Drift Notes: What’s happening in AI this week?

Research firm Pangram found that more than 40% of long-form LinkedIn posts are AI-generated, compared with around 30% on X and Medium. LinkedIn’s editor-in-chief, Dan Roth, says the posts all start to sound the same, which is a business problem: if people see too much generic content, they may leave for other platforms. So LinkedIn is using its own AI, with help from experts who teach it what authentic, useful posts look like, to weed out the generic ones. It has also scaled back its own AI rewriting tools and set up “content therapists” to coach executives and subject-matter experts on posting like humans.

LinkedIn isn’t anti-AI, but posts that haven’t been edited or personalized are less likely to reach a wide audience. Even the perfectly bulleted post with a hook up top is starting to read as a red flag. For AI trainers, it’s a familiar idea: a human eye that can tell authentic from generic is what teaches a model the difference, and here experts are literally training LinkedIn’s model on what good looks like.

As AI models need more specialized knowledge, highly paid experts like radiologists and attorneys are joining an industry long associated with gig work. LinkedIn ranked data annotators fourth among the fastest-growing occupations in the US this year. Recruiters are searching for software engineers, condensed matter experts, and civil rights attorneys, and some AI companies are hiring in-house “AI tutors” in subjects as varied as finance, Finnish, and comedy.

Companies are racing to secure specialized, privately held data that sets their models apart, instead of relying on what’s publicly available online. For AI trainers, it’s a sign that subject-matter depth is increasingly what counts.

Off the coast of Shanghai, a wind-powered data center sits about 10 meters underwater, using cold seawater for cooling instead of the fresh water a land-based facility would need. Built by Shanghai HiCloud Technology with state-owned partners at a cost of about CNY 1.6 billion (roughly $240 million), it’s wired directly to offshore wind turbines and is reportedly 95% powered by green electricity. One expert estimates underwater setups use 18–28% less total power for the same servers, and a sister site off Hainan has reportedly run for 40 months without incident.

Experts say it’s too early to call it green. The seawater returns to the ocean warmer, the company hasn’t released figures on heat output, and environmental reviews so far focus on construction rather than long-term operation.

It could speed up stem cell engineering and make lab-grown lung tissue more accurate, which would help researchers study diseases like asthma and pulmonary fibrosis. There’s a familiar lesson here known well by AI trainers: A model is only as good as the expert-built data behind it.

Anthropic announced the Anthropic Public Benefit Fund, a new entity that will receive 5% of its shares, subject to stockholder approval. The board can add 2% more for every 50% increase in the company’s value above a set baseline, so the fund grows as Anthropic does. Anthropic says it could make large donations to traditional nonprofits or direct distributions to specific populations, and the fund can’t be used for lobbying or political donations. Combined with existing pledges from its co-founders and employees, roughly 20% of Anthropic’s shares are now dedicated to the public good.

It’s a sign AI companies are starting to plan for the economic fallout of their own technology, though how the money gets spent is still undecided. It also comes as Anthropic prepares for a possible IPO as soon as November.