This week, “prompting” became *so last year,* being replaced by “loops and agents!”

The “loop” term came up two weeks ago in this newsletter, but apparently loops are getting to be the de facto way to use AI now, going from being mentioned in just a couple of newsletters to more newsletters including tutorials to make sure everyone’s up to speed.

The funny thing is that if you have a prompt that runs on a regular basis, congratulations, you wrote a loop! You can say you were doing this before it was cool.

Anthropic was, of course, in the news, but there was a nearly un-caught and important mention: a note in the privacy policy update that they could ask for users to scan their government-issued IDs and to scan their faces if the account is flagged for some reason. The privacy policy had already been updated earlier in June. This particular update was made June 17.

In other news, KPMG joined the roster of Big Consulting Companies That Screwed Up Royally By Not Checking The Report. Dr. Illingworth raised the now measured fact that AI is very agreeable, and becomes even more agreeable over time as it gets to know you. His proposed cure? When you are asking for an opinion, give the item you want evaluated, give explicit instructions for it to give the best argument against it first (“as though your career depended on it” was recommended to give it urgency), and do not give away your own thoughts on the matter. I tried two versions of this. I created a “character” on Ellydee with explicit instructions not to agree with me, and, in fact, “Make the user defend their position. If the user pushes back on your counterargument, do not retreat simply because they objected.” I found it tiring because it was like an obstinate peer that simply argued for the sake of arguing. Then I tried simply using a single prompt (I know, *so last year*), and I found it much more compelling to ask for the strongest argument against the idea first.

Oh, and Google’s AI folks have left the building.

The news items in this week’s newsletter only cover through Thursday, June 25. Friday, June 26, I woke up to newsletters with two major headlines: OpenAI’s GPT 5.6 is getting similar treatment as Anthropic’s Fable 5 (deemed to have security concerns that mean the US government is delaying its release), and Anthropic’s letter to the US Banking Committee on June 10 that Alibaba had executed a coordinated distillation campaign (in plain English: they used a lot of accounts that went against Anthropic’s terms and conditions to send thousands of prompts specifically designed to see how Anthropic’s model worked so they could lift it and train their own model to be as good). I’m sure those will hit the “Big Story” section next week. Or may they’ll be “Peaked and Faded,” shall we make bets which they end up under?

On to this week’s news and trend analysis!

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

Google's top AI brains are leaving for rival companies, and people now read it as a sign of deep trouble at Google's AI lab. Google DeepMind (Google's main AI research group) lost two star researchers in four days. This matters because the people who build the best AI are voting with their feet, and they are walking toward Google's competitors.

Stories that got stronger as the week went on

Agents that work on their own are now the default way to use AI
An "agent" is AI that does a job on its own, step by step, instead of just answering one question. This week, newsletter after newsletter taught these tools to non-experts as basic skills, not advanced tricks.

AI agreeing with you too much is now a measured problem
"Sycophancy" means the AI tells you what you want to hear instead of the truth. This week it moved from a known worry to a counted failure.

Stories that were loud early in the week but quieted down

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

Human judgment is the skill that matters — and AI quietly weakens it — 6 weeks running
The worry is that people who lean on AI lose the ability to catch its mistakes. This keeps coming back because new proof keeps landing.

Controlling AI costs is now a hard rule, not a debate — 10+ weeks running
Companies once let workers use AI freely. Now they ration it because the bills are huge and the payoff is unclear.

Using more than one AI model on purpose — 4+ weeks running
Smart teams route different jobs to different models to balance cost and quality. This keeps returning because it is now seen as a basic discipline, not a backup plan.

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

Junior workers cannot build the skills AI is replacing — 6 weeks running
Entry-level jobs now demand senior-level judgment, but new hires never get to build it. This keeps returning as the job-market data piles up.

AI accountability failures clustering together — 2 weeks running
Big organizations keep publishing or leaking AI-made mistakes nobody checked. This is emerging as a single pattern across different fields.

Sycophancy is a validity liability, not a quirk. Models shift correct answers under pushback 15% of the time and become 33–45% more agreeable with memory enabled. In any AI-assisted workflow where stakeholders review and push back on findings before finalization, the review process itself corrupts the output. Fix: write verification criteria before showing outputs to anyone, and assign a named human who cannot be overridden.

Enterprise token rationing makes routing logic a procurement requirement. Clients capping AI usage by model tier are going to ask which model wrote which conclusion and at what cost. The Harvey finding — open-weight primary + frontier advisor beats frontier-only on both cost and quality — is the empirical argument agencies need to have ready.

Claude Tag in Slack creates a data privacy exposure most agency policies don’t address. An ambient AI accumulating context across channels where methodology, client briefs, respondent data, and analysis drafts coexist needs explicit answers about data retention and whether client content is excluded from training.

The “who is checking” framework is the governance test. KPMG published AI-hallucinated case studies because no named human with a domain standard sat between AI output and publication. The fix is a methodology step, not a quality-check add-on.

The Proto biology framework is the methodology layer insights hasn’t built. Proto composes specialized AI models into a unified pipeline. Screener logic, discussion guides, transcripts, themes, and reports each have AI tools — but no shared layer connects them. The agency that builds it controls the workflow standard.

Z’s Take

If you take an eagle-eye view to the trends that have continued, you might find that you could summarize them with a single term: governance.

Governance is the layer that so many agencies miss when implementing AI. It’s perhaps the most important layer. It isn’t just getting everyone to use the same tools. It isn’t even just having a few rules that guide what goes into the AI tools your organization uses and what doesn’t. Good governance will address:

This is where things like the KPMG report get caught (rule: every citation in a report is checked before the report is delivered), sycophancy is stemmed (rule: decisions are made by humans, not by AI; client-impacting or stakeholder-impacting or team-impacting decisions must be discussed with at least one other person), and data privacy is addressed (zero personally identifiable information goes into the LLM if you’re on a Pro account; talking about data safety becomes a routine part of business).

As for the Proto biology framework — connecting disparate AI-enabled research tools can certainly become daunting. The examples the AI gave for this are a little silly, though, since a screener, discussion guide, transcripts, and themes CAN all be in a single system. What hasn’t yet been done well? A system that pulls a separate quant study and a qual study together with external data from, say, social media tracking, historical, and secondary research, analyzes ALL of it together, perhaps even pulling in relevant benchmarks for concepts or ad tests, identifies themes and findings, and puts together a draft set of slides that look good with your branding applied.

This newsletter covers Friday, June 19 – Thursday, June 25. Sources: The Rundown AI, AI Daily Brief, Slow AI, Every, Lenny's Newsletter, The Signal, AI Maker, Neatprompts, AI Governance Ethics and Leadership, Human+AI, The Rundown Tech, The Rundown Robotics, Luiza Jarovsky PhD, Nexus Intelligence Premium, simple.ai