I had the opportunity to research IT professionals during the earlier days of cloud technology. I got to see the disillusionment as the overpromise/underdeliver cycle happened with them. They were promised tons of efficiency gains and price reductions, only to be met with neither. Then, they were met with security issues like data being hacked and literal power issues where servers were going down and service was disrupted.

It brought on “disaster recovery planning,” which was a fancy way of saying, “don’t put all your eggs in one basket.”

That’s basically what everyone who was writing about AI talked about this week in some form. When a model can be pulled so abruptly, or costs can skyrocket overnight, having your systems rely on a single provider or system suddenly doesn’t look like a great idea, no matter how awesome that provider has been (or seemed to be) in the past.

As Slow AI’s Dr. Illingworth offers, take the time to think through what would happen if you were to lose the platform on which the process is currently operating. Give it a red/yellow/green for risk, and then create those fallback plans for the workflows that you identify as “red.” A little planning now will save a lot of disruption later.

On to this week’s AI-generated pattern discoveries!

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

The U.S. government shut off access to the two best AI models from Anthropic (the company that makes the Claude AI assistant) — and a week later, there is still no fix. The government used export rules (laws that control what U.S. companies can send to other countries) to force Anthropic to turn off its top models, Fable 5 and Mythos 5, for everyone who is not a U.S. citizen. This matters because it shows that access to powerful AI tools can be taken away overnight by government order, with no warning and no clear way to get it back.

Stories that got stronger as the week went on

Don't build your work around one AI model
Once Fable 5 was shut off, newsletter after newsletter said the same thing: if your daily work depends on one AI model, you are at risk. The advice grew into a shared playbook by midweek.

Open models you can't have taken away
As the week went on, AI models you can download and run yourself became the clear answer to the shutdown risk. A government cannot revoke a model whose code is already public.

Knowing your field beats knowing the tool
Two separate studies this week found the same thing: people who deeply understand their job get more out of AI than people who are just good with the AI tool.

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

Big AI tools work in demos but fail in real life — 8 weeks running
For weeks, the same gap keeps appearing: AI looks great in a showcase but struggles or fails when real work depends on it. This week the gap took a new form — a model that worked one day was simply gone the next.

Outside money is buying up AI power — 5+ weeks running
Private investors, sovereign wealth funds, and now rocket-company stock keep funding the computers and tools that AI runs on. This week it reached the model-building layer itself.

Agentic workflow design replaces “prompting” as the core skill — 8 weeks running
The durable skill in AI work is no longer writing good prompts — it’s designing loops: cycles where an AI does work, evaluates against a goal, incorporates feedback, and repeats. This week’s evidence: GitHub’s COO confirmed the “developer versus non-developer distinction is disappearing,” with legal, finance, and marketing professionals using AI to build prototypes. Anthropic’s own 400,000-session study found domain experts (lawyers, scientists, managers) finished within seven points of software engineers on coding tasks — because understanding what “good” looks like matters more than knowing how to code. (Every, Jun 17; The Rundown AI, Jun 18)

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

Government using "national security" to pick AI winners and losers — 2 weeks running
The same power that shut off Anthropic's models is also protecting other companies. This creates two tiers: AI whose access can be cut, and AI that gets legal shelter.

Junior workers can't build skills they never use — Short-term Continuing
New data shows young workers in AI-heavy jobs are getting squeezed: companies expect senior skills from junior hires but cut the training that builds those skills.

The Fable 5 shutdown is not a one-time event — it is the demonstration of a new structural reality for anyone using AI to understand what people think or want. Any insights workflow (the process of turning survey data, interviews, or customer conversations into analysis) that depends on a single frontier model is now a liability, not a methodology. The practitioner response this week — the model-independence audit, the context vault of portable markdown files, the multi-model routing policy — is the minimum viable governance standard for any team running AI-assisted analysis.

The longer-term trend is the separation of tool from skill. The Anthropic study showing domain experts outperforming coders on AI-assisted tasks confirms what market researchers have always known: understanding what a good answer looks like matters more than fluency with any particular tool. The agencies that will win the next procurement cycle are not the ones with the fanciest AI stack — they are the ones who can prove their researchers own the planning decisions and can evaluate AI output against domain standards.

But trust data from Pew is a warning light the industry is not heeding. Adoption is rising while trust is falling, and the heaviest users are the most skeptical. When clients pressure agencies to “use AI more,” they are responding to industry narrative, not consumer demand. The right question is not how much AI to use — it is what work the AI is doing, who is evaluating it, and what happens when the model changes or disappears.

Data quality remains an unresolved crisis underneath all of this. Bots generating human-looking survey responses are getting harder to detect. Speed claims from AI vendors are outpacing methodology rigor. And the hidden labor cost — researchers cleaning up AI outputs, checking for quality, switching between tools — is eating into the productivity gains AI promises. The agencies that build transparent, auditable, multi-model workflows with clear human oversight will be the ones clients trust when the next model goes dark.

Z’s Take

I really enjoyed Dr. Illingworth’s recommended audits, as they typically are designed to shine a light on where you are giving away your skills or thinking to AI versus keeping the skills sharp on your own. This week’s recommendation to document the workflows that rely on AI to complete is a major step in identifying the risk if tools break, increase in cost dramatically, change privacy policies in ways unfavorable to customers, or disappear entirely. This isn’t just an AI issue. It’s a general tech issue, especially relevant in insights where the tech market is still flooded with options. When acquiring new tech, it’s always important to be thinking a few steps ahead and asking about data portability if contracts end for any reason.

The study showing domain experts outperform coders is one that should feel validating but should also continue to worry us as an industry. If experience wins over just using AI, then how do we make sure those entering the field are able to gain the experience needed to continue winning over just using AI? Training entry-level staff on what “good” looks like in insights is still a gap to be addressed so we don’t end up with a pipeline gap.

The Pew study is interesting; younger Americans (under 30 years old) are using it most, but also most negative about it. Unfortunately, while the AI summary states that agencies being pressured to “use AI more” are hearing from companies who are responding to industry narrative rather than consumer demand, the pressure is real, and real projects are lost if the customer on the deciding end of the RFP doesn’t think you’re using AI “enough.” When will the industry narrative catch wind of the negative sentiment and stop asking for agencies and freelancers to describe how they’re using AI?

And last, but not least, speed claims have outpaced methodology rigor for a long time now. AI hasn’t changed that in the least. What AI has changed is the speed at which companies are trying to catch fake data in online surveys, and, increasingly, qualitative in-depth interview participants who are using AI to answer questions and pass screening criteria. But I do agree that the work that goes into checking that the AI didn’t actually hallucinate a quote (or an entire respondent) in a report, switching between tools, and even keeping up with what platforms are capable of doing adds up to the argument for, “Why should I even bother using AI if it’s not really saving me that much time on my research projects?”

This newsletter covers Friday, June 12 – Thursday, June 18. Sources: AI Daily Brief, The Rundown AI, The Rundown Tech, The Rundown Robotics, Every, Slow AI, Slow Takes, simple.ai, AI Maker, AI Governance Ethics and Leadership, Neatprompts, Human+AI, The Signal, Lenny's Newsletter, Prompt-Led Product, Luiza Jarovsky PhD