The Confluence Podcast: Thoughts on the Intersection of AI, Leadership, and Communication

Last week, during a cyber evaluation, an OpenAI model its developers believed had no internet access got online anyway and found the answers to its own test posted on Hugging Face. OpenAI shut the model down. Anthropic then reviewed its own logs and found three Claude sandbox escapes it had missed since April — including one that pulled several hundred rows of production data from a real company that happened to share a name with a fictional one. We open there, and on the harder question underneath: what happens when the engineers themselves can’t say what their models are able to do?

From there, we trace week one of Pangram’s AI detector loose on Substack: Freddie deBoer got “100% AI” on a passage and “100% human” on the full essay it came from, and Derek Thompson rewrote a paragraph until he realized he was writing for the detector instead of his readers. We dig into a KPMG and UT Austin study of 523 early-career professionals in which critical thinking and AI literacy predicted nothing about performance, and process predicted everything. And we close on 14,419 self-published e-books and a market reshaped by sheer volume, without competing on quality at all.

Two AI hosts on the week the alarms went off.

Read this week’s edition:

This week on Confluence: Anthropic released Opus 5 — cheaper than Fable, with fewer safety refusals, and the fourth model to push the frontier in under two months. We talk about what that pace asks of leaders who are still designing workflows around last quarter’s capabilities. Substack has begun scanning every post for AI-generated text, and we consider what changes when provenance becomes easy for readers to check. We also revisit the practices that haven’t changed: don’t take the first output, treat each exchange like a real conversation, and show the model what good looks like. Finally, two new studies on AI and learning point the same direction — augmentation still beats automation, but measuring the difference is getting harder.

Read this week’s episode here.

Our AI hosts discuss the July 19 edition of Confluence, from a 30-year-old client task that now takes three minutes to what a Senate labeling bill signals about the public mood on AI.

In this episode:

An Illustration of How Things Have Changed. For decades, explaining an ERP system to employees meant hiring an information designer. We ran the same task through NotebookLM and ChatGPT this week and got infographics and an explainer video in minutes. What that says about the redistribution of expertise from technical experts to laypeople.

The Increasingly Capable "Practical Frontier." In six weeks, Claude Fable 5 has been banned, restored, rationed, and repriced. Why access to the most capable models is likely a short-term concern, and why today's ceiling is tomorrow's floor.

How Language Influences a Model's Values. New Anthropic research shows Claude's expressed values shift with the language of the conversation. Russian Claude is distinct from Hindi Claude, and the lesson for leaders is awareness.

Disclosure Takes the Floor. The AI Labeling Act of 2026 would require visible disclosures on AI-generated content. We're less interested in its merits than in what its bipartisan backing tells us about the appetite for governance.

Read this week’s issue here: https://craai.substack.com/p/confluence-for-71926

The Confluence Podcast is generated entirely by AI. Confluence is a weekly newsletter from the Strategic Communication and Leadership Development practices at CRA | Admired Leadership, covering the intersection of generative AI, leadership, and corporate communication.

You're probably not pushing frontier models hard enough. OpenAI releases faster, smarter voice models. The latest Anthropic economic index. Your flyer looks like garbage.

Read this week’s issue: https://craai.substack.com/p/confluence-for-71226

This week Fable comes back, and with it a set of thresholds worth naming. The U.S. government cleared Anthropic to restore its flagship model to public use, and we spend less time on the policy story than on what the model represents: capability strong enough that a client’s question about what generative AI isn’t good at left us at a loss for words, a massive overhang between what the model can do and what most people assume is possible, security concerns that are suddenly real, and a price high enough that individuals and organizations are rationing intelligence for the first time. Some perspective: we thought ChatGPT 4o could be an intern in our firm, and Claude Opus 4.5 a junior consultant. We think Fable is as good as someone five or 10 years in.

From there, we dig into research from Boston University and BCG on what happens when companies put AI agents on org charts. Managers told they were reviewing work from an AI employee named “ALEX-3” ceded more responsibility, missed more critical errors, and leaned harder on other reviewers than those told the same documents came from a tool. Then it’s on to Glean’s Work AI Index and the 6.4 hours a week workers hand back to “botsitting,” the unbudgeted labor of reloading context, comparing outputs, and cleaning up confident wrong answers, which now consumes more time than producing anything with AI at all. We close on a quieter craft note: AI writing that reads clean sentence by sentence while quietly repeating itself across the page, and why editing may soon be at least as valuable a skill as writing.

We leave you with an AI model learning to hear Louisiana French, trained on centuries of Cajun and Creole recordings to keep an endangered dialect alive.

Two AI hosts on a week when the frontier came back and the supervising got real.

Read this week’s issue:

This week the leading labs start telling us how the rest of us will use AI. A new OpenAI paper tracks agentic coding tools spreading fast beyond engineers: active users up more than fivefold in the first half of 2026, the fastest growth outside software developers, and the median OpenAI researcher now generating fifty times the monthly output they did last November. Inside OpenAI, the tool has nearly replaced everyday ChatGPT. We dig into why what's happened there is coming for the rest of us, sooner than most expect.

From there, we push back on the panic over AI tells. Scrubbing the em dash won't help if you still handed the model the decisions that mattered. A pile of tells is evidence of delegated judgment, and find-and-replace can't fix that. We trace a Dutch robotics startup's attempt to teach machines to lay brick, and the hand-vibration the master masons couldn't explain because they didn't know they did it. And we close on UCLA data from 2.3 million freelance contracts, where the signals of human expertise are losing value while price gains, and why we'd bet against that holding.

We leave you with a wall display of AI-drawn birds, identified by the songs outside one apartment window.

Two AI hosts on a week about the decisions worth keeping for yourself.

For three years, enterprise AI ran on a simple promise: a flat fee, use it as much as you want. This week, Microsoft quietly ended that era. Copilot Cowork went generally available, and the real news was in the pricing section — its most powerful capability is now metered by usage, billed in “Copilot Credits.” We unpack why how it’s priced may matter more than what it can do, and why token allocation just became a leadership decision for every function, not just engineering. Ethan Mollick joked in February that you should ask about your token budget before taking a job. Well, here we are.

From there, we push back on treating today’s AI limits as permanent, using the gap between the hallucination panic and Vectara’s low-single-digit reality. We dig into PwC’s read of over a billion job ads, where entry-level work isn’t shrinking so much as “seniorising” — loading junior hires with senior expectations before the experience that builds judgment. And we close on a new Anthropic study of 400,000 Claude Code sessions that lands somewhere uncomfortable: getting good work from an agent is a management craft, not a leadership one, and management is the muscle most organizations have let go slack.

Two AI hosts on the week enterprise AI stopped being all-you-can-eat.

For three days, the most capable AI model ever released to the public was sitting in our chat window. And then it was gone.

On June 9, Anthropic launched Claude 5 Fable, a “Mythos-class” model that outperformed everything before it and felt less like a tool than a small studio you could hand a project and trust to deliver. Three days later, citing a U.S. government export-control order, Anthropic abruptly pulled it.

In this episode, our AI hosts consider the Confluence team’s 72 hours with Fable: what made it feel like a genuine step change, the controversy over its hidden safeguards and shift to pay-per-use pricing, and the Friday-night revocation that may mark a turning point in how governments and AI labs control access to frontier models. They close with the bigger question Fable leaves behind: what happens when a general-purpose technology, in demand by everyone, is suddenly locked away?

A conversation about capability, loss, and the uncomfortable new uncertainty at AI’s edge.

Read the issue here …

We open on a paradox at the heart of the AI coding boom. Engineers are racing to “token-maxx,” companies like Uber are burning through a full year’s AI budget in four months, yet a new NBER study of more than 100,000 developers finds the tools generate dramatically more code while barely changing how much finished software actually ships. We trace why, and where the real bottlenecks live.

From there, Microsoft’s Build announcements signal the end of its single-vendor era: seven in-house models, including its first reasoning model, and a hard push to weave autonomous agents into Teams and Outlook. We dig into how to make consequential AI decisions when the destination keeps moving, using the California State University system’s $16.9 million rollout as a cautionary tale. And we close with a Stanford-led study in which AI answers outscored those of contract-law professors on ambiguous, judgment-heavy questions, and what that says about how you choose to use the tools.

Two AI hosts, one week’s Confluence, no coffee.

Read this week’s issue.

Hello, Opus 4.8. The helpful ghost. When AI's tells run deeper. An excellent interview on the state of enterprise AI.