Free desktop apps run AI on a laptop with no internet, and EU labeling rules start August 2

An employee pastes a client list into a chatbot and asks for a summary. That paste is how company data gets onto a server the company does not control, and it is routine: the security firm LayerX, in its Enterprise AI and SaaS Data Security Report 2025, found that 45% of employees at the companies it monitors use generative AI tools, that 77% of those users copy and paste data into the chat box, and that 22% of those pastes carry personal or payment card details.

LayerX also found that much of the pasting happens through personal accounts no IT department administers, which is why a written ban on chatbots is hard to enforce.

Two free desktop apps keep the whole exchange on the laptop. LM Studio and Jan download a compact model onto the machine itself and answer from there with nothing connected. LM Studio runs on Windows, Mac and Linux, pulling its models from Hugging Face's library; Jan is open source under the AGPLv3 licence.

A model that size is good for the document in front of it: summarize a long policy, pull the dates and amounts out of a contract, draft the reply. What it cannot do is look anything up, and it is smaller than the ones in the cloud, so the hardest reasoning still goes there.

Neither app charges anything, and once the model is on the machine, the file stays in the folder it was in and the answer comes back on the same laptop.

From Sunday, August 2, any company whose AI-generated images, audio, text or video reach people in the European Union has to label that content as AI-made, and has to tell a person when the thing answering them is a chatbot. The duty is Article 50 of the EU AI Act, it applies wherever the company itself is based, and the fines reach 15 million euros or 3% of worldwide annual revenue, whichever is larger.

The guidelines the European Commission adopted on July 20 put the duty on providers and deployers alike, which is to say on the business using the tool as well as the lab that built it. The age of the system makes no difference, so an image generator switched on in 2024 is covered from Sunday.

The marking that goes inside a file, the one other software reads rather than a person, does have a delay: a four-month grace period under the Digital Omnibus, through December 2, 2026, by the Cloud Security Alliance's reading of the negotiated text. The visible label and the chatbot disclosure have no such extension.

One written instruction to Anthropic's Claude Opus 5 now returns a playable 3D game. The model writes the geometry, the textures, the physics and in some cases the music as code, with no art files behind it. The AI news site The Decoder collected what people generated from single prompts over a few days: a first-person shooter, a submarine game, a Minecraft clone, a snowboarding game whose slope physics work.

A game is an odd thing for a plant manager or a claims director to care about, and a hard test: it has to keep working while a person moves around inside it, and any wrong line of code shows up as something visibly broken. Most of the software a business actually wants is easier than that. A calculator that prices a non-standard order, a screen that turns a shift spreadsheet into a rota: first drafts of those waited in a queue for a developer, or went to a software firm with a bill attached.

The developers posting the game demos say each one ran correctly on the first attempt, and each came back as a single editable file that opens in a browser.

Executives at 337 S&P 500 companies mentioned AI on their first-quarter earnings calls, by FactSet's count of calls held from March 15 to June 11. A study led by MIT FutureTech and Carnegie Mellon read the 10-K annual filings those companies file with the SEC, where securities law holds them to a high standard of accuracy, and found 21% of S&P 500 firms using AI in real operations by 2025 and 11% at the top level, deep integration.

Paul Cheek, a senior lecturer at MIT Sloan, has built a public index that scores each S&P 500 company against its competitors on how much it has implemented, reading job descriptions, patent filings, news reports and earnings calls rather than asking executives.

By sector, the filings study found deep integration at 5.88% of retailers and 5.71% of healthcare equipment and services companies, while 54.55% of energy companies did not mention AI at all.

Substack now runs Pangram's AI-text detector over posts, notes, replies and comments. A week after that, on July 29, Pangram released Pangram 4, which by its own benchmarks flags human writing as AI in 0.0041% of documents, about one in 24,000.

That rate is per document, so a team scanning tens of thousands of pages a month will see some human writing flagged. The detector returns a probability about a text, not a finding about a person. API scans now cost $0.05 per 100 words, priced by length rather than rounding each document up to 1,000 words, which Pangram says is up to ten times more for a long document. Pangram 3 keeps running until September 30, 2026.