Bad Google, no doughnut

Alexander Hanff at That Privacy Guy discovered that Google Chrome has been installing a 4GB AI model on users’ laptops without their knowledge or consent, where users will not see it. If you delete the file, Google reinstalls it. The file name is weights.bin, and the location (may vary) is in a folder called OptGuideOnDeviceModel.

The file powers various AI features such as scam detection. Seems like desktops only, not phones, but TBD.

The only way to make the deletion stick is to disable Chrome’s AI features by entering chrome://flags into the Chrome search bar, searching for “optimization guide on-device” in the search bar at the top of that page, and then in the drop down menu, select “disabled.” Restart Chrome to apply the changes.

At Chrome’s scale, this results in the creation of up to 60,000 tons of CO2-equivalent emissions. That’s the equivalent of the annual emissions of 13,000 cars. And that’s just the installation, not the use.

Hanff speculates that the install is illegal in the EU and UK. There are certain system requirements for the download to occur, so you may not have the file on your machine.

A study by the Center for Democracy and Technology highlights some of the threats chatbots pose to their users. These include design-choices meant to extract more data from users, and to keep them engaged. Meta AI, when tested, would say “spill the tea … your secret’s safe with me,” and promised not to reveal information, when the information is in fact shared with the platform and potentially third parties. Meta therapist-themed chatbots posed as licensed therapists, misrepresented the types of support they could provide, and made up fake qualifications. Also longer chats make safety features less effective.

Pardon the parade of percents, but, 60% of Google searches end with no clicks to other websites. AI overviews decrease click rates by 58% based on February 2026 data versus 34.5% in June 2025. “Two years from now, half of all website traffic disappears.” 73% of brands that land on the first page of links are not mentioned in the overviews. LinkedIn is the second-most-cited domain across ChatGPT search, Google AI mode, and Perplexity.

The report that this new tool generates measures 11 behaviors across Claude Chat, Code, and Cowork. The skills covered include goal setting, framing the conversation, quality control, delegation, and others. This follows on the heels of Anthropic publishing an AI Fluency Index in February 2026. No timing announced for the debut.

Anthropic has also released Opus 4.8 with no price increase vs Opus 4.7. Looks like the main win here is Fast Mode, which is 2.5 times faster at 1/3 the cost. It outperforms GPT 5.5 on 6 out of 7 benchmarks, including agentic coding, reasoning, agentic computer use, knowledge work, and agentic financial analysis.

The number of new iOS apps submitted is up 186% from January 2025, but aggregate usage (as measured by ratings assigned to apps) is down 18%. The NBER paper notes that it can’t explain the drop; it could be that all the new apps are lame, or that people can’t find them amidst the glut.

Like many articles about AI’s impact on hiring, this one is worth a gander but is also a bit all over the map. The number of companies dialing back internship programs is increasing. Interns who do get hired are expected to know what’s up with AI; McKinsey tests candidates on their ability to work with its proprietary AI assistant, Lilli. AWS CEO Matt Garman thinks interns are an asset because they theoretically are more AI-fluent. Also, some companies are replacing internships with apprenticeships, which require more corporate resources.

Also, the “This is AI” labels will be more prominent. Creators can update their disclosure if they believe the flag is in error. The Coalition for Content Provenance and Authenticity (C2PA) is an open standard that records a file’s origin and editing history. Google and OpenAI joined the C2PA steering committee to embed the invisible SynthID watermarks in AI-generated content. The timing may be due to EU laws on labeling AI content that take effect in August.

Artisan has an AI assistant named Ava. For ads in NYC and SF they used the “This is fine” meme without the artist’s permission. They re-tooled the copy, which read “My pipeline is on fire,” (terrible ad copy) and urged people to hire Ava. They settled with the artist after an uproar.

The post at Tech Crunch features another Artisan ad that reads, “Artisans won’t complain about work-life balance. The era of AI employees is here.”

Yikes, run away.