Ole Reissmann | AI, Journalism, Media Strategy

Local AI autocomplete for writers: Cotypist and Cotabby

While you type, the computer guesses your next word. Press tab to accept it. The more you use it, the better it gets. By computer, I mean a small AI model running entirely on your machine. Private, fast, and free of API costs.

Typing with the help of Cotypist and Gemma 4

This can feel like actual magic. I’ve been chasing something like this since I first saw autocomplete in code editors. I even tried using Cursor for prose, pulled my own context via MCP… and ran out of tokens. There are better options now for Mac users:

Cotypist looks polished and costs $8/month or $72/year.

Cotabby is the free, open-source alternative.

Both apps suggest a local LLM or let you choose: a small Qwen 3, Apple’s built-in model, or Gemma 4 for better suggestions, if your Mac has enough RAM to run it. You can also bring your own model. And you can steer the suggestions by adding custom instructions and tell the apps how many words they should suggest, from 2-4 to 10 and more.

Both apps are a bit hacky in how they locate the cursor on your screen and overlay suggestions. With some apps, this can break the layout. I’ve seen at least two more recent entries to this new category of auto-complete apps, and there will certainly be more. Until Apple builds something like this right into the OS.

4.6.2026 read →

“We cannot watch as AI companies attempt to permanently dismantle the rights that give us control over the work we create. We cannot sit by as this work is used to build replacement products that undermine our ability to earn the audience and revenue necessary to continue reporting the news.” (New York Times publisher A.G. Sulzberger)

A.G. Sulzberger, speaking at the WAN-IFRA World News Media Congress in Marseille, called AI companies' use of copyrighted journalism "brazen theft at unprecedented scale" and urged news organizations to stop being passive about it.
Five of the top 10 sites used to train major AI models belong to news publishers, yet less than half of one percent of US AI investment goes toward compensating content creators.
Sulzberger argues fair use doctrine does not cover substitutive copying and regurgitation at this scale, and that paying for content would cost far less than the data centers and power plants AI companies are already building.

3.6.2026 visit →

75 cards to help your newsroom figure out AI before someone makes a bad call: A free kit covering ethics, workflows, and use cases, developed with 15 journalists and trainers across multiple countries. (DW Akademie and MethodKit)

DW Akademie and MethodKit released a free, downloadable 75-card kit to help newsrooms think through AI adoption, covering ethics, workflows, and use cases, developed with 15 journalists and trainers across multiple countries.
Cards are grouped into three areas: big picture questions, workflow analysis, and designing AI solutions, meant to structure team conversations rather than prescribe specific tools.
A physical version is not yet available, so for now this is a print-it-yourself resource, funded by Germany's Federal Ministry for Economic Cooperation and Development.

3.6.2026 visit →

UK publishers will be able to opt out of Google’s AI Overviews, per CMA, the UK competition watchdog.

The UK's Competition and Markets Authority has imposed a conduct requirement on Google, giving publishers the right to opt out of their content being used to power AI search features like AI Overviews, a world first under the UK's digital markets regime.
Google must also properly attribute publisher content with clear links in AI-generated results, and publishers can opt out of content being used for AI model fine-tuning.
Google has nine months to comply, must submit six-monthly reports, and the CMA is signaling further action on Google's search business in the coming weeks.

3.6.2026 visit →

Muscle memory for agents: The Printing Press. An agent-native CLI for any API, app, or site from a single prompt. The library already has 170-plus CLIs across travel, CRM, media, and more.

The Printing Press generates agent-native command-line tools for any API, website, or app from a single prompt, outputting a Go binary, Claude Code skill, and MCP server.
The pitch buried in the demos: a local SQLite mirror answering compound queries in 50ms does things a linear API cannot.

1.6.2026 visit →

Three publisher coalitions have already committed to three incompatible models, and whichever one becomes the default makes the other two marginal. (Ulrike Langer, News Machines)

Three publisher coalitions (SPUR, News/Media Alliance via Bria, and dpa) have committed to three incompatible monetization models: pay-per-interaction, blanket licensing, and publisher-owned distribution infrastructure. AWS has shipped agent-payment infrastructure via Bedrock AgentCore, but it only solves the agent side of the transaction. Publishers must build x402 endpoints themselves to receive payments, and almost none have done so yet.
The gap between Altman's podcast vision and what actually shipped from AWS is a useful reminder that infrastructure shapes outcomes long before policy does.
Whichever coalition model becomes the default will likely make the other two marginal, which means the choice publishers make now about which coalition to join is also a bet on which payment architecture wins.

1.6.2026 visit →

Six provocations from Reuters Institute researcher Felix Simon: The low-hanging fruit of AI in journalism has been picked, human-in-the-loop is becoming a polite fiction, and “responsible AI” labels won’t fix the trust problem.

Felix Simon presented six provocations at Nordic AI in Media Summit 2026 arguing the easy AI wins in journalism are already done: transcription, translation, summarisation, copy-editing are table stakes, not advantages.
The harder problems are structural: human-in-the-loop is becoming performative, AI may replace entry-level roles rather than just augmenting them, and audience trust in AI-produced news is lower than in almost any other sector.
The useful frame here is that "responsible AI" labeling won't move the needle on trust because audiences judge AI in news through the lens of institutional trust, not just tool quality.

1.6.2026 visit →

The Economist built a ChatGPT app that lets users query its Trump approval rating tracker. Maybe younger audiences will discover The Economist inside ChatGPT rather than through Google. (Andrew Deck, Nieman Lab)

The Economist launched a ChatGPT app called "The Economist - Graphs," making it the first major consumer news publication in OpenAI's app store, which opened to third-party developers in December.
The app is limited to U.S. polling data from its Trump approval rating tracker, a deliberate choice to avoid exposing paywalled content or undercutting subscriptions.
The goal is not traffic but brand discovery among younger audiences who now treat ChatGPT as a first stop for information rather than Google.
Another example for a media company would be Tubi where you can discover movies and shows and play them in an iFrame.

1.6.2026 visit →

Forget social sharing buttons: Hacks/Hackers rebuilt its site for crawlers and agents. “Ask AI” buttons, an MCP server, and an llms.txt file. (Burt Herman, Hacks/Hackers)

Hacks/Hackers rebuilt its site with "Ask AI" buttons, an MCP server, and an llms.txt file so AI agents can query its archives directly rather than scraping broken HTML.
The piece is a practical walkthrough: buttons hand users off to ChatGPT or Claude with pre-filled prompts, while the MCP server exposes ten tools for structured content access with full-text responses.
The most telling detail from their logs is that agents are already quietly crawling sites just to catalog what MCP tools are exposed, before anyone asks them to do anything.

1.6.2026 visit →

Local AI autocomplete for writers: Cotypist and Cotabby

THEFUTURE: Brazen Theft and Local Magic