In this issue: AI browsers surf on your behalf while media companies wrestle with the subscriber question. Sonali Verma on why human connection beats chatbots in breaking news. Plus: My battle to teach Claude not to mangle quotes (spoiler: it involved a lot of CAPS LOCK).
What we’re talking about: How do you prove you’re not a bot? Soon, it might not be enough to click “I’m not a robot.” The growing cry to block AI scrapers is pushing us toward eyeball scans (thanks, Sam Altman’s Worldcoin) and ID verification (already live in the UK for certain websites).
How did we get here? Let’s look at the evolution of how we access the web: Phase 1: A user, a browser, a website. Simple. Clean. Human. Phase 2:
Platforms like Google and LinkedIn sit between users and content. Sometimes they host the content directly. Other times, they scrape it and repackage it—think Google’s AI Overviews turning your articles into summaries. Phase 3:
What if the scraping isn’t done by some distant platform in a data center somewhere in the desert, but from your own device, with your logins, your cookies, your digital identity? AI-enabled browsers like Perplexity’s Comet and Dia from The Browser Company do this. They surf the web on your behalf, sometimes hidden behind a chat interface.
The future: For now, AI-enabled browsers still use AI in the cloud to get things done. As AI models become smaller and devices more capable of crunching AI workloads locally, I suspect we’ll see more interactions delegated. It’s gonna be on your phone. As a bonus, this approach would be more privacy focused (think: Apple).
Media companies face the question: Do we allow our real human users to access our websites with the help of AI? If a subscriber asks Siri to get him the latest news, and the phone summarises and reads it to the user, is this something we enable or try to prohibit? Some good answers on LinkedIn:
Esa Mäkinen, Helsingin Sanomat’s Executive editor: “There is difference between subs and ad funded media brands on this issue.”
Daniel Pleus, NOZ/mh:n Medien: “Own the customer relationship (subscriptions!) and make sure users can consume our content wherever they are, not just on our own websites.”
David Caswell on the “by humans, for humans” approach: “It’s the equivalent of hand-made physical goods in a world of industrial automation – high-quality, high-margin, expensive, status-associated and also not relevant to most people/societies or in most situations.”
Welcome, future newsroom leaders: JournalismAI has recruited 20 participants from 17 nations for its “Skills Lab”, empowering non-technical staff in the responsible use of AI.
Next: Why human connection still wins despite all the AI hype. Sonali Verma, the International News Media Association’s GenAI Initiative Lead, explains why you still want your reporters calling sources during breaking news, not a chatbot.
I saw a couple of things that really grabbed my attention: The first, from Gartner, points out that when it comes to AI, most companies are investing in efficiency plays, the low-hanging fruit — but these are simply table stakes to remain competitive and rarely pay off because returns really depend on scaling and adoption. (I see this every day at news brands, and I hear the tales from the trenches about how hard adoption is as well.) What companies should do instead is focus on deploying AI for use cases that create new products that boost revenue or profitability, because that is where they will see real ROI.
The second is an investigation which points out that AI answer engines can provide pretty good summaries of paywalled content — not by breaking the paywall but by cobbling together the contents of the article from what is visible on social media posts and similar sources on the internet that bots are allowed to crawl. That really made me think about the impact on future business models at news media brands. Here we are, thinking that subscriptions will help at least partly pay for the loss of search traffic — and here is an easy way for people to circumvent the paywall to get the gist of an article.
What's one fact about AI that everyone should know?
Despite the rapid pace of change and astonishing new models, GenAI is not going to replace a first-rate reporter or columnist. If you get wind of big breaking news, are you going to ask the machine to call well-placed sources — or are you going to ask your beat reporter, who has spent years building a relationship of trust with those sources? If there is a natural disaster, are you going to get the bot to make calls — or are you going to ask a reporter who knows how to ask questions with sensitivity and empathy and also knows when to stop pushing for answers? Human connection and human judgment are real things. When I want a trustworthy opinion on a topic, I want to see what my favourite columnist has to say about it, not an AI summarisation. You can use AI to get commodity information — but not as a source for the kind of journalism that wins awards and has a real impact on communities.
What's a good hobby to pick up?
Singing! It’s an immersive, meditative experience because it forces you to focus on your breathing while you are reading both music and words — it is incredibly hard for your mind to wander. It releases all kinds of endorphins. It feels good to learn something new, and it’s good for your brain as well. And it gives joy to both the singer and the people around them. If you sing in a group, you are collectively producing a beautiful sound that none of you could produce individually. Pick a genre, find a teacher or a choir that you can sign up for, and do it.
Hands on: How do you tell a chatbot to back off? Apparently, you switch to annoying teacher-mode.
I spent some hours working on a prompt for Claude Sonnet 4 that suggests edits to text. Not a complete rewrite, just suggestions on my terms. The problem? The machine kept mangling quotes. Which it shouldn’t do. Ever.
But getting an AI to actually follow instructions is like trying to train a particularly stubborn person who thinks they know better than you do. This is what I came up with (the actual prompt is in German, translated here for your entertainment):
ABSOLUTE PROHIBITION: NEVER ALTER QUOTATIONS!
CRITICALLY IMPORTANT: Everything between »…« or "…" or „…" is SACRED and UNTOUCHABLE
Even if a quotation contains "horror trade deal" or other emotionally charged terms
Even if a quotation violates all 12 rules
QUOTATIONS ARE NEVER CHANGED - NO EXCEPTIONS!
DEFINITION: What is a quotation?
A quotation begins with » or " or „ and ends with « or " or "
EVERYTHING in between - EVERY SINGLE WORD - remains EXACTLY as it is!
But that wasn’t enough. I had to add this to my step-by-step instructions:
Step 0: ACTIVATE QUOTATION PROTECTION - without output, internal only
Before you do ANYTHING:
Identify ALL quotations (»…« / "…" / „…") without output
These areas are FORBIDDEN ZONES
Apply the editing rules ONLY to text OUTSIDE the quotations
Step 1: Quotation Identification - without output
Mentally mark all text portions between quotation marks as QUOTATIONS
These areas are exempt from ALL rules
Check ONLY the text outside the quotation marks
And finally, at the end of the prompt:
UNCHANGEABLE ELEMENTS
QUOTATIONS ("…", »…«, „…"): Preserve word for word, even with errors!
Subjunctive mood in indirect speech: Maintain
Formatting: Paragraphs and blank lines identical
Source citations: "according to XY", "as reported by" preserve
After testing, this seems to do the trick. But I can’t help thinking: No wonder some journalists are critical or even fed up with AI. These machines don’t understand basic editorial principles. I’m starting to wonder if I should just build some deterministic mechanism to strip out quotes before I even bother praying that an LLM will actually listen to me.
Behind the scenes: Thanks for joining the first reader survey. The initial feedback is already being implemented: links get more explanation, but the newsletter gets a bit shorter overall. On it! If you want to help shape THEFUTURE too, it’s never too late.
