As a reminder (especially for newer readers) - this newsletter is built by:

I started the newsletter because I didn’t see other newsletters taking this trend tracking approach, nor taking a market research industry lens to any of the AI news. I wanted to play with the AI tools and see how well they’d do tracking trends and analyzing the news; I also was reading all these newsletters and knew I was missing trends by reading them individually.

While this is, I hope, informative for folks, I’m REALLY excited to launch two new services this week.

First, a weekly AI Tips for MRX newsletter specifically geared for market research professionals. The tips will sometimes be specific to an LLM - mainly Microsoft Copilot, Claude, or ChatGPT - and sometimes general tips that apply to any LLM. You can sign up (and tell your friends to sign up) here.

Second, I am officially launching MRXplorer’s AI classes! There are three tiers: AI for beginners, AI for intermediate users, and AI for team leaders. You can take these as individual classes or as part of a cohort. Classes and cohorts start in September. See more at mrxplorer.com.

The AI learning community for market researchers is coming soon!

Now for the trends…

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

AI turned from a tool you ask questions into a coworker that runs work loops on its own — and the big AI companies went to war over who owns that loop. The past seven days redefined how people work with AI. Instead of typing prompts and waiting for answers, the new model is “tending the loop”: you set up a goal, the AI runs a continuous cycle of gathering, deciding, acting, and learning, and you step in only to approve, edit, or redirect. By midweek, OpenAI and Anthropic (the company that makes the Claude AI assistant) both made moves to own the screen where that loop lives, triggering a backlash that proved the shift is real.

Stories that got stronger as the week went on — or are new this week

Nobel‑prize‑winning economists told the world to take AI job disruption seriously — now.
A statement titled “We Must Act Now,” signed by 16 Nobel laureates and over 200 researchers, warned that AI could reshape the labor market faster than the Industrial Revolution. The statement landed Tuesday and was covered the same day by two major newsletters. On Wednesday, the companion “AI 2040” plan — a detailed scenario for a U.S.–China slowdown on frontier AI training — gave the warning a concrete policy blueprint. The statement deliberately avoids prescribing specific laws, which helped it attract a broad coalition from across the AI safety spectrum. (The Rundown AI, 2026-07-14; AI Daily Brief, 2026-07-14; AI Daily Brief, 2026-07-15)

AI infrastructure got hit with both financial and physical regulation in a single 48‑hour window.
The Bank of England designated Amazon Web Services, Google, Microsoft, and Oracle as “critical third parties” that must pass financial‑stability stress tests. New York’s governor froze permits for new large data centers for 12 months. And Demis Hassabis, the CEO of Google DeepMind (a leading AI research lab), proposed a U.S. model‑vetting body modeled on FINRA, the agency that oversees stockbrokers. The combined effect: the cost of running AI just went up, and the hardware layer that makes AI possible is now a political issue, not just an engineering one. (AI Governance, 2026-07-14; The Rundown Tech, 2026-07-14; The Rundown AI, 2026-07-15)

Signals appearing in 4 or more of the last 8 weeks (Long‑term Continuing) — keep brief

Multi‑model routing as operational must‑do — 6+ weeks running
No organization can afford to build on a single AI model. This week, Every’s business‑operations team documented a live case study where they routed coding to Fable, analysis to Codex, and cost‑sensitive work to a cheaper model — all inside one project. The case study turns “use multiple models” from a hedge into a working template. (Every, 2026-07-15)

Human judgment as the scarce, non‑negotiable layer — 7+ weeks running
The “tending the loop” paradigm made this concrete: the human’s job is to set standards, catch errors, and exercise taste on AI outputs. Slow AI’s new field guide for dismantling false consciousness claims (published Wednesday) reinforces that any AI that says it “understands” is only showing functional access — not felt experience. That distinction is the same one needed when an AI claims to “understand” what a survey respondent meant. (Slow AI, 2026-07-15)

Token cost governance as daily discipline — 12+ weeks running
The new infrastructure regulations and the price‑war model launches this week (Meta’s Muse Spark 1.1 at $1.25 per million tokens, Grok 4.5 at one‑fifth the cost of comparable frontier models) mean the cost of running AI is simultaneously dropping at the token level and rising at the hardware‑compliance level. Teams that do not track cost per task will be caught by surprise. (Neatprompts, 2026-07-10; AI Daily Brief, 2026-07-10; AI Governance, 2026-07-14)

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

The authenticity backlash is forcing disclosure rules — 3 weeks running
AI‑generated content is flooding every channel, and audiences are learning to spot it immediately. This week, the European Union published a new Code of Practice for labeling AI content, and LinkedIn quietly rolled out an “Authenticity Update.” For any organization that produces research, the message is: AI‑first analysis without a human‑differentiated insight will soon be as recognizable — and as damaging — as an AI‑written LinkedIn post. (The Signal, 2026-07-12; Joanna Byerley, 2026-07-12)

Anthropomorphism rules are arriving from unexpected directions — first significant appearance
China’s new law on human‑like AI interaction took effect July 15, banning AI services that induce emotional dependence or manipulate users. On the same day, a detailed critique of Anthropic’s research on a “global workspace” inside Claude argued that the company is marketing functional access as consciousness. Taken together, any AI that claims to “understand” or “care” is now under both regulatory and ethical fire. (Luiza Jarovsky, 2026-07-15; Slow AI, 2026-07-15)

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

The long‑term trend is that AI is no longer a tool you use for a task — it is a coworker that runs work loops while you supervise. For insights teams, that means every AI‑assisted deliverable must have a named human reviewer who checks outputs against domain standards, not just against the prompt. The short‑term emerging trend is that the cost and availability of the AI infrastructure underneath those loops is now subject to sudden government action — data‑center freezes, financial stress tests, and model‑vetting proposals all arrived this week with no warning. This week’s arc tied the two together: the platform war over the agentic home screen shows that the loop is the new battleground, and the infrastructure regulations show that the hardware to run those loops is no longer a safe assumption. For an insights agency, the minimum viable response is a published routing policy that names which model does which task, what the fallback is, and who verifies the output — because the procurement question has permanently changed from “are you using AI” to “which model did what, at what cost, and who checked it.” Agencies that skip this step are selling speed without a quality guarantee, and that position becomes harder to defend every week.

Z’s Take

I think the idea of “AI as coworker you supervise” is still just a bit of a stretch in some cases. As a frequent AI user, can I just say how often these things don’t work? AI is the only coworker we keep paying for despite its failure rate. It’s the only tool that gets a glowing review despite getting factual answers wrong 11-15% of the time in straightforward cases, and 50-80% of the time in specialized domains.

There are some cases where AI is genuinely getting better; we’re also learning how to use AI better. It still takes an awful lot of supervision depending on the task it’s given.

I think the bigger trends to watch are both the various ways in which governments and private entities are starting to emerge with “hey, let’s slow this tech pace down a bit” plans; and the increase in “authenticity” tools to identify human-made from AI-made content.

The Bank of England move to add cloud providers to the list of Critical Third Parties shows the increasing concern about the impact of cyber attacks and tech failures, which I can only think comes from AI model advancement and reports of those models revealing critical issues; concerns from security specialists about how much data is being leaked - knowingly and unknowingly - to AI providers via LLMs; and hackers finding more ways to mess with systems by using AI themselves. Add the list of economists calling for the world to pay closer attention to the level of market disruption coming about due to tech advancements, and we might be reaching a critical mass of groups calling for a slow-down for systems to catch up to the new realities.

And the pushback on Meta for saying anything posted on Instagram public accounts was fair game for anyone else to use in creating AI content plus the LinkedIn “authenticity update” shows people are getting increasingly tired of the AI slop that is posted without acknowledging the content is AI-generated. (I personally think it’s one thing to post something from AI and acknowledge it’s AI and another to pretend it’s your own content. At least be upfront about what’s you and what’s AI.)

What does this mean for market research? Perhaps if enough entities call for a more thoughtful approach to what the tech advancements mean to society, not only will there be interesting research being done addressing that question, but we will also see more emphasis on research quality over research quantity, and we will have time to establish the new paths needed for entry-level professionals to gain what is currently mid-career expertise, and time to re-envision insights careers overall.

This newsletter covers Friday, July 10 – Thursday, July 16. Sources: Every, The Rundown AI, simple.ai, AI Daily Brief, The Signal, Lenny’s Newsletter, Slow AI, AI Snack Club, Neatprompts, Unpromptable, Joanna Byerley, AI Maker, Nicolle Weeks, AI Governance Ethics and Leadership, The Rundown Tech, Luiza Jarovsky, Slow Takes