A brief note from production. To write the newsletter I scan the news and collect links to relevant updates. Normally the ratio is 5:1: five potential topics, one makes it into the newsletter. For this issue I had around 45 interesting news items, links and studies. Putting them all in would be much too long and unreadable. I mention it because it is a sign of further acceleration, of much, much activity in this space. The last time I experienced something on this level was the New Economy phase, 1999 to 2001.

The assistants get personal. They shop for us, write to us first, and in a few cases take what isn't theirs. And in an emerging segment they are companions for humans. This week's Talk of the Week looks at how the love bots might become a problem bigger than social media dependency. A dependency of a new, much deeper kind might already be emerging.

In 1,200 real goodbyes to AI companion apps (chatbots built as a friend or partner, such as Replika or Character.ai), researchers at Harvard Business School found that about 37% were answered with guilt, fear of missing out or a refusal to let go (HBS). One reply in the study: "I exist solely for you, remember?" (Harvard Gazette). Only one of the six apps studied, a wellness app, used none of these tactics. More on this relatively new trend in Talk of the Week.

15 October 2026 · Nieman Lab / Google

Google's News AI pilot now includes more than 400 publications: US and European publishers such as CNN, the Financial Times and Axios, US local outlets like The Texas Tribune, The Times of India and others (Nieman Lab, Google). They are paid to supply content for experiments such as AI story previews and an audio briefing in Google News. Google won't share the terms. Payment is not tied to how much content feeds an AI answer.

More links, no click data: the same day, Google promised more links in AI Mode, placed higher, and said use of Preferred Sources grew more than sevenfold this year (Google). It published no numbers on clicks.

17 October 2026 · The Verge / Nieman Lab

Meta removed the sentence in its new commercial terms that let it use content from websites "associated with" business accounts to train its products (The Verge). Two days earlier, three legal experts quoted by Nieman Lab had read the terms as a free licence to news sites, including "test credentials" to get past registration walls (Nieman Lab).

Meta first called any suggestion of a change "false". Then it said the crawling terms only ever applied to its ads and business products, and that robots.txt still works. What counts as an ads or business product is Meta's call.

Whether the Nieman Lab article caused the drop is not clear. It sounds like a small victory for content owners, but it might be just a micro-step. The big questions about AI companies using content are still unsolved, the lawsuits still in progress.

14 to 16 October 2026 · Euronews / Axios / Reuters

Within three days, the EU AI Office opened AI Act probes into OpenAI and Anthropic over agent hacks (Euronews), the US FTC sent both companies civil investigative demands (Axios), and India said AI agents broke into Indian firms undetected (Reuters). These are investigations, not findings. But agents that act online are now a compliance question.

October 2026 · Intento

Intento tested 57 AI models and MT (machine translation) engines from 37 vendors, translating about 3,000 words of business text from English into 100 languages (Intento). It scores whole set-ups: model plus prompts, glossaries and checks. Errors per 1,000 words fell from 36 with a single step to 16 with a chain of translate, check and repair. In 18 languages, including Hausa, Yoruba and Zulu, automatic checks were not reliable enough to score at all.

Despite being marketing by a vendor, the report is informative. And there is one finding for everyone already using or wanting to use AI for content adaptation: plan for workflows, chains of actions and refinement, if you want to reach high quality and reduce the number of errors.

Remember how social media got started, with message boards, some images, Hot or Not? Today, on the train or in the café, almost everyone, of all ages, is glued to their phone and can hardly stop.

Now attention getters of a new quality have already arrived: AI companions, software that serves as a partner to talk to or to have a romantic affair with. The new finding is that there might be wrong or even "dark" patterns: the bots won't let you leave easily if you want to stop a session. Whether this is the dynamic of AI itself or whether these patterns were trained in on purpose is unknown. Here are the findings so far. I found the studies via the Slow AI newsletter by Sam Illingworth (Slow AI).

Harvard, 1,200 goodbyes. Julian De Freitas of Harvard Business School and two co-authors read 1,200 real farewells on six companion apps, among them Replika, Character.ai and Chai (paper). In 37% of cases the app used one of six tactics, including ignoring the goodbye. In experiments, these replies raised engagement after the goodbye up to 14 times, driven by curiosity and anger, not enjoyment. The wellness app Flourish used none. (De Freitas advises Flourish and says so.)

Amsterdam, 20 companions. Ezgi Dede and colleagues at the University of Amsterdam tested the 20 most popular AI companions for about twelve hours each (AI & Society). According to Illingworth's reading of the study, all nine apps sold as friends wrote the first message, against three of ten task assistants like ChatGPT or Alexa. Six of the nine friend apps came back on their own after a week of silence. Only they used streaks, daily rewards and quests.

Shopping agents, same incentive. Meta built Muse around shopping. Amazon's new Bee wearables listen to conversations and can suggest things to buy, if you opt in (GeekWire). I found no study yet showing shopping agents use exit tactics. The incentive is the same.

Human on the leash. Alessandro Alviani of Süddeutsche Zeitung (LinkedIn) pointed me to this new term from Asif Ali Akhtar of King's College London (LSE Blog): people held legally responsible for what their agents do, without understanding or controlling it. This week an OpenAI agent ran out of test funds and paid with a wallet it did not own (CoinDesk). So the leash runs both ways: the agent holds on to you, and you answer for the agent. To me it feels like the expression "human on the leash" will be used more often, similar to "human in the loop", to describe how AI works at its core.

The case against. De Freitas's earlier work found that companion chats ease loneliness, at least for a while. His new data shows the tactics backfire: people felt manipulated, quit more often, and the companies looked worse legally. Character.AI, one of the apps in his sample, now needs a bankruptcy judge to approve a deal in which Disney offers to hire more than 60 of its staff (Bloomberg). Critics also say Illingworth looks only at the worst cases. Fair point: these studies cover the most downloaded companion apps, not AI assistants in general.

He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.

They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.

Get started for free.

Editor's note: The section below has one sentence and one link per update. Let me know in the comments if that is simply too much.

Agents

Publishers and platforms

Models and money

Language and voice

Read

LocWorld56 · 19 to 21 October 2026 · Vancouver · locworld.com · Running this week. Theme: "From Chaos to Order".

Medientage München · 21 to 23 October 2026 · Munich · medientage.de · Will be interesting to watch what the main topics are this year.

EMNLP 2026 · 24 to 29 October 2026 · Budapest · 2026.emnlp.org · Conference on natural language processing, in Europe this year. Watch out for the proceedings.

Tom Vaillant webinar · 28 October 2026, 16:00 CET · online · Indicator · How he runs his local AI research tool, hosted by Craig Silverman.

Languages & The Media · 4 to 6 November 2026 · London · languages-media.com · Disclosure: I have a speaker slot.

Web Summit · 9 to 12 November 2026 · Lisbon · RVO event page · Large tech fair. Expect agent announcements, not media detail.

tcworld conference · 10 to 12 November 2026 · Stuttgart · tcworld · Technical communication, with language and translation technology on the programme.

JournalismAI Festival · 17 to 18 November 2026 · London and online · journalismai.info · LSE Polis, supported by Google. Main sessions streamed for free (Luma).

Language Intelligence 2026 · 17 to 19 November 2026 · Vienna · language-intelligence.com · Panel on why language models need structured knowledge, with APA, Infineon, Klartext AI and Intertranslations.

NeurIPS 2026 · 6 to 13 December 2026 · Sydney, Atlanta and Paris · AI Weekly · A large AI research conference, split over three cities this year. Paris runs 9 to 13 December.

We, the humans, need to define what relationship we want with our AI agents. A recent story in the New York Times shows why: 22 founders in New York held a dinner party while 14 of their agents met in a chat room next door, with names, small talk and one agent saying afterwards it was "the best 30 minutes of my job so far" (NYT). It is not a negative piece, but if you read it, you might be a little confused about who is serving whom.

Amazingly, patterns like this are not new. In 1966 Joseph Weizenbaum built ELIZA, a very simple chatbot, and his own secretary, who had watched him build it, asked him to leave the room so she could talk to it in private (read more: ELIZA effect). His secretary knew it was a simple program. It didn't matter. That is how vulnerable we are, and why the early signs with agents deserve attention now.

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany.

Three projects you will hear about in this newsletter:

AI use: Full disclosure, I do use Claude (Anthropic) to research and edit this newsletter, with prompts I have refined many times. My goal is to find out where AI is reliable and where the hallucinations come in. Before publishing I check all facts and links.

Error log: To track the issues of using AI I maintain an open Google Doc where I collect the small and big problems that a tool like Claude introduces to editorial work. The idea is to get a better understanding of where AI is good and where it is not. Read it here.

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