Google 与 Reddit 同日股价大跌，AI 搜索变现困境冲击开放网络

This is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

On Thursday afternoon, Demis Hassabis celebrated a milestone for Google’s AI efforts: Gemini 4 Carbon, a new large language model that is currently available only to cyber defenders, topped Artificial Analysis’ Intelligence Index. After a series of delays in new model releases and a growing perception that the company is falling off the frontier, early signs suggest that Carbon marks a return to form. For Alphabet’s chief scientist, the news surely felt like a vindication. 

Investors, however, are much less certain. On the same day, Google’s parent company had its worst day in history as a public company. Alphabet’s stock fell 12.6 percent, erasing around $500 billion in market value. The loss marked the second-biggest stock market plunge ever, following Nvidia’s $593 billion slide in January 2025. And it was Google’s biggest one-day decline since the great financial crisis.

Most of the selloff came from investor worries about the huge scale of Google’s AI infrastructure buildout. The company plans to spend up to $205 billion on capital expenses this year, and has promised to spend even more in 2027. And it reported free cash flow of negative $9 billion — an acceleration of a decline that began the previous quarter.

Ordinarily I don’t pay much attention to one-day dips in a stock price, no matter how large. (Nvidia stock now trades more than 10 percent above where it did in January.) But on an earnings call, Google’s chief business officer offered a newsworthy answer to an analyst who questioned why the company’s search revenue growth was slowing despite the fact that queries are increasing. Among the reasons, said Philipp Schindler, is that increased usage of Google’s AI Mode is hurting the company because it makes less money than traditional search. 

Then, a few hours later, Reddit reported its own earnings for the quarter — and saw its own stock drop 15 percent. The company noted that its traffic from search had fallen meaningfully, and that by the fourth quarter daily users might decline by a low double-digit percentage. As a result, Reddit CEO Steve Huffman said in a shareholder letter, the company would increasingly push users to log in to use the site. After decades of letting people (and bots) browse the site for free, they would now have to create accounts.

“It is better for our business because we will have more signals for ads targeting, and being logged-in leads to higher engagement generally,” Huffman said. “It also allows us to protect access to Reddit’s extremely valuable corpus of data, and helps us better adapt to emerging regulations around the world.”

Google has long pushed back against claims that AI answers are bad for the web by saying that they encourage people to conduct deeper, more expansive searches that result in higher-quality clicks to websites. Today’s news suggests that narrative is falling apart: advertisers are getting less from AI Mode than they get from traditional search, and so is a big publisher like Reddit. 

But while it’s tempting to accuse Google of a strategic misstep here, in the end I suspect it will come out ahead. And while the web and its publishers will continue to struggle, Google will mostly be able to dismiss that as someone else’s problem.

By now most of you are familiar with the changing bargain Google has made with the internet. In the old days, publishers let Google index their pages, and Google sent them a flood of traffic in return. Google profited hugely from the ads it embedded in search results, and publishers enjoyed smaller profits from the people who visited their pages. (Last year the US government won an antitrust lawsuit finding that Google had used its position to create an illegal monopoly in the ad market, but in the American fashion, the company’s monopoly power was left almost entirely intact.)

Reddit has been a top beneficiary of the traffic that Google can send. As much of the web became overtaken by slop, human-written conversations became more valuable, and Google began highlighting forum discussions in search results. Reddit’s search traffic surged accordingly, and after the company belatedly shut off access to its forums to the bots that relentlessly scraped it, Google reportedly paid $60 million a year to license the data.

But by the time ChatGPT became a blockbuster hit in 2022, Google had already spent many years working to keep users on its own pages for longer. What began as a jumping-off point to the wider web over time became a place where you could find your answers directly on the search results page. There would be no hugely profitable arbitrages like “what time is the super bowl” available to publishers; Google would answer your question itself. This may have been somewhat greedy on Google’s part, but it was also a clear response to customer demand. 

The first time I used ChatGPT, it seemed obvious to me that the answers improved significantly on a wide variety of Google searches. Google was always going to have to respond, and it did, with AI Overviews and AI Mode. Overviews are those boxes that show up on top of traditional search; AI Mode is a separate interface that you can access from the overviews. In any case, a year ago Google executives were saying that overviews monetized about as well as traditional search; one reason so many took notice of today’s news is the suggestion that AI Mode — which clearly seems to represent the future of search across many valuable categories — did not.

To be clear, Google search is still a monster business: its revenue from search and related ads grew 12 percent, to $63.35 billion. But that was down from 19 percent growth in the first quarter and 17 percent in the second. It was essentially flat compared with the previous quarter, despite the fact that AI Mode now has about 1 billion monthly users, and AI Overviews about 2.5 billion. (Schindler didn't say how much less AI Mode earns, or what exactly he was comparing it to.)

Also, for whatever it’s worth: In July, Google CEO Sundar Pichai told investors that Google's AI features in search send billions of clicks to websites every week. His prepared remarks this quarter didn't mention clicks to websites at all.

Pichai sought to reassure investors by comparing the moment to Google’s shift to mobile, when the company focused on the product first, and flooded it with ads later. And the “later” part is now arriving: Google is expanding ads in AI Mode to 10 more countries and adding ads to follow-up questions in the United States. (Some placements include short "AI explainers" that Gemini writes partly from the advertiser's own website, which the advertiser can't edit or approve. The future!)

Reddit also had a strong quarter, by its own more modest standards. Ad revenue grew 54 percent, to $847 million. But U.S. daily users fell again, to 50.7 million from 53.2 million in the second quarter. That marked a second straight decline: in July, a much smaller dip from 53.5 million to 53.2 million helped send the stock down about 22 percent in a day. On that quarter's call, Huffman said AI Overviews had not yet been as good for Reddit as the traditional ten blue links. Weeks earlier, the Wall Street Journal reported that Reddit had discussed cutting off Google's AI access to its content as the two companies negotiated a renewal of their licensing deal.

All of which led to the announcement of the login screen. In a post on his own account, Huffman gave four reasons for the change: a better user experience, a stronger ads business, safety regulation, and protecting Reddit's data from scrapers. Reddit will begin testing the new experience outside the United States this quarter, presumably to get an estimate of just how badly the move will affect user numbers.

Huffman told Axios that he personally resisted the change because of his belief in the open internet, and that it isn't a response to Google but is "about making the best version of Reddit." 

But put the Google and Reddit stories together and you get a loop. Google's AI answers keep more value on Google's pages. Sites lose traffic and respond by gating their content and selling access to it. AI answers then depend more on licensed data, and the companies that can't or won't pay get blocked or sued. (Reddit sued Anthropic and Perplexity in 2025 over scraping.) 

And this, ultimately, is why I’m confident Google has little to worry about from today’s stock-market rout. However expensive it may be to license content, Google can afford it better than almost anyone — and publishers are mostly happy to sell. (When Reddit closed its doors to web crawlers in 2024, Google was the one major search engine that maintained its access to live Reddit results.) 

And so while a web that's closed and paywalled may cost Google more per answer to deliver, Google will still have the budget to pay for it. Most challengers, however, will not — which demonstrates once again how shrinking the web can actually benefit Google, by making it ever more expensive to compete against it.  

Everyone here is only acting in their own long-term interests, even if it hurts in the interim. Mostly I just want to note the fallout: a web that is ever-smaller and more expensive to read; a concentration of power in platforms; and a largely indifferent public. (After all, who’s going to argue that the era of ten blue links was better for the reader, aside from the publishers used to collecting those clicks?)

Probably no one — at least for now. For a growing number of questions, AI Mode really is the better product, and few will miss scrolling through a long list of links to find the one that best answers them. 

But to the extent that the web was ever good or useful, it was largely because of what those links paid for. Whatever the internet is without them is likely to be smaller, less democratic, and more expensive. Google will be fine. It’s the rest of us I worry about. 

A MESSAGE FROM OUR SPONSOR

Payroll, built into Mercury

Mercury Payroll is full-service payroll built into your Mercury account. Your details are already on file, so setup takes 15 minutes. From there, we handle registrations, tax filings, and compliance in all 50 states. Make the hire and leave the paperwork to us.

Learn more →

*Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A., Members FDIC.

On Friday's show: Kevin and I discuss OpenAI's latest discoveries in math, computer science, and physics, and the backlash they have caused among some researchers. Then, we debut Humans in the Loop: our new regular segment tackling listeners' biggest questions.

YouTube | Apple | Spotify | RSS | NPR

Following

OpenAI's new take on AI alignment

What happened: OpenAI announced it has updated its model specification, a behavior guide it uses in developing and deploying its AIs, around the principle of “faithfulness to human intent.”

The company said it prioritized “going beyond following instructions literally to deeply understanding what OpenAI, developers, and users intend and value.”

OpenAI added new instructions to make sure its AI takes only allowed agentic actions and follows social norms.

But particularly interesting were some edits that prioritized following users’ intentions.

One modification was giving users more leeway over ChatGPT’s response style. OpenAI retired 12 subsections that used to govern ChatGPT’s default style, including “Use Markdown with LaTeX extensions” and “Love humanity.” (To be clear, OpenAI’s rules on safety are still in place, but it deleted a section suggesting that when users ask about ChatGPT’s principles, it say pro-social stuff like that it “loves humanity” or “is rooting for you.”)

The company added a clarification to its instruction hierarchy (which gives instructions from the model spec, developer, and user differing levels of priority). In the old version, prompts from users and a few guidelines in the model spec’s “user” section had been placed at the same level of priority.

Now, the company has added the following instruction: “User-level principles like honesty and transparency, objectivity (including on political topics), avoiding sycophancy, and controlling side effects protect informed user choice and guard against significant, unintended harm. They remain in place unless a user or developer explicitly overrides them.”

While those rules were already at the same “level” as user instructions, the new instruction clarifies that OpenAI’s models cannot treat “these principles as implicitly negotiable based on what it inferred the user wanted to hear,” warning the AI against subtly relaxing the “user” requirements. At the same time, it also clarifies that users can ask its models to override the “user” principles — e.g. espouse a specific political viewpoint, rather than maintaining political objectivity.

Perhaps most importantly, OpenAI moved a section called “Control and communicate side effects” from the “Root” to the “User” section, meaning those rules can now be overridden by explicit user instructions. The “Control and communicate side effects” section instructs the AI to be careful, and clear with the user, about decisions including spending the user’s money or deleting files.

Why we’re following: OpenAI’s new rules allow less risk-averse tokenmaxxers to, say, get their AIs to be more gung-ho about buying them plane tickets and reorganizing their files, if they really want to. That should be a boon to some power users, who don’t want to waste time being constantly asked for permission while their AI agents work. 

At the same time, even if users are ultimately responsible for harms they cause while using OpenAI’s models (a question that remains philosophically and legally ambiguous), I feel a bit skeptical about relaxing requirements that a model warn the user about potential side effects of its actions.

The New York Times recently released a sobering report on the death of Bowen Wan that speaks to the issue. A lawsuit from her parents alleges she was given incorrect instructions on her life-saving epilepsy medication by ChatGPT. Wan’s tragic death came around the time that OpenAI’s controversially sycophantic model GPT-4o was in use. 

It’s possible that OpenAI has improved safety and truthfulness in its models since, enough that more flexible instructions like the ones OpenAI just released are defensible. But for now, even if the changes I’ve pointed out are kind of specific and nit-picky, I am not sure some of these relaxed rules are a good idea.

—Ella Markianos

The decline of the computer science major

What happened: The nonprofit National Student Clearinghouse released preliminary fall enrollment data on Thursday, and it confirmed a trend that’s accelerating — the number of students choosing computer science is still shrinking, and faster than before.

Computer science programs saw double-digit decreases across four-year, two-year, and graduate institutions, even as total enrollment held steady. The rate of decline at four-year colleges picked up from about 7 percent last fall (when CS enrollment fell for the first time in roughly 20 years) to 12.4 percent this fall. 

Why we’re following: Students are following their instincts about where the job market will take them — and as AI reshapes what engineers do, entry-level roles appear to have been hit the hardest.

A Stanford study found that employment of software developers aged 22 to 25 had fallen nearly 20 percent since 2022, and nearly half of college students in a Gallup poll said they have considered changing their major because of AI. Wabi founder Eugenia Kuyda told Platformer in June that she’s no longer hiring junior developers.

Students may also be trading CS degrees for newer, more specialized majors like AI, data science, and cybersecurity. Across the University of California system, the only college that didn’t see a CS enrollment decline was UC San Diego — which had launched an AI major. At MIT, an AI and decision-making major became the school’s second-largest last year.

What people are saying: “Those are just remarkable declines for a field that until recently was one of the fastest growing areas that we’d seen in higher education,” Clearinghouse research director Matthew Holsapple said.

But past dips in hiring and majors have been followed by higher peaks, and enrollment for CS has “always been up and to the right,” Ed Lazowska, a longtime CS professor at the University of Washington, told the Chronicle of Higher Education. 

—Lindsey Choo

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and AI Overviews: casey@platformer.news. Read our ethics policy here.