
That said, if someone created an app suite basic pre-iPhone "smartphone" ~ integrated email, SMS, calendar, tasks, address book/phone , maybe some perks I'd definitely give a long look at it. Unfortunately this is what Goog does best
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.