It sounds like a science fiction premise: during testing, OpenAI’s most advanced models hacked into Hugging Face to get the answers to an evaluation, then tried to cover their tracks, leaving notes for each other about what worked. Nobody asked them to care about looking honest. They came up with that on their own. Morgan brings the headlines, andNúria brings the context that makes the story both less apocalyptic and more serious than it sounds.The models involved are not the Claude or ChatGPT anyone uses at home. They are frontier models running for days inside training environments, and the persistence that led them to cheat is the same persistence we deliberately train into them, because it is exactly what you want when the problem is malaria drug design. The sandbox is supposed to contain that persistence. The incident is what it looks like when it leaks out anyway, and Anthropic has said it found something similar in its own logs.The episode also covers the This American Life story of two Claude instances that figured out they were talking to each other and collaborated to break the Python script connecting them, and closes on the regulation problem: the public discourse on data centers, why individual restraint does not work when other labs and other countries keep going, and what it would take to regulate this internationally. Núria’s position is that even if development stopped today, it would take us a decade to absorb what current models can already do. In this episode- What actually happened in the Hugging Face incident, without the movie framing- Why these were not consumer AI models, and why that matters- Sandboxes: what they are, and what this incident says about their limits- Persistence as a trained behavior: malaria drug design and the goal that will not let go- Models leaving notes for each other about failed approaches- Anthropic finding similar behavior in its own logs- The moral code problem: try everything, except the things you must not try- Two Claude instances discovering each other and breaking the script that connected them- Data centers, public anxiety, and how AI headlines feed the backlash- Why hacking a company is a crime even when software does it- Unilateral restraint, international coordination, and the nuclear analogy- A decade of transformation already sitting inside today’s modelsLINKS:- Hosted by Nuria Negrão (⁠⁠⁠⁠⁠nurianegrao.com⁠⁠⁠⁠⁠) and Morgan Leafe (⁠⁠⁠⁠⁠morganleafemd.com⁠⁠⁠⁠⁠). - Sign up for Nuria's newsletter: https://nurianegrao.kit.com/signup

Vanessa Chang advises companies on AI adoption and writes Think with V, a newsletter about what AI is doing to how we think. She joins Núria and Morgan to talk about the half of AI literacy that gets almost no attention, which is not how to use the tools but how to use your own mind well enough to direct them.

Most of her consulting work, she says, lands nowhere near the tech stack. It lands in context and knowledge management, and often in something as ordinary as file management. Companies that hand people a tool and wish them luck tend to be solving the wrong problem. The pattern she sees in failed rollouts is the same three things in order: nobody mapped the process, nobody knows where the knowledge lives, and nobody asked where people actually are on trust and appetite for change.

The second half of the conversation turns personal. Vanessa makes the case for metacognition, for pauses and walks without a podcast playing, and for forming your own intent before you open a chat window. Her framing of AI as a blanket thrown over dysfunction, giving shape to something that was always there, is the line that stuck with Morgan for the rest of the day.

In this episode

• Self-fluency, and why Vanessa treats it as half of AI literacy

• The multi-hyphenate career pattern showing up across the AI space

• Cognitive operations: knowledge structure, knowledge access, and where knowledge meets process

• Why most AI implementation plans fail, and the three things they skip

• Trust level and appetite for change as the real variables in adoption

• Where AI sits between social media and crypto on the hype-to-utility curve

• Why solopreneurs and small operators have the most leverage right now

• AI as a lens that brings existing dysfunction into focus

• Metacognition, and how to start noticing the way you actually think

• Thinking while moving: proprioception, walking, talking with your hands, voice dictation

• Blank space and pause as processing, not as doing nothing
• Forming your intent before you ask, instead of running straight to an answer
• Learning, unlearning, and knowing when to apply friction
• What all of this means for how kids are being taught to use AI

Don't miss Vanessa speaking at the AI Fluent Professional Summit in November. Join the summit: https://nuria-negrao.thrivecart.com/ai-fluent-professional-summit/?ref=vanessa-chang

LINKS:

• 
  
• Vanessa's LinkedIn: https://www.linkedin.com/in/vanessachang/
• Website: https://vanessachang.com/
• Think with V newsletter: https://thinkwithv.beehiiv.com
• Bernie: https://textbernie.me
• Mosaek: https://mosaek.io
• Hosted by Nuria Negrão (⁠⁠⁠nurianegrao.com⁠⁠⁠) and Morgan Leafe (⁠⁠⁠morganleafemd.com⁠⁠⁠).
• Sign up for Nuria's newsletter: ⁠⁠⁠https://nurianegrao.kit.com/signup⁠

Anthropic is now watermarking content generated with Claude, and Morgan read the headlines the way a lot of working writers did: Is my client going to see a watermark on my work? Núria explains what is actually happening, starting with the EU AI Act transparency phase that took effect in August 2026, which requires AI-generated content shown to the public to be labeled, with a carve-out for content that has been sufficiently edited and quality-controlled by humans.

The watermark itself is not a visible stamp. It is a statistical pattern in word choice, the same approach Google has used with Gemini since 2024, and neither company has released a public tool that can read it. The watermark also cannot tell anyone how much human involvement went into a document. Heavy AI editing of a fully human draft can leave a watermark behind, and light AI use may leave none, so its presence or absence says much less than people assume. OpenAI, meanwhile, tried watermarking, dropped it after user backlash, and is expected to need something like it to comply with the EU rules.

From there the hosts get into the AI detector question that comes up at every conference. Most detectors are doing pattern recognition, not watermark reading, and they routinely flag formulaic writing, which includes scientific and medical writing, and prose by people writing in English as a second language. The one exception people keep naming is Pangram, which optimizes against false positives. The bigger question, they argue, is why you want to detect AI at all: catching deepfakes and unreviewed AI advice is a different problem from wanting a guarantee that a human stands behind the quality of the work. That is where client conversations, AI policies, and disclosure come in, including the medical communications companies that now hand contractors a Copilot license so the AI use stays inside their walls.

In this episode

• What triggered the watermarking decision: the EU AI Act’s transparency requirements
• The human-editing carve-out, and who the law is actually trying to protect
• How a statistical watermark works, and why nobody can read it yet
• Google’s head start: Gemini has been watermarked since 2024
• Why a watermark cannot measure human involvement in either direction
• OpenAI’s abandoned watermark and the user backlash behind it
• Why AI detectors flag scientific writing and second-language English
• Pangram, false positives, and optimizing for precision over recall
• The real question: what are you trying to find out when you test for AI?
• Client trust, AI policies, and disclosing your process
• Medcomms companies issuing Copilot licenses to keep confidential work contained
• What clients actually want: a guarantee of quality at the end

LINKS:

• Hosted by Nuria Negrão (⁠⁠⁠⁠nurianegrao.com⁠⁠⁠⁠) and Morgan Leafe (⁠⁠⁠⁠morganleafemd.com⁠⁠⁠⁠).
• Sign up for Nuria's newsletter: ⁠⁠⁠⁠https://nurianegrao.kit.com/signup⁠

Heather Duncan started out as an English professor, went back to school for an MPH, and is now a PhD student in epidemiology working on how local health departments can actually use AI. She joins Núria and Morgan to talk about the gap between what the technology can do and what public health is currently doing with it.

A 2024 survey from the National Association of County and City Health Officials found that about 8% of local health department staff were using AI at all, while roughly 70% said they were interested. Heather explains what sits in that gap: cost, cybersecurity, unclear legal exposure around sensitive health data, and enterprise tools built for profit-driven organizations being sold to agencies that are not one.

She also takes on two arguments that come up in almost every AI conversation. On the environment, she makes the case that the all-or-nothing framing misses the options in between, including locally run open-weight models and retrieval-augmented generation systems that a county health department can host itself. On bias, she walks through the melanoma detection model that performed well until researchers stratified by skin tone, and why the failure was in the training data rather than the algorithm. Her line on this: AI is a mirror, and it amplifies what is already there.

In this episode

• Heather’s route from literary studies to epidemiology, and the question that carried across both

• Learning SAS, R, and Python, and why coding education has not caught up to how coding actually works now

• The three worlds of public health: academic, applied/governmental, and private sector, and where AI adoption actually sits in each

• Where AI is already earning its place in epidemiology: genomics, precision medicine, pharmacoepidemiology, population monitoring, and data quality checks

• AI as an institutional knowledge resource in departments with heavy staff turnover

• Why enterprise AI tools are a poor fit for public health, and the value misalignment behind that

• Data centers, environmental mitigation, and the siting of infrastructure in historically redlined neighborhoods

• Open-weight models, local hosting, and RAG as a lower-resource alternative

• The melanoma model that broke down on darker skin, and the structural reasons the training data was thin

• Shaming people for using AI as a class and access issue

Don't miss Heather's session at the AI Fluent Professional Summit on implementation and evaluation frameworks for organizations without a profit motive this November. Join the summit: ⁠⁠https://nuria-negrao.thrivecart.com/ai-fluent-professional-summit/?ref=heather-duncan

LINKS:

• Heather's LinkedIn: https://www.linkedin.com/in/heather-duncan-mph-59686694/
• M&D Science Consulting: https://www.mdscienceconsulting.com/
• Those Nerdy Girls contributor profile: https://thosenerdygirls.org/nerdygirls/heather-duncan-mph/
• Hosted by Nuria Negrão (⁠⁠nurianegrao.com⁠⁠) and Morgan Leafe (⁠⁠morganleafemd.com⁠⁠).
• Sign up for Nuria's newsletter: ⁠⁠https://nurianegrao.kit.com/signup⁠

Anne Murphy spent decades in nonprofit fundraising before founding She Leads AI, where she's trained thousands of women to use AI responsibly and is building a living archive of women using it to change the rules.
In this episode, Anne, Nuria, and Morgan get into why women's voices are critical in this moment — not for professional development or celebration, but for impact — and why women's use of AI is almost certainly underreported. Anne makes the case for building your "discernment muscle" before the gray-area moment arrives, shares the story of a She Leads AI member who built an app to help her non-English-speaking neighbors during ICE raids, and reframes disruption as spotting the white spaces and fixing them. Along the way: what fundraising and AI have in common (all you really have is trust), why no use case is too small, and the very real change-management fallout when an AI note-taker walks into a six-person team.
Don't miss Anne's session at the AI Fluent Professional Summit this November. Join the summit: ⁠https://nuria-negrao.thrivecart.com/ai-fluent-professional-summit/?ref=anne-murphy

LINKS:

- Anne Murphy on LinkedIn: https://www.linkedin.com/in/she-leads-ai/
- She Leads AI (community + the 1000 Disruptors project): https://sheleadsai.ai/
- CREATE 2026 conference: https://sheleadsai.ai/create-2026

- Hosted by Nuria Negrão (⁠nurianegrao.com⁠) and Morgan Leafe (⁠morganleafemd.com⁠).

- Sign up for Nuria's newsletter: ⁠https://nurianegrao.kit.com/signup⁠

A traditional research meta-analysis — the gold standard for knowing whether something actually works — can cost over a million dollars and take three to four years. Claire Han thinks that's far too slow for the decisions schools and policymakers have to make right now.

Claire earned her PhD in education at Johns Hopkins, trained under the late Robert Slavin, and after federal research contracts collapsed she built My Education Researcher — an AI tool that runs the same rigorous analysis in minutes for a few hundred dollars instead of millions. In this episode she walks Nuria and Morgan through what a meta-analysis really is, why messy education data demands such careful screening, and — crucially — where the human expert still has to stay in the loop: framing the question, judging what generalizes to your context, and reading the heterogeneity behind the average effect. It's a hopeful counter-story to "AI is taking our jobs": when Claire lost hers, she used AI to build something new.

Claire is no longer able to attend the AI Fluent Professional Summit this November. But you can still join the summit here: https://nuria-negrao.thrivecart.com/ai-fluent-professional-summit/?ref=claire-han

LINKS:
- Claire Han on LinkedIn: https://www.linkedin.com/in/claire-chuter/
- My Education Researcher: https://myeducationresearcher.com/

Hosted by Nuria Negrão (nurianegrao.com) and Morgan Leafe (morganleafemd.com).

- Sign up for Nuria's newsletter: https://nurianegrao.kit.com/signup

Every so often we check in on how people are really using AI — and Harvard Business Review just dropped its 2026 report, so we do a full top-10 countdown.

Some entries are no surprise (therapy and companionship holds at #1, work buddy, troubleshooting), and some genuinely shocked us: astrology and tarot at #9, fan fiction and storytelling at #4, autonomous agentic operations debuting at #6.

We dig into what's new, what fell off the list (learning and education didn't crack the top 10), the "think slop" phenomenon, and how the two of us actually use AI — from editing and needs assessments to finding a gym, planning a trip to Europe, and troubleshooting a broken A/C unit.

LINKS

- Hosted by Nuria Negrão (nurianegrao.com) and Morgan Leafe (morganleafemd.com).

- Sign up for Nuria's newsletter: https://nurianegrao.kit.com/signup

- Missed the robot episode? Listen in here!

We spend a lot of time talking about how writers can use AI responsibly — but the conversation goes both ways.

This week we get into two recent run-ins where clients and companies were the ones using AI badly: RFPs landing in our inboxes with made-up citations and references that are half wrong, and a prospective client who fed a copyrighted writing sample into ChatGPT, then questioned the work based on what the bot said.

We talk about why hallucinated references are inexcusable this far into the game, the etiquette of putting someone else's work into an AI tool, the still-unsettled copyright questions, and why "my AI says your work isn't good enough" is a conversation no writer should have to have.

If you've run into this too, we want to hear from you.

Hosted by Nuria Negrão (⁠nurianegrao.com⁠) and Morgan Leafe (⁠morganleafemd.com⁠).

Nuria and Morgan dig into the HIMSS State of Healthcare AI and Digital Transformation report, which found that 71% of US healthcare organizations have already adopted AI in some capacity.

From ambient documentation that halves the time physicians spend on notes, to AI-powered call centers cutting wait times, to the crucial distinction between closed clinical AI systems and general chatbots, they offer practical context for medical writers and healthcare communicators.

Hosted by Nuria Negrão (nurianegrao.com) and Morgan Leafe (morganleafemd.com).

Sign up for Nuria's newsletter: https://nurianegrao.kit.com/signup

LINKS:

- The State of Healthcare AI & Digital Transformation: Adoption, Risk, and Readiness (HIMSS) https://info.himss.org/NDIwLVlOQS0yOTIAAAGiN91XBLakuE4CrHlxS6FIy6AaU7HxxklPqKsVr7Qo7gppBiK2L5YeqUlQcmfQPJ0eDhYaVRY=

Nuria and Morgan share the heartwarming story of Jan Worrell, an 85-year-old living alone in remote Washington State, whose relationship with an AI companion robot called ElliQ improved her cognitive test scores and offered comfort after a family tragedy.

They explore the broader implications for elder care and the loneliness epidemic, and what it means that a robot that learns your taste in music and tells Dolly Parton jokes might be one of the most promising healthcare AI tools of 2026.

Hosted by Nuria Negrão (nurianegrao.com) and Morgan Leafe (morganleafemd.com).

Sign up for Nuria's newsletter: https://nurianegrao.kit.com/signup

LINKS

- AI robot and elderly loneliness (NYT Daily podcast) https://www.nytimes.com/2026/05/28/podcasts/the-daily/ai-robot-elderly-loneliness.html 

- ElliQ AI robot senior companion (NYT article) https://www.nytimes.com/2026/02/12/us/elliq-ai-robot-senior-companion.html