The question this week is whether AI can give humans on this planet a better view of the world, not a worse one. Most writing on AI and information assumes the next chapter is darker than the last. There is another reading worth taking seriously. Talk of the Week steps back from the fight and asks: could AI actually make information better?

This is what the New York Times has spent on its lawsuits against OpenAI, Microsoft and Perplexity since December 2023. Chairman A.G. Sulzberger disclosed the figure on 1 June in the opening keynote of the WAN-IFRA World News Media Congress in Marseille, set against the USD 2 billion the NYT spent producing nearly half a million pieces of journalism in 2025. He called AI training on news content "brazen theft of intellectual property." (Press Gazette · 1 June 2026)

Why care? The legal fights look familiar. When YouTube launched in 2005, the music and television industries spent years suing it before settling into a working relationship; similar arcs ran through Napster, Google Books and Spotify. What we are watching now is the same pattern, one industry on: publishers and AI platforms have never worked together at this scale before, and the terms of that relationship are being argued out by lawyers because no one has written them yet. It will take time. The outcome that makes journalism economically sustainable is still being negotiated.

3 June 2026 · Press Gazette

What's new: The UK Competition and Markets Authority issued the first binding order anywhere requiring Google to let publishers block their content from feeding AI Overviews — without losing position in conventional search. Publishers can also opt out of having content used to fine-tune Google's models, and Google must attribute sources with clear links. The order comes under the Digital Markets, Competition and Consumers Act 2024.

Why care: Until last Wednesday, publishers had one choice — let Google use your content for AI summaries, or disappear from search. The opt-out breaks that. Google says it will roll the controls out globally, which would make this a UK regulator effectively setting terms for everyone. CMA chief Sarah Cardell called it "a world-first requirement."

Reality check: The Professional Publishers Association made the obvious point — an opt-out preserves your content, it does not recover the traffic already gone.

June 2 2026 · Microsoft AI

What's new: Microsoft AI shipped MAI-Transcribe-1.5, a speech-to-text model covering 43 languages, claimed to run roughly five times faster than competitors. Released alongside MAI-Voice-2 and MAI-Image-2.5. All trained in-house, not by OpenAI.

Why care: Last week we covered Microsoft making Anthropic's Claude the default in Excel and Powerpoint. This week the in-house stack adds transcription. The pattern is clearer now: Microsoft is brokering, not picking a single winner — OpenAI for some surfaces, Anthropic for others, in-house where the application is narrow and language coverage matters.

Reality check: "Five times faster" is vendor-stated. Forty-three languages is not 100, and "covered" and "production-quality" mean different things. The number worth watching is word error rate per language — not yet published.

Late May 2026 · Press Gazette AI mistakes tracker

What's new: The Sydney Morning Herald removed an opinion piece by Prof. Cath Ellis of Western Sydney University after it emerged AI had been used to write it. Ellis pushed back: she had written down her own thoughts and asked an AI assistant to structure them. "It was written with AI. There's a really big difference there."

Why care: Her distinction is the one most newsrooms have not landed on yet. There is a continuum between written by, structured with, edited with, and verified with AI. Treating all four as the same will keep getting pieces removed.

Reality check: No shared standard means no shared answer. The Babylon disclosure block at the bottom is one attempt - my attempt. There are dozens of others, none compatible.

For most of the twentieth century, a serious person could read three newspapers in the morning and have a defensible view of the world. The model was incomplete — only what editors picked made it onto the page — but it was a shared starting point.

That approach has been breaking for twenty years. The question is whether something better might replace it, or only something worse.

The positive view: The strongest version of the better case goes like this. AI's distinctive contribution to information is not access — millions can already reach a public-health study or a budget — but comprehension. Almost no one can engage with either of those on their own. An AI intermediary that translates, contextualises, and adapts content to what a specific person needs to know does something newspapers never did at scale: it meets the reader where the reader actually is. Shuwei Fang at Harvard's Shorenstein Center calls this the comprehension gap, and argues that closing it could expand the market for good journalism, not shrink it (Reuters Institute, March 2026).

Done well, the new system would offer horizontal — an honest overview across more domains than any one newsroom can cover — and vertical — depth on the things that matter to the reader, assembled from multiple sources. A personal information diet rather than a publication. The technical idea underneath this — news decomposed into smaller semantic units that recombine for different readers — was laid out in David Caswell's 2019 paper in Digital Journalism; he and Fang then mapped the implications in their AI in Journalism Futures report.

On the truth question, fact-checkers are blunter than the discourse usually allows. Chris Morris of Full Fact, at a Reuters Institute panel: "There are things this technology can do for us which allow us to address the problem at a huge scale, in ways that even a newsroom of a hundred people couldn't do." Brazilian fact-checker Aos Fatos is already moving from after-the-fact verification toward live, real-time context delivery — its forthcoming tool is called Busca Fatos.

Now the counter-case, which gets less airtime than it should. Dodds, Zamith and Lewis argue in Journalism this year that an AI-mediated environment risks not better-informed readers but burned-out, detached ones — the cognitive load gets pushed back onto the reader at every step. A separate March 2026 study by Petrovic Garcia and Perry found people who consume more AI-generated news believe the stories are better, while reporting that what they see is less diverse. Easier to consume, narrower at the same time.

A system that could give most people genuinely better information is buildable. So are the failure modes. Which one we get depends less on the technology than on who decides what counts as relevant — and whether those people are accountable to readers or advertisers.

The 36-publisher coalition. SPUR formalised its WAN-IFRA partnership in Marseille on 3 June and now bundles 36 publishers; the European Publishers Council joined too. Worth tracking as the body most likely to set the terms of the next licensing round, rather than the platforms doing it one deal at a time. (WAN-IFRA, 3 June 2026)

Journalists, training their replacements. The Reuters Institute spoke to four freelance journalists in Canada, India, Germany and the US now working as AI trainers — judging chatbot output for clarity, accuracy and tone. Two found the AI work paid more than journalism; two found freelance writing still paid better. From Khaleda Khan, ex-xAI: "While I was working there, I was thinking — am I helping make something that will replace me?" (Reuters Institute, 5 June 2026)

"Chat is dead." OpenAI plans to relaunch ChatGPT as a "super app" in the coming weeks — a gateway leading free users into paid products like its coding tool Codex, per the Financial Times. The line comes from a senior OpenAI employee. The strategic context: getting closer to profitability before an expected IPO, and catching Anthropic on business customers. (TechCrunch, 7 June 2026)

Free or subscription? ChatGPT alone has around 50 million paying subscribers worldwide against close to a billion weekly active users — about 5% convert to paid. Adding Claude, Gemini, Copilot and the rest does not change the picture much: the vast majority of AI users do not pay directly. Whether AI's future is free (ad-supported), embedded in productivity suites you already pay for, or standalone subscription remains the open question of the next eighteen months. (TechCrunch, 27 February 2026 · Nieman Lab, 4 June 2026)

EAMT 2026 · 15–18 June 2026 · Tilburg, Netherlands · The 26th European Association for Machine Translation conference. Workshops 15 June, main conference 16–18.

GIJN webinar: "From the Panama Papers to the Epstein Files" · 18 June 2026 · 15:00 CET · ICIJ's Pierre Romera Zhang, Panama Papers' Bastian Obermayer and others on data investigation when forgery is cheap.

DW Global Media Forum · 23–24 June 2026 · Bonn, World Conference Center · DW's annual international media conference, ~2,000 participants from 120+ countries; theme this year is Journalism out loud. plain X (DW Innovation / Priberam) will have a stand on site for the first time — disclosure as in About & Disclosure below.

RSI — recursive self-improvement. The new term to know. The Economist used it this week for the closed loop where an AI model builds its own successor, which builds the next one, with no human in the loop. Anthropic co-founder Jack Clark puts the odds of that loop closing by end of 2028 at 60%.

Worth knowing for two reasons. The gap between "AI helps researchers" and "AI replaces researchers" is narrower than the public debate assumes — Andrej Karpathy recently had an AI agent improve his Nanochat training time by 18% in two days, unsupervised. And on 5 June, Anthropic itself called for the option to "slow or temporarily pause frontier AI development" — striking from the market leader, the same week it is preparing one of the largest IPOs in history. Both can be true.

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany. I co-founded Datawrapper, a charting tool used in many newsrooms.

Four projects you will hear about in this newsletter:

I cover all four with the same critical lens applied to competitors.

AI use: I use Claude (Anthropic) for research and to edit this newsletter, based on refined and specific prompts. My goal is to understand where the AI performs and where it fails. I learn something every week. Responsibility for stated facts, names, and links is entirely mine.

babylon-newsletter.com · Every Tuesday · 7,000 languages. AI works for 20.