The Great Middleman Revival

Newsletter sent 5.8.2025 by oler

In this issue: While humans disappear into AI search engines, bots offer micropayments. Cloudflare wants to be your middleman (for a cut). Newsroom Robots’ Nikita Roy on reclaiming audience relationships. Plus: I tested ChatGPT Agent and Perplexity Comet on real tasks—and learned something about AI bias.

What we’re talking about: Publishers have spent years saying no to single-article purchases. Why sell one piece for cents when you could build a lasting relationship instead? (Except for that €5 garlic oil test report—that’s different.)

But lately, more and more human users get trapped in AI search engines, never to be seen again on publishers’ websites. Meanwhile, AI bots roam the open web, outnumbering humans. Instead of turning away these bots, could publishers make money from them?

Several companies are trying to build exactly that. One of the loudest and most influential: Cloudflare, the content delivery network sitting between you and 20 percent of the web. The middleman monitors declining traffic and wants to help: When a bot hits an article, Cloudflare would negotiate payment on the publisher’s behalf—taking a 20-30 percent cut, obviously.

Micropayments in combination with the beloved fee-skimming intermediary? Sign me up!

Meanwhile, the New York Times reportedly earns $20-25 million annually from Amazon for content access, according to the Wall Street Journal. There’s money to be made in journalism—the question is how much, and for whom.

What else I’ve been reading:

AI & Journalism Links

What I Learned from Making a ‘Liquid Content’ Machine (Clare Spencer, Generative AI in the Newsroom)

“Transparency” and “disclosure” are buzzwords in the AI and journalism space – this report looks at the state of global research. (Center for News, Technology & Innovation)

From an overview how newsrooms tackle bias in large language models: Humans and AI bots have biases. But the machine won’t be offended when you call it out. (Ramaa Sharma, Reuters Institute)

$200+ monthly fees: Welcome to the two-tier AI landscape favoring the well-off. (Reece Rogers, Wired)

“Despite appearances, an LLM does not actually output text”: The Guardian’s Joseph Lochlann Smith with a myth-busting deep dive. (Medium)

Image generation, without the “AI look”: Flux.1 Krea is an open weights model with opinionated aesthetics.

“There’s no getting around the decline in traffic”: Another apocalyptic roundup on what’s happening with search. (Klaudia Jaźwińska, Columbia Journalism Review)

Hands on: I tried ChatGPT Agent and Perplexity Comet to actually get things done—and I’m not sure if I’m witnessing the future of productivity or just an elaborate way to make simple tasks more complicated.

The premise sounds great: tell an AI what you want, and it’ll just… do it. Biggest difference between the two: ChatGPT Agent is a sandboxed browser inside your browser. Perplexity Comet is the actual browser, carrying over all your logins from Chrome.

1. The Great Train Booking Experiment: I asked both agents to book me the cheapest train from Hamburg to Copenhagen. Simple enough, right? I mean, I could knock this out in under a minute on the Deutsche Bahn website.

ChatGPT Agent found the better deal at €63.26, but watching it navigate the Bahn website was like watching someone perform surgery with oven mitts. Ten excruciating minutes of clicking cookie banners and scrolling through options. Perplexity Comet found the same connection in 3:40 minutes—but missed on savings, resulting in a €70.94 fare.

2. Lost in Translation: I asked for the biggest news in Germany in Axios-style bullet points, and both agents… used only English sources? In 2025? When AI can translate faster than I can say “Guten Tag”?

Perplexity Comet took 20 seconds and served up some articles, among them week-old news about a train derailment. ChatGPT Agent took nine minutes (nine!) but delivered actually witty, recent summaries. It appeared to read different sources, but used only two in the end.

3. The Guest List Gender Test: I asked both agents to help me find potential newsletter guests by checking past issues and my LinkedIn connections. I alternate between male and female guests, a rather obvious pattern.

Perplexity Comet got close: In less than two minutes, it crawled through my newsletter archive, scanned my LinkedIn connections, checked out 14 profile pages, and delivered a table with 13 solid suggestions. Five women out of 13.

ChatGPT Agent spent minutes of opening browsers, methodically reading through all newsletter issues, getting confused by non-existent archive pages, making me log into LinkedIn, and painstakingly clicking through connection profiles and their experience-sections like it was conducting a digital census. The result after eight minutes? Nine solid entries, but only one woman.

Only one.

Now, I’m not saying ChatGPT Agent is secretly plotting against gender diversity in newsletter interviews. But ChatGPT Agent took four times longer, yet somehow managed to recreate exactly the kind of unconscious bias that many of us are actively trying to avoid. Makes you wonder: When we hand over these kinds of decisions to AI, whose patterns are we really perpetuating?

Help me help you: There are two more examples on my blog, buying stuff and making playlists. Maybe I need better use cases. Maybe I’m thinking too small. While I don’t have any invites to share (yet), I’m happy to work on your behalf: What tasks would you test these agents with? Send me your most ambitious prompt ideas!

Speaking of newsletter guests: This week, it’s one of the most sought-after keynote speakers on AI and journalism, host of the Newsroom Robot podcast, journalist turned media advisor Nikita Roy. She has an important message for everyone who thinks that all of this is just another pivot to video.

Three Questions with Nikita Roy

My summer of AI: The New ABC of AI · Who’s who in the German-speaking LinkedIn universe · 10 Facts Every Journalist Should Know · Essential AI and Journalism Resources

A very good post from Louis on Threads: “using llms to make decisions, call it chat roulette.”

This is THEFUTURE.

Subscribe to THEFUTURE

Get this newsletter in your inbox. For free.

The previous issue is Magic Tokens and Modern Luddites, the next issue is Jumping Bunnies and Shrinking Context.