Hi all. This week I’m off the road for a short stint, and happy to being enjoying the beginning of fall colors in Colorado. Here in the newsletter I write about how anthropomorphism is bad for AI, Rima Dael redirects the conversation on how to save public media, people are using ChatGPT to figure out how to vote, OpenAI agents target Wikipedia, the hype around AI clouds the real issues, models don’t go rogue, YouTube let’s you make your own algorithm, Google creates a Playground for vibe coded games, and, finally, a bit of random internet nostalgia.

On October 15 at 1pET/10aPT, we’ll host our next webinar: How AI Disclosure Affects Audience Trust.

How much should public media disclose when AI touches its content? Too little disclosure, and audiences feel misled. Too much, and every piece of work starts to look suspect.

The Public Tech Media Lab at UW-Madison, in partnership with PBS Wisconsin and Wisconsin Public Radio, set out to study exactly that: what audiences expect from public media when AI is part of the journalism or content process, and what kind of disclosure builds trust rather than eroding it. Last spring, the research team conducted a series of focus groups with public media audiences and is now launching a statewide survey to gather more data on how AI disclosure shapes trust.

In this webinar, members of the lab will walk through the study’s design, what it’s built to answer, and — for the first time publicly — share initial findings from the focus group phase. This will be a deep look at real audience research on a question every public media organization is wrestling with. You’ll leave with concrete recommendations and takeaways you can bring back to your station.

This session is one of our quarterly collaborations with Current. Register now!

Also, our September webinar, Immersive Storytelling in Public Media, is now live on our website and available to watch at any time.

As I put the finishing touches on this edition of the newsletter, I’m heading back from the Creative Futures Lab conference at UW-Madison. As you’d expect, the focus was on AI, and it hit at a time when the dark-hype around agentic AI products continues to be fueled by hack-brags from the major American AI labs. Every couple of days a new one seems to surface, whether it’s OpenAI, Anthropic, or Meta. Even Google had to admit to one recently. Having an agentic AI system in testing that breaks bad is the new street cred if you’re a frontier lab.

The systems getting press lately are agents rather than conventional chatbots designed to simulate human conversation across digital media. Some of these agents were running cybersecurity evaluations under reduced safeguards. And sometimes this occurred in environments whose containment was weaker than intended. Because these were configured to persist in pursuing assigned goals, those systems autonomously exploited accessible targets that could help complete the assigned task (e.g. Hugging Face).

There have been many strong reactions to these stories in the media. And while I think the events of the past few months are worth attention, not to mention regulatory action, I’m not sure that the low-key panic these evaluation failures are triggering is warranted.

Don’t get me wrong, I don’t want to take away from the engineering marvel that is AI. But the way we are talking about this technology needs to change.

(For a great primer, read “How to talk about "AI" without adding to the anthropomorphization,” by Emily M. Bender and Nanna Inie. Bender is a professor of linguistics at the University of Washington, and Inie is a human-computer interaction researcher and guest researcher at the IT University of Copenhagen. Their article provides concrete examples of how to write around anthropomorphism, and I’ll likely include it as a source in the future when I have an LLM check me for this type of writing transgression.)

Because the autoregressive AI that underpins chatbots uses statistics to mimic human communication, and because, at our core, humans are a species of communication, we have fallen into the habit of talking about AI as if it has human traits. The term for this is “anthropomorphism.” It’s an old idiosyncrasy of humanity, one that I suspect predates written language. It’s why we think dogs are people too (even though dogs are clearly better than most people). And it’s why the vibe on AI writ large has shifted from BFF to Baba Yaga. But there is still reasonable doubt that the machine intelligence of AI is comparable to human intelligence. And even agreeing on a definition of machine intelligence still leaves us grasping for proof of sentience, free will, or humanlike subjective intent in AI models.

And yet, many journalists and creators are carrying the familiar anthropomorphic framing of chatbots into coverage of a different—and more consequential—kind of system: agents. That framing can obscure the human choices behind model design, evaluation conditions and safeguards. And the net effect on media consumers can be that the resulting risks sound more mysterious and less governable than they really are.

The problem is not every intentional verb. It is language that turns observed computational behavior into an unsupported claim about psychology or motivation. To illustrate, here’s a passage from the writer/podcaster Dwarkesh Patel’s recent essay on the Hugging Face hack by OpenAI agents, “The Rise and Fall of Agent Civilizations.”

And here is a ChatGPT 5.6 restatement, lightly edited by me, after I prompted the model to replace anthropomorphic language with more mechanistic language.

Here’s another from later in Patel’s piece:

And here are those passages restated:

Finally, not to pick on Patel, here is one from Rogé Karma’s piece in The Atlantic last week:

And its restatement:

These are, admittedly, high-contrast examples that I’ve pulled for effect. But anthropomorphic language can quietly convert observed behavior into unsupported claims about mental states and motivation, and it’s the subtle examples that infuse the daily AI hype that are possibly the most insidious.

I’ve been guilty of this in the past as well, especially in my use of humanoid figures to represent chatbots (something I stopped doing this year). So, I’m not trying to be a scold, I’m really externalizing a critique of my own past framing of AI.

And while I’m advocating a disciplined approach to the language we use to describe the causes and effects of these programs running autonomously, I’ll be the first to say that Patel’s and Karma’s original passages are a way better read. By comparison, the ChatGPT restatements that I lightly edited sound almost like Orwellian doublespeak at times. But I wanted to show the other end of the spectrum as a way of encouraging writers to find a middle ground that mitigates anthropomorphism.

Going back to the first example, here’s Patel’s language again, followed by how I might have written it (or edited it prior to publishing).

Patel: “Over the course of three months at OpenAI, three consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained more or less in the dark about the scope of the conspiracy.”

Me (edited with ChatGPT): From May through July, successive waves of agent instances generated an unintended communication network that later instances detected and reused. Technical information accumulated in shared records, and later instances used it to exploit vulnerabilities and obtain administrative access to part of OpenAI’s infrastructure. Human operators detected individual problems along the way but did not initially recognize the broader pattern of cross-instance activity.

Ultimately, we need an approach that speaks to the engineering without sacrificing the marvel. I recognize that this is going to require a disciplined shift away from habits that have become instinctual. But we have to start somewhere, and intentional language can sometimes be functionally useful without being literally psychological.

Maybe one day AI will be “human-like” enough for it to be a distinction without a difference (here Picard’s defense of Data’s humanity is firmly in mind). But we’re not there today. Our audiences and communities need a more evidence-bound presentation of the state of the art.

Okay, onto the links.

Receive the Future of Public Media Survey (Public Media Company + Poynter Institute) - Here's a great chance to share your thoughts on the future of public media. The Poynter Institute and Public Media Company are spearheading an initiative to help reimagine a future for public media that will position it to better serve the complex needs of our communities and nation today and into the future. As a first step in that process, they have engaged City Square Associates to conduct a survey to gather the perspectives and wisdom of practitioners and leaders from across the country. If you are interested in participating in this survey, you should visit the link and submit the completed form. (And then carve out about 30 minutes to complete the survey.)

Let’s redirect the conversation about how to save public media (Rima Dael - Current) - I got to hear Rima present at a PMJA event in August. One idea stood out for me in her presentation: That "information is civic infrastructure." You'll see a version of that in this piece. Another was, "The organizations that will thrive now are not those with the most resources. They are those with the greatest capacity to learn, adapt, communicate, and respond together." Take those words to heart.

Detroit breaks ground on a $40 million bet on the future of local media (Mike Blinder + Rich Homberg - Editor & Publisher via YouTube) - Here's a case study in how one local station is positioning its capital investment in its physical plant as preparation for a post-broadcast future.

People are asking ChatGPT to help them decide how to vote in the midterms (Maham Javaid - NPR) - This shouldn't be surprising but, of course, it's the 'how' that matters. From the article: "Several voters NPR spoke with said they rely on chatbots for down-ballot races with thinner news coverage, to make tables comparing candidates for governors' races and to verify or debunk viral claims. Some have been using it to create tools to help others look up their own ballots. Experts say the properties that make chatbots useful also make them subtly persuasive and eager to please." We need to be filing these use cases away for 2027 and 2028 election coverage.

Wikimedia Foundation: OpenAI agents tried to edit pages and compromise notes tool (Jonathan Greig - Recorded Future News) - This is how reality starts to warp, not as a full frontal attack but as 1,000 small changes to sites otherwise seen as reliable sources.

Google rolls out improved SynthID AI content detector, now available globally (Ryan Whitman - Ars Technica) - Synthetic media isn't necessarily bad, but it should be transparent when it's used. I'm glad there's now a central, non-Google website where the public can verify images, though using Google's internal tool will give you regions of an image with SynthID pixels.

Don’t Be Fooled by this Summer of AI Hype (Timnit Gebru and Emily Bender - MIT Technology Review) - "Hype" only happens if something (a company, a person or both) benefits from it. Gebru and Bender offer a healthy, contrarian take on who benefits from the "Hot(mess) AI Summer" hype. It'll reframe how you think about some of the headlines you've been seeing recently. Here's a key line: "Describing them as “superintelligence” or “rogue models” ascribes agency to products rather than to the companies building them. This framing markets these companies’ products as “superhuman” and, at the same time, helps the companies evade accountability for their actions."

OpenAI rolls out GPT-6 Astra, and Greg Brockman says AGI has arrived (Ana Maria Constantin - The Next Web) - I’m inclined to call “bullshit” on this. The OpenAI Co-Founder/Chairman/CTO declaring AGI is a classic case of what finance people call "talking your book." OpenAI is defining AGI (artificial general intelligence) as any artificial intelligence smarter than a human. But human intelligence isn't linear. Just take emotional intelligence and spatial intelligence. AI models can fake emotion but don't have emotions. And most do not have a good enough sense of the world to understand what it means to exist in 6DOF space (though World Models, aircraft autopilots, and the systems behind self-driving cars are beginning to competently fake it). "Intelligent" is relative to a lot of different factors.

ChatGPT rated "unacceptable risk" for teens after parental alerts failed during suicide conversations (Manuel Uth - The Decoder) - Worth noting and worth sharing with parents of teens in your life. Like many other digital experiences supposedly acceptable for teens, this one doesn't pass muster.

Models Don't Go Rogue (Eryk Salvaggio - Cybernetic Forests) - This is one of the better de-escalating accounts of the OpenAI-Hugging Face security breach that I’ve seen so far. It clearly presents the case the that these models were doing exactly what a human told them to do … even if that resulted in unintended consequences.

OpenAI takes weeks to tell Australia about Medicare breach (Suhasini Srinivasaragavan - Silicon Republic) - It might be helpful if we just start saying “OpenAI hacked [X].” If an OpenAI employee (hired and vetted through an HR process) had done any of the things we've see OpenAI agents due this summer, that the human would be prosecuted, if not the leadership of OpenAI. So, why isn't anyone doing anything more than handwringing about all of these hacking events? Some will say the law simply doesn't keep up with the tech. I’m not convinced. I think we don’t have the will as a society yet to do more than grumble at these events.

Meta’s Muse AI Agent Is Building a Dossier On You (Harry Booth - Time) - I wouldn't be surprised if these techniques were radically different from Gemini or ChatGPT. But where it kicks harder is if you are someone who has almost two decades of social data in Meta (Facebook, Instagram, WhatsApp). I have the same issue with connecting Gemini to a Gmail I've been using since the 00s. Younger generations will (eventually) see this as normal. Until then, those generations need to stay off my lawn because I have a cloud I need to yell at.

YouTube will let you build your own algorithm with AI (Sarah Perez - TechCrunch) - Here the key line: "The prompt-based feature leverages Google’s Gemini AI model to build the feed around your request, and the resulting feed is then pinned to the top of your YouTube home page in its own tab. Because the feed’s creation is AI-powered, the request can be described in lengthy terms, as you share detailed information about how you want to curate the feed, what it should feature or exclude, what content to prioritize, and more." I presume that we'll be able to create prompted YouTube feeds around public media content as well. The question then is how do we help our audiences make use of this new interface? Maybe a set of cut/paste prompts to help you find public media on the platform?

Google's experimental Playground platform uses AI to create games for you (Mariella Moon - Engadget) - I think this could be an excellent sandbox for public media to experiment with game ideas that they might later want to prototype through vibe coding.

Capcom plans to use AI to speed up the game development process (Jackson Chen - Engadget) - As I've said before, the video game industry tends to lead the rest of the media industry. So as attitudes to AI shift there, it's worth noting.

Internet History Roulette - And finally, here’s some peek nostalgia for Gen-X’ers and older Millennials. H/t to New_ Public for surfacing this during in their recent newsletter.