A little note about the behind-the-scenes for this newsletter. I originally built the newsletter using Claude to run the analysis and to write the newsletter. When costs rose, I spent a weekend testing other models to find a cheaper model that could replace Claude. My options at that time were many, and I tried them all, but the results showed that Claude still won for ability to follow the prompt and write the outputs.

But models do improve quickly. Over a matter of two weeks, I was able to move off of Claude and over to Gemini for the daily digests, with DeepSeek writing the weekly newsletter. My initial test of this system worked well, with data being accurate, including citations. I was good to go!

Or so I thought.

I caught Gemini hallucinating like crazy this week. When the newsletter draft was done, I was confused by how much it looked like last week’s newsletter. The stories were nearly identical - and I do still read the newsletters I get in my email, so I knew that there was much more that had happened this week in the world of AI. Why wasn’t the newsletter picking that up? Then I saw references to data that had already been cited last week, but being reported as cited by the same newsletter again this week. I checked my email, and the newsletter in question hadn’t even published on the date listed! I checked daily digests, only to find too many of them recycling last week’s data or just incomplete!

I tried updating the prompts to be more specific, including specifically listing every newsletter source. It didn’t work.

So, I swapped out Gemini for the same model used to write the weekly newsletter, re-ran the analyses for the entire week, checked all the data, found it was accurate this time, and then re-ran the newsletter. WOW! What a difference!

This just goes to re-emphasize: just because you have an agent that works the first time doesn’t mean it works the second or the 15th time. If the output is going to an audience, check it each time for accuracy. It’s worth doing for your own peace of mind.

On to this week’s AI news!

What moved in AI this week — plain English, weekly arc

Voice AI crossed from a demo you try once to a tool people use every day. OpenAI released GPT-Live, a voice assistant that listens, talks, and thinks at the same time — and it delegates hard questions to a stronger model in the background without the user ever noticing. Early testers describe it as a shift from software to a "live cognitive presence." The same week, a new survey of nearly 6,000 tech workers showed the workforce is splitting into two groups: those amplified by AI and those shaken by it.

Stories that got stronger as the week went on

Anthropic found a hidden "workspace" inside Claude that shows what the model is really thinking — and the discovery spread from a research paper to an operational tool in one day.
Anthropic (the company that makes the Claude AI assistant) discovered a small set of internal concepts the model uses to steer its answers, separate from the text it shows the user. The research team proved they could read and rewrite this hidden workspace, changing the model's answers without touching the visible output. This turns AI from a black box into something that can be audited.

Token cost governance hardened from a practitioner concern to a named business metric: "revenue per million tokens."
Three newsletters on Wednesday independently converged on the same message: managing what AI costs per task is now the central operational challenge, not a technical footnote. A new efficiency metric emerged that could replace "revenue per employee" as the way companies measure productivity.

China may restrict overseas access to its best AI models, making the AI supply chain a two-sided risk.
The U.S. has already limited access to the most powerful American models. Now Beijing is discussing the same thing in reverse — potentially cutting off open-weight models like Qwen and GLM that Western companies have adopted as cheap alternatives.

Signals appearing in 4 or more of the last 8 weeks (Long-term Continuing)

Sovereign AI access is now a two-sided fracture — 5+ weeks running
Governments on both sides of the Pacific can now shut off access to frontier AI models. This started as a U.S. action against Anthropic in June. It is now a structural feature with China potentially mirroring the same restrictions.

Token cost governance as operational discipline — 11+ weeks running
Managing how much computing power each AI task consumes is now a core business practice, not a technical afterthought. The conversation has shifted from raw capability to cost-per-task.

Human judgment as the scarce, atrophying layer — 6+ weeks running
The skill that matters most is verifying whether AI outputs are correct — not generating them. The verification tax keeps growing.

Signals appearing in 2–3 of the last 4 weeks (Short-term Continuing or Emerging) — keep brief

Authenticity tension between AI-generated and human-created content — 2 weeks running
Instagram's head calls AI "a tailwind for authenticity" while practitioners publicly list their AI non-negotiables and model reviewers note that even the best writing models require heavy human editing to avoid sounding generic.

Multi-model routing as governance doctrine — 5+ weeks running
Using multiple AI models for different tasks — and switching when one becomes unavailable or too expensive — is becoming a required practice. This week, the bidirectional access fracture made routing a compliance layer, not just a cost tool.

Why any of this matters if your job involves understanding what people think or want

The long-term trend is clear: AI access is no longer stable, and the risk now comes from both directions. Any research methodology that depends on a single frontier model — American or Chinese — carries a supply-chain risk no service agreement covers.

The emerging short-term trend is the workforce splitting into "amplified" and "shaken" groups, with burnout rising as AI becomes ambient and always-on. This week's voice AI launch and interpretability breakthrough make both trends concrete: the tools are getting more powerful and more opaque at the same time.

Anthropic's hidden workspace discovery means the claim that "we don't know how the AI arrived at this conclusion" is no longer technically true — it is a disclosure choice. Agencies that do not surface the model's reasoning layer invite clients to ask why they are not using the available audit tools. Meanwhile, Slow AI's liability framework gives insights buyers the vocabulary to demand a named human verification step before any AI-generated analysis reaches them. An AI-moderated qualitative report or auto-generated tag set that arrives without a documented domain-expert review is a structural liability, not a speed advantage. The "revenue per million tokens" metric means clients will soon demand that agencies show how AI spend translates to insight value — not just that AI was used.

Z’s Take

Well, this week sure got interesting! Just as groups started moving to Chinese models for their AI work, China took a page from the US government and started talking to Chinese AI companies about gating releases of future models and limiting foreign access to models.

Sovereign AI - or AI models that a group can train themselves, own themselves, and not have to rely on anyone else to provide - is likely to become the next big thing. Microsoft is already putting their bet behind it with a business unit dedicated to helping companies do just that: build their own AI models.

For market research, this gets really interesting because most market research tech companies have been relying on models provided by frontier companies like Anthropic or OpenAI to power the systems and services that they provide to the industry. So, if cost becomes an issue because those models are becoming extremely expensive to run, and they switch to models that are cheaper to use, such as those from China, but then China pulls those models or limits access to those models, then does that mean we're going to see a new wave of market research companies that are training their own AI models to power their tools themselves, so that they don't have to worry about outsourcing the LLM?

Switching topics: I have to wonder about the conclusion that AI makes here about “the report or …tag set that arrives without a documented domain-expert review becomes a structural liability and not a speed advantage.” One of the things that we see in this industry is that “good enough” results tends to be the deciding factor for a lot of the technology that gets adopted. If technology can deliver a result that is good enough, then why spend extra time and extra money on something that provides a result that is even better? We saw this with panel companies that provided results to companies fielding surveys that used to take weeks to complete and now were taking days. Data quality took a back seat because the survey answers were good enough to make decisions from, and as much as buyers wanted to say they were concerned about data quality, when faced with increasing the cost to increase the data quality or keeping their “good-enough” quality, most stayed with “good enough.” So, will “documented domain-expert reviews” become a requirement for companies buying data? Probably not. What will be a requirement is domain-expert application of that data to the business and knowing what data matter to the business.

This newsletter covers Friday, July 3 – Thursday, July 9. Sources: The Rundown AI, AI Daily Brief, Every, Lenny's Newsletter, Slow AI, AI Governance Ethics and Leadership, Neatprompts, AI Maker, AI Snack Club, Nicolle Weeks, The Rundown Tech, Luiza Jarovsky PhD, QuestionPro, On New Terms, Gen Purpose, Prompt-Led Product, The Rundown Robotics, Slow Takes