The machines have arrived faster than any technology before them. US adoption of generative AI outpaced both the personal computer and the internet at the same age, and yet inside most organisations the picture is split: a few people run everything through AI, most try it occasionally, and a solid block wants nothing to do with it.

That split has a name, and it is 64 years old. This issue is about what the innovation-adoption curve tells you to do with the people at the back of it: be friendly, take the concerns seriously, and stop asking them for permission.

That is how many pages some AI crawlers take from a website for every single visitor they send back, according to Cloudflare's new attribution dashboard, launched 1 July. The ratios run from about 118 to 1 at the friendly end up to nearly 50,000 to 1. Cloudflare sits in front of a large share of the web, so it can see both directions of the exchange: what the bots take, and what they give back.

Why care? The number is Cloudflare's own, and the examples are illustrative rather than from a named customer, so treat it as a vendor-measured ratio, not an audited one. But the direction matches every independent measurement of the same exchange.

Source: Cloudflare blog, 1 July 2026. Ratio analysis: PPC Land, 5 July 2026.

12 July 2026 · The Media Stack / LinkedIn

What's new: Louis Dreyfus, CEO of Le Monde, says readers arriving from ChatGPT convert to paying subscribers at 17 times the rate of readers arriving from Facebook, in an interview with John Rahim's The Media Stack.

Why care? This is the first hard conversion figure from inside a publisher AI licensing deal. The deal structure matters as much as the number: Le Monde ring-fenced its agency wire, photography and video, took a guaranteed payment instead of a revenue share, and passes 25 percent of the income to its journalists.

Current status: One publisher, numbers are publisher-stated, no independent verification. Dreyfus was the first French publisher to sign with OpenAI and remains the only one.

9 July 2026 · OpenAI / Florent Daudens

What's new: OpenAI launched GPT-Live, a voice model that listens and speaks at the same time, handles interruptions, and switches languages mid-conversation.

Why care? The language switching is the Babylon angle: live, unannounced movement between languages inside one conversation has been a demo promise for years. And Florent Daudens, who had early access, reports he turned voice on by default for the first time. His frame: voice replaces the first gesture. You do not open an article, you ask.

Current status: Early-access impressions, no published benchmarks, no per-language quality data. Daudens names the real product problem himself: an answer synthesised from five sources sounds smooth, and you cannot have footnotes with voice.

11 July 2026 · Rasmus Kleis Nielsen / LinkedIn

What's new: Rasmus Kleis Nielsen, professor at the University of Copenhagen and former director of the Reuters Institute, sketched four scenarios for how answer-engine optimisation plays out for news publishers, splitting the field on two axes: few or many licensing deals, and low or high value to those who hold them.

Why care? His blunt read is that the current incentive structure is "fine for PR, uncertain for news publishers." Everyone doing strategic communication has a clear reason to feed the answer engines; for newsrooms the payoff depends on how AI firms, publishers and regulators move next, and three of his four scenarios leave most publishers worse off than search once did.

Current status: A framework floated for discussion, not a finding, and Nielsen asks openly for counter-examples.

Why you should not try to convince the hold-outs.

The pattern is older than the new technology: In 1943, two sociologists studied how Iowa farmers took up hybrid seed corn and drew a curve that has described every major technology since. Everett M. Rogers formalized the observation in 1962 and called it the “diffusion of innovation”. According to the model different groups adopt a new at different times: The first group are the innovators, at about 2.5 percent. Early adopters, 13.5 per cent. Followed by the early majority and a late majority of 34 percent each. And at the back, 16 percent who adopt last or never. Rogers built the model from more than 5,000 diffusion studies. The percentages are idealised, but the shape keeps showing up. Note that the curve says nothing about how fast an innovation is taken up, nor does it imply that every innovation actually moves through all groups. Only the biggest new offerings get to 100 per cent, or almost 100.

The curve, as drawn since Rogers: the bell is who adopts when, the S-curve is the market share they add up to. Source: Wikipedia, Diffusion of innovations.

Here is what the shape means for an AI project in 2026: roughly one in six of the people your project touches will not be convinced by anything you say. They are called “hold-outs” or “laggards”, but both words mean the same group. Their objections do not carry information you can build with. Their approval is not a milestone on any path your project can actually take. And yet a remarkable amount of organisational AI effort goes into exactly this group: the extra town hall, the softened rollout, the feature nobody asked for except the person who will not use the product anyway.

The current evidence says focus wins. MIT's Project NANDA reviewed enterprise AI deployments after an estimated 30 to 40 billion dollars of spending and found that only about 5 percent of integrated pilots produce measurable value. The finding is widely cited and its methodology is contested, so hold it loosely. These are the very early days of AI at scale and like other innovations (cars, railways, electricity, the computer) it takes time to become impactful. But its description of the winners matches what McKinsey found from the other direction: 88 percent of organisations now use AI somewhere, only about 6 percent get real financial impact, and the difference is not the model. The winners pick one pain point, one willing team, and execute. They do not run a referendum first.

So the working method looks like this:

Two notes, because this argument is usually made sloppily. First, the theory's own authors disagree about the details. Geoffrey Moore built a business classic on the "chasm" between early adopters and the early majority; Rogers himself said past research shows no support for that chasm existing. The curve is a map, not a law. Second, Rogers named the field's biggest flaw himself: pro-innovation bias, the assumption that everyone should adopt everything, quickly. He thought that assumption was wrong.

And it is true, big new changes come with benefits and they create new problems. This is where the hold-outs are right. Early critics of the automobile said it would poison the air. The convenience won out and the car scaled to everyone anyway. Still, the critics were right: MIT researchers attribute roughly 53,000 early deaths a year in the US to road-transport exhaust alone. The social-media generation is running the same argument about the negative effects of social media on children now.

The serious concerns should of course not be entirely ignored. But they go into the risk register and get worked on, eventually. The key suggestion is to first focus on the benefits. Just note that there is a notable rise of “shadow AI” - where employees use AI, whether their employer likes it or not. Such usage patterns are a sign that a technology clearly found a market, whether the sceptics like it or not. For sure AI will have both positive and negative effects. But the diffusion curve has been drawn ten thousand times since the Iowa corn fields, and it has never once been redrawn by the last 16 percent.

Be friendly to the hold-outs. Do not let them write the project plan.

Three publishers, three AI strategies, one rule. Ulrike Langer compares how TIME, The Economist and Dow Jones decided what to expose to AI agents: TIME opens everything and sells the visibility data, The Economist exposes only what is already free, Dow Jones hands over the journalism and guards the proprietary data underneath. The rule: the strategy follows from what each company actually sells.

The subscription number under the subscription number. Thomas Baekdal splits the Reuters Institute payment data into one-off payments versus real ongoing subscriptions. France's headline 12 percent who "pay for news" collapses to 6.2 percent with an actual subscription, and among 25 to 34 year olds in the US the rate has halved since 2020.

The pressure is not AI. It is your clients' AI. Bruno Herrmann, who spent years buying language services inside HP, Nielsen and IQVIA, argues the real force reshaping the language industry is clients transforming their own content operations. Suppliers who defend the old model become cost centres; the ones who move upstream into the client's workflow become partners.

LocWorld56 · 19 to 21 October 2026 · Vancouver · locworld.com · The localisation industry's largest event, with the Multilingual AI track.

Languages & The Media · 4 to 6 November 2026 · London · languages-media.com · "Moving Images That Move Audiences: Localising with Intent." Disclosure: I have a speaker slot.

Test a multi-model platform, even if you only ever need one answer at a time. Tools like Mammouth.ai (issue #8) put the major models behind one subscription, which turns every question into a small comparison test: you see which engine currently handles your language, your domain, your kind of task best. It is the personal version of the smart model routing covered in issue #17, and the cheapest way to stop believing any single vendor's story about itself.

I am Mirko Lorenz. I work as an innovation manager on language technology projects at Deutsche Welle in Germany. I co-founded Datawrapper, a data visualization tool.

Three projects you will hear about in this newsletter:

AI use: I use Claude (Anthropic) to research and edit this newsletter, with prompts I have refined many times. The goal is to get help, and to find where AI makes mistakes. It makes them; a diligent author can catch them. Responsibility for stated facts, names, and links is mine. I also keep an open Google Doc tracking problems, mistakes by AI or by me. This is to learn more, week by week.

babylon-newsletter.com · 7,000 languages in the world, AI works for 20.