ChatGPT Health, a $1.5B Copyright Payout, and AI Detection on Substack
Dear Friend,
In our AI Smart Labs session this week, the whole team built a newsletter together inside Claude. Everyone ran the same workflow: a research step that gathers and verifies the week’s stories, a human review checkpoint, and a writer step that turns the approved links into a draft.
We also fixed a habit that was quietly costing us work. Chat history is not storage, so anything worth keeping now gets exported straight to a Google Doc where the team can comment, edit, and track versions.
A process is only done when anyone on the team can run it without the person who built it.
Here are the top stories from this week.
What AI Consulting Really Costs a Small Company
Every consulting conversation we have starts with the same unspoken question: what is this going to cost me? This article answers it with real numbers. AI consulting for a small company runs $25,000 to $150,000 depending on scope, and companies that maximize the tools they already own, like Microsoft Copilot and Google Gemini, spend 40 to 60% less than the ones that buy new platforms. You also get the questions to ask any consultant before you sign, and the red flags that mean you should walk.
Our Most Requested Service: AI Strategy and Consulting
Clients keep coming to us with the same picture. They know AI can help, but they are not sure which tools to trust, what order to tackle projects in, or how to keep their data safe while the team adjusts.
In a strategy engagement, we work with your whole team to pick the tools, the projects, and the order to tackle them in. We weigh team capacity, budget, security, and your AI use policy before anything gets built, so the plan fits the business you actually run.
You walk away with a practical 12 month roadmap and a guide through implementation, not a pile of subscriptions and guesswork.
Get your strategy session today
OpenAI Opens ChatGPT Health to Every Adult in the US
OpenAI rolled out ChatGPT Health to all US users 18 and older on every plan, from Free to Pro. People already ask ChatGPT more than 300 million health questions a week, up from 230 million in January. You can now connect Apple Health and medical records from US hospital systems, One Medical, and Function Health, and ChatGPT will use that context in any conversation. OpenAI says connected health data is never used to train its models or target ads.
My Take: 70%. That is how many health questions happened outside OpenAI’s dedicated health hub during testing, so the company tore down the wall it spent 6 months building. People do not file their health in a separate drawer, they mention a sore knee while planning a weekend, and OpenAI just admitted the separate-app era of AI is over. Context that follows you is the product now, and the price is how much of your life you connect. We felt the fragile side of connected context in our own stack this week when a blog post silently failed to publish because a single integration had disconnected.Check the wiring before I trust the output.
Source: OpenAI
MIT Teaches AI to Turn Flat Drawings Into 3D Models You Can Test
Researchers at MIT built a system called GIFT that teaches vision-language AI models to write accurate CAD programs from 2D designs. An engineer can take a flat drawing of a part and get a working 3D model to run through virtual crash and durability tests. The generated programs are far more accurate and functional than what these models produced before.
My Take: Put this next to last week’s MIT story, where students who had never seen the inside of a jet engine fired a working one in 4 weeks with AI copilots. The expensive middle of building, turning an idea into something you can test, is collapsing toward free. The moat is no longer whether you can build the prototype, it is knowing which prototype is worth building. We watched the same shift in our own shop this week when newsletter production became a workflow anyone on the team can run. The steps got cheap. The HUMAN review checkpoint is where the value remains.
Source: MIT News
Washington Threatens Sanctions Over a Chinese AI Model
Treasury Secretary Scott Bessent said sanctions and Entity List designations are on the table after the White House accused China’s Moonshot of secretly distilling Anthropic’s Fable model to build its new Kimi K3. Distillation means training a smaller model on the outputs of a bigger one, and it is normal engineering when you have permission. Bessent also said the US will examine Chinese open source models for signs of IP theft. In plain terms, the government now treats the knowledge inside an AI model like a controlled export, and a model you build on can get caught in a trade fight.
My Take: Anthropic is the quiet winner here. The US government just treated its model weights as a national asset worth defending, and that is a moat no marketing budget can buy. The cost lands on the rest of us, because once a model becomes a geopolitical asset, access and pricing follow politics instead of product. This is why we documented exactly what instructions and context feed each of our tools, so any workflow can move to Gemini or Claude in an afternoon. We even route work by model now, one for the newsletter and another for proposals. No single vendor holds everything we do.
Source: TechCrunch
Authors Start Getting Paid in Anthropic’s $1.5 Billion Book Settlement
A federal judge gave final approval to Anthropic’s $1.5 billion settlement with authors and publishers, the largest copyright settlement in US history. Rights holders get about $3,000 for each of roughly 500,000 pirated books, and 92% of those eligible opted in. The court had already ruled that training AI on legally bought books is fair use. In plain terms, Anthropic is paying for how it got the books, not for the training itself.
My Take: The authors won the check and lost the war. $3,000 a book sounds like justice until you notice the fair use ruling still stands, which means every lab reads this as a green light with a receipt requirement. The legal fight has moved from whether AI can learn from your content to whether the content was taken or bought, and provenance is now the whole game. Inside our team the rule is plain: just because we can get information does not mean we own the rights to it, and our AI use policy spells out what goes into a tool and what never does.
Source: Publishers Weekly
Substack Lets Readers Scan Posts for AI Writing
Substack built AI detection from Pangram into its app. Readers can scan any post, note, reply, or comment over 100 words and see an estimate of how much a person wrote versus AI. It covers content published from July 21 on, and writers can scan their own drafts, dispute results they think are wrong, and add a note explaining how they work. Pangram’s data shows more than 40% of long-form posts on LinkedIn flag as fully AI-generated.
My Take: Human writing used to be assumed, and now it gets verified, which makes verified human a premium label the way organic became one for food. I expect other platforms to copy this within a year, and when they do, sounding like everyone else becomes a cost you can measure. We have seen fully human written pieces get flagged at 98% AI. It’s a nice idea and I don’t believe the technology is really there yet unless people are bad at or neglect to craft a really good AI voice.
We retune our newsletter voice in almost every AI Smart Labs session, feeding the tools real samples of our writing, and speaking, and cutting the generic lines that creep back in, because the models drift and a person has to catch it. Would your content pass a scan your readers can now run?
Source: Substack
Tool picks of the week
Voice Instructions Creator: Analyzes your writing to find your tone and style, then builds custom voice instructions for your AI tools. With Substack letting readers scan for AI writing, this is how we keep our drafts sounding like us instead of like a model.
ClickUp: A platform to manage tasks, notes, and chats in one place. This week we are moving our newsletter SOP and model-selection guide into it so the whole team runs the same playbook.
Cyclewise.AI: A personalized wellness platform for women with tailored fitness, nutrition, and well-being insights. A fitting pick the week ChatGPT Health opened to everyone, if you want AI on your health without connecting your medical records to a chatbot.
Got a burning question, a fresh take, or just want to share your latest AI wins? Hit us up at [email protected]. Your insights keep this community growing and thriving!
See you in the Lab,
– Nicole A. Donnelly
Founder, AI Smart Ventures
AI Strategy – AI Training – AI Consulting – AI Implementation
