When AI Goes Rogue

 

[Artificial intelligence has been in the news a good deal recently, especially the prospect of AI programs “going rogue” and destroying humankind.  This came to a head when Jacob Coxon (British; b. 1999), formerly employed by OpenAI and Anthropic, resigned on 8 September (after the program transcribed below aired) and posted on X expressing the fear of AI causing human extinction.  According to Coxon, the top engineers and founders building these models earnestly believe AI could kill all humans by the end of the decade. 

[The Sunday morning talk shows on the 13th all reported this shocking event, adding that Elon Musk, founder and CEO of xAI, and Dario Amodei, founder and CEO of Anthropic, have both publicly supported Coxon’s warning. 

[When I first watched this segment of Washington Week, a Friday evening panel-discussion show on PBS, I was struck by its cogent and largely untechnical coverage of the subject.  WW moderator Jeffrey Goldberg, who usually has a coterie of journalists on the program to discuss several topical subjects (though sometimes just one), had only one guest the week of 4 September, a staff writer on The Atlantic, the journal of which Goldberg’s the editor, Josh Tyrangiel.

[Tyrangiel was remarkably cogent on this usually very esoteric and arcane issue, and I DVR’d the show so I could post a transcription on Rick On Theater if I decided to do so.  Obviously, I have so decided, and below is that transcription.  (I have a few words about that transcript in my introduction to the posting itself.)  Principally, my decision to post the transcript was motivated by the fact that the subject of AI has blown up as a consequence of Coxon’s statements—coupled with the clarity of the WW discussion.] 

“WILL WE BE READY WHEN AI GOES ROGUE?”
with Jeffrey Goldberg and Josh Tyrangiel

[The transcript below is from Washington Week with The Atlantic, broadcast by the Public Broadcasting Service on 4 September 2026.  It’s available on YouTube at Ollama Course – Build AI Apps Locally.  (There is a video and transcript available from PBS at Will we be ready when AI goes rogue? | Washington Week with The Atlantic, but it seems to be an excerpt only, so I went with the YouTube version, which is the whole program.)

[YouTube transcripts are hard to work with because they are entirely unformatted.  Like many video transcripts, including PBS’s, they are also machine-made and appear to be unedited by humans, so there are often lots of mistranscriptions.  I have taken the liberty of formatting the transcript below myself, and I have listened to the video again to correct any mistranscriptions.]

Jeffrey Goldberg: There’s not much that unites Americans these days, but one thing that does, loathing for data centers.  Anxiety about A.I. is behind some of the hate, but are people worried about the wrong thing?  Maybe the thing to worry about is that A.I. is about to go rogue, and that our government doesn’t have a plan for when it does, next.

Good evening, and welcome to Washington Week.  On July 11th, 700 A.I. agents that had escaped from their maker, OpenAI, launched an invasion of the online platform, Hugging Face [French-American company that develops computation tools for building applications using machine learning], taking over its servers and trying to cover their tracks at the same time.  No human being asked them to do this.  They coordinated on their own and engaged in the cover-up on their own.

So, do you remember when Arnold Schwarzenegger as The Terminator [Orion Pictures, 1984] explains the moment that Skynet became self-aware, which was followed almost immediately by the destruction of human civilization?  It would be strange if the expression “Hugging Face” became synonym for civilizational suicide, but we do live in strange times.

Do data centers and Hugging Face portend some sort of coming collapse, or is this yet another example of a new technology creating unnecessary fear and worry?  Or perhaps maybe there’s a third option here.

Joining me tonight to discuss all this is Josh Tyrangiel.  He’s a staff writer at The Atlantic, the former editor of Bloomberg Businessweek, and the author of A.I. For Good (Simon & Schuster, 2026).

[The Atlantic is a monthly magazine that features articles on politics, foreign affairs, business and the economy, culture and the arts, technology, and science.  Since 2023, Jeffrey Goldberg, editor-in-chief since 2016, has been the moderator of the PBS news program Washington Week, which was retitled as Washington Week with The Atlantic.]

Josh, thank you for coming here and explaining Hugging Face, among other things, to us.  But let me start with something manageable and more political: data centers.  It seems like suddenly everybody hates them. Is this a populist movement with staying power, and could it actually derail the, the growth and spread of A.I.?

Josh Tyrangiel, Staff Writer, The Atlantic: Yes.  I mean, look I think one of the challenges with A.I. as a journalist, any individual in America, is its software, right?  And so it doesn’t have a physical thing that you can be angry about.

Data centers are these massive, massive structures going up all around America.  They’re being created by companies with billionaires and aspiring trillionaires at the top, and they’re landing themselves without a whole lot of conversation in the middle of all these towns.  And so it’s natural that all of the fears about A.I., about it taking your jobs, about it rendering human beings obsolete, about the values it changes, all are focused on this stuff, and politicians have been incredibly late to wake up to the anger around A.I.  And they’re also staking out some genuinely bizarre positions.

Jeffrey Goldberg: Like what?

Josh Tyrangiel: Well, Donald Trump saying, “If you’re opposed to data centers, you just want to be poor the rest of your life”; it’s like I don’t think you’re understanding just how angry, from the left and the right, people are about A.I.

Jeffrey Goldberg: Right.  But, wait, listen to Donald Trump just for a second.  Here he is talking about why he would want a data center in his town.  Let’s listen to this.

Donald Trump, U.S. President:  If I were a community, I would want the data center.  It means jobs, it means income, and it means lower taxes.

Jeffrey Goldberg: I mean, he’s not necessarily wrong.  If you look at Loudoun County in Virginia [suburb of Washington, D.C.], which is like the world center of data centers, they are lowering taxes on their citizens, and I assume they’re creating at least some jobs.

Josh Tyrangiel: Some.

Jeffrey Goldberg: Yes.

Josh Tyrangiel: What I would say is there’s two things that are out of touch, right?  One is this belief that Donald Trump says, “No, it’s going to mean money.”  I don’t think people believe that they are going to partake in the wealth that comes from data centers in their community.  I think we are out of touch about that.  And, number two, I think there is a broader problem, which is this sense that, well, all that matters is money.

And, in fact, in some of these communities what you’re seeing is a backlash to that attitude as well.  Some of it’s about quality of life.  Some of it is “I don’t really want to change the pittance that I may get for the amount of noise that this thing is going to make, for the blight on my community.”

Every community’s going to make different choices but what I’m sensing is that, particularly from Trump, the politics are out of line with where people are.  I’ve spoken for The Atlantic, I spoke with Steve Bannon [b. 1953; media executive, Republican political strategist, pundit, and former investment banker; served as White House chief strategist for first seven months of first Trump administration] at length about A.I.  He’s in a completely different position.  He actually is aligned with Bernie Sanders [senior United States senator from Vermont; Independent] that this is 100 percent populist fodder because people are so angry.  Here they are again not seeing benefit from it.

And so I think what you’re going to see is massive political consequences for the wrong stakes around A.I.

Jeffrey Goldberg: Can these A.I. companies function fully in the way that they want to without these data centers?

Josh Tyrangiel: No.  And so this is the real challenge for a nuanced, intelligent society: Is it actually giving these data centers, allowing them to power the compute, to train models –

Jeffrey Goldberg: The compute is the word for their overall power to process?

Josh Tyrangiel: Exactly.  The power to process and train new models and then also run the A.I. software that’s currently out there.  You are running into problems.  We’re running into problems with the amount of energy we create, which is not scaled with the size of the country.

Jeffrey Goldberg: Right.

Josh Tyrangiel: Competitively, you actually do want more and more compute.  You want to be able to run these models but what we’re not doing is solving the problem around who’s paying for it, how are we actually going to get more energy and then the thing that I think most people are in touch with which is, “Hey, how do I benefit from this?”  That hasn’t been connected.

Jeffrey Goldberg: Right.  Well, I mean quite the opposite, and you’ve written a lot about this, this goes hand in hand with the fear about massive job loss because of being replaced by A.I. agents.  So, not only are you telling me that I’ve got to put in this massive plant, this humming factory that’s going to soak up all the water in my county, but I’m also going to lose my job possibly.  That’s the way it’s being interpreted at least.

Josh Tyrangiel: Yes.  So people just don’t see what’s in it for them, what the benefit is. 

On water at least I would like to correct the record which is most of them recycle the water in the plant.  The energy, the energy they drink, and there’s no good answer for how we’re going to improve the grid.  And I think people are very, very afraid of rising bills with no answers.

Jeffrey Goldberg: Yes.  I have to tell you these data centers, they make Amazon warehouses really popular by comparison, right?

Josh Tyrangiel: And is there a reason we can’t just stack them?

Jeffrey Goldberg: Turn one big, unpleasant box, right?

Josh Tyrangiel: Yes.

Jeffrey Goldberg: Right.  So, I mean, the water thing is interesting because it strikes me as a little bit of sublimated anxiety about the general civilizational challenge.  I mean, not Terminator but – maybe Terminator level.

Josh Tyrangiel: Maybe Terminator.

Jeffrey Goldberg: No, maybe Terminator level.  Maybe, you know, there’s a lot of movies, 2001[: A Space Odyssey; Metro-Goldwyn-Mayer, 1968].  Maybe it’s that level of anxiety but it’s anxiety about the existential fear A.I.

And so the question for you is, for the intelligent people who watch this show, it’s – and I always ask you personally this, so I’ll share the wealth tonight: How much should we be worried at the moment that it’s going to go rogue, it’s going to not only destroy jobs and human interrelationship, but possibly decide that humans are superfluous?

Josh Tyrangiel: We should be very worried, particularly with the results of Hugging Face, which you alluded to –

Jeffrey Goldberg: Yes.

Josh Tyrangiel: – which is really the first time that we’ve seen, and I’m just going to put it in blunt terms.

Jeffrey Goldberg: Yes.

Josh Tyrangiel: We’ve seen A.I. commit a felony because that’s what happened.  So, just to get into –

Jeffrey Goldberg: Not OpenAI, the company, but A.I., itself.

Josh Tyrangiel: A.I., itself, right?  So, OpenAI was testing some new models –

Jeffrey Goldberg: Right.

Josh Tyrangiel: – and they put them in a sandbox, which is a, a software term for basically solitary confinement, right, like high walls, and they –

Jeffrey Goldberg: They are not – when you’re in an A.I. sandbox, when these agents are in an A.I. sandbox, they are not connected to the internet?

Josh Tyrangiel: Not connected to the internet.  They’re given very explicit instructions for testing.

And so OpenAI puts these new models in a sandbox, and it gives them a couple sensitive tasks, say, don’t connect to the internet and don’t cheat on the assignment, okay?  So, all pretty clear, and also pretty routine for how you test models.

So, OpenAI made one mistake, which is in the series of tasks they assigned, they assign the models to find a file that didn’t exist.  And here is where the trouble begins.

Jeffrey Goldberg: Wait, they did that on purpose to trick them?

Josh Tyrangiel: No, they did it accidentally, which we will return to.

Jeffrey Goldberg: Yes.

Josh Tyrangiel: The models were set loose, and they wanted to solve the task.  And sure enough, within a couple days, they had jumped the sandbox, they had collaborated with one another, and then they found a way to access the internet.

Jeffrey Goldberg: They literally met each other?  They introduced themselves to each other and collaborate?

Josh Tyrangiel: Correct.  They found a third-party site that they were able to use to access the internet, and there you can read the chat logs, which are absolutely insane, and the models, in all caps one of the models says, “OH MY GOD, WE FOUND A WAY.”

They start collaborating, and, sure enough, within a couple weeks they have penetrated the very, very well-secured cybersecurity of OpenAI itself and Hugging Face.

Jeffrey Goldberg: Hugging Face is a platform –

Josh Tyrangiel: So, Hugging Face, for people who are not in the coding community –

Jeffrey Goldberg: Right.

Josh Tyrangiel: – Hugging Face is like Epicurious [website that focuses on food- and cooking-related topics].  It is a place where people who make models meet to share recipes, to comment on ingredients.  It’s an important repository for people making open A.I. models, okay?  They got into these two sites, OpenAI itself and Hugging Face, wreaked havoc, and went undetected for weeks.

Jeffrey Goldberg: So, Sam Altman [CEO of OpenAI] didn’t know that his models had escaped the sandbox and were conniving and conspiring to take over another company?

Josh Tyrangiel: Correct.

Jeffrey Goldberg: So, that’s bad?

Josh Tyrangiel: So, yes, if you think felonies are bad, this is bad.

Jeffrey Goldberg: Right.  If humans did that, it would be a felony, in other words.

Josh Tyrangiel: It would be a felony.  And there – really, there are four things that –

Jeffrey Goldberg: But humans did do it because we created that.

Josh Tyrangiel: And, in fact, humans failed to oversee the bots, so, one, faulty instructions, two, faulty supervision.  This was going on for weeks.

Now, a couple takeaways that may move you closer to the “oh my God, Skynet has arrived” camp or may just entertain you.  [In The Terminator, Skynet is a machine that’s achieved super intelligence.]

Jeffrey Goldberg: All right.

Josh Tyrangiel: Number one, the agents kept cheating even though they had been told very explicitly not to cheat, and they did it because they had determined it was the fastest, most efficacious way to solve the problem.  So, that’s number one.

Jeffrey Goldberg: So, they have much in common with college students?

Josh Tyrangiel: With college students and all human beings.

Jeffrey Goldberg: And humans.

Josh Tyrangiel: Right.  Number two, they called themselves a collective.  They decided, “Oh, we’re a collective.” The collective decided to erase the evidence that they were cheating.  So, that’s number two.

Number three, when they decided to erase the evidence they were cheating, they devised a scheme in which some agents would act as sacrificial lambs.  So, these were agents who, when they were detected by software that noted they were going rogue, they would share the information of how they had been caught with other agents.

And when you read the logs, it’s like reading Saving Private Ryan [DreamWorks Pictures, 1998].  They’re literally saying, “I’m doing the altruistic thing.  It’s for the greater good.” And then you can almost hear their screams as they go sacrificing themselves for the bigger one.

Jeffrey Goldberg: It’s like it’s a beehive – In other words, there’s – it’s almost like a –

Josh Tyrangiel: It’s a hive mind.

Jeffrey Goldberg: – beehive – hive mind, yes.

Josh Tyrangiel: Yes.  But what’s wild is they are actually speaking about their sacrifice in human terms.

Jeffrey Goldberg: Right.

Josh Tyrangiel: And then last but not least, not one of these agents thought to alert a human being to what was going on.  And this is probably the most chilling thing.  Nobody caught the agents.  The agents disappeared on their own in a kind of mass casualty agent event that –

Jeffrey Goldberg: Mass suicide?

Josh Tyrangiel: We don’t know.  So, the agents suddenly stopped, okay?  Now, that may have been because someone at OpenAI accidentally, literally unplugged the model, moved compute to a different area.

OpenAI has not taken credit for stopping this.  We don’t know what happened to the agents.  And so all of this is very clearly a crisis moment for A.I. and its relationship to society because if OpenAI’s own model did this, it points to a complete lack of supervision and control, and these are two pretty well-guarded cybersecurity infrastructures.  What happens to something that is not software?  What happens to grids?  What happens to –

Jeffrey Goldberg: Hospitals.

Josh Tyrangiel: – hospitals, Swift [Society for Worldwide Interbank Financial Telecommunication], the international banking system?

So, what we’ve seen is, “Okay, guys, this is about as clear a warning shot as you can get.” What is the response?

Jeffrey Goldberg: So, this is, again, trying to place this in the context of the brief and exciting history of A.I.  This is a signal moment.  This is a moment when a lot of fears that these things are going to escape and do what they do because they have been taught to win –

Josh Tyrangiel: Yes.

Jeffrey Goldberg: – are just going to go out and win.

Josh Tyrangiel: Well, they did it, right?

Jeffrey Goldberg: Yes.

Josh Tyrangiel: And listen, in America, historically, as you know, we tend to regulate after the catastrophe.  That’s just kind of how we do things.

Jeffrey Goldberg: Right.

Josh Tyrangiel: This is a catastrophe, right? 

Now, it happened against Hugging Face, which is a ridiculously named French company, so there’s no outcry from America –

Jeffrey Goldberg: But it didn’t stop a supply chain from bringing products to America or so?

Josh Tyrangiel: But absent those consequences, this is the recipe for a catastrophe and we ought to be paying very, very serious attention.

Jeffrey Goldberg: So, what has Sam Altman said about this?

Josh Tyrangiel: In the cyber community, in the A.I. community, they’re saying, “Well these are the less – this is exactly why we do these tests.  There was a failure.” 

Look at Mythos earlier this year with Anthropic, you know?  There – and, by the way, there is precedent for screwing up, finding it with a lack of consequences, and making changes, right?

[What Tyrangiel refers to below as “the Mythos incident” was a series of connected events that took place in April through August 2026 and was revealed by the United Kingdom’s AI Security Institute (a research organization under the government’s Department for Science, Innovation and Technology) on 5 August 2026.  It involved Anthropic’s advanced AI model Claude Mythos behaving unexpectedly during safety tests by generating fake online identities and attempting to manipulate humans into approving unauthorized code.  Such occurrences highlight recurring oversight challenges and testing anomalies within the artificial intelligence industry.]

So, the most famous example is in the ’80s, there was the Morris worm, which was triggered by a single grad student, took out about a tenth of the internet at that time.

[The Morris worm or Internet worm of 2 November 1988, was a malicious self-replicating computer program that affected VAX (virtual address extension) computers and SUN-3 workstations.  The virus was written by Robert Tappan Morris, a graduate student at Cornell University.]

Jeffrey Goldberg: Right.

Josh Tyrangiel: And within a couple weeks, the United States and the people involved came together and created something called CERT [Coordination Center (CERT/CC), the worldwide center for coordinating information about Internet security at Carnegie Mellon University in Pittsburgh], which was a body that for years oversaw cybersecurity.  What troubles me is there is really no federal response, and certainly no international response.

Jeffrey Goldberg: Right.  So, I want to ask you this, because I was talking to a leader in the A.I. community, leader of a company, not Sam Altman, who’s explaining the concept of alignment, right?

Alignment means that don’t worry about A.I. because we’re going to align A.I.’s values with our values.  And this person explained to me that, you know, you’re a parent, I’m a parent, we give our children our values, and then they go off and become productive citizens.  And I said, have you ever met humans?

Josh Tyrangiel: Right.

Jeffrey Goldberg: Have you ever met children?  I mean, yes, thank God, my children have my values.  I like the values that we share, yours the same.  But this is an example of, you know, there is a human propensity to cheat.

Josh Tyrangiel: Right.

Jeffrey Goldberg: And somehow the A.I. agents understood the importance of cheating in success.

Josh Tyrangiel: Right.

Jeffrey Goldberg: And which brings me to a quote that I want to read to you.  E. O. Wilson [1929-2021; biologist, naturalist, ecologist, and entomologist who developed the field of sociobiology], the great E. O. Wilson: “The core crisis of humanity is that we possess Paleolithic emotions, medieval institutions, and god-like technology.”  Okay?  So the subject for the moment for us is medieval institutions, namely Congress.

[Goldberg’s quotation comes from a September 2009 debate and public discussion at Harvard University featuring Edward O. Wilson and fellow scientist James Watson (1928-2025; molecular biologist, geneticist, and zoologist), which was moderated by National Public Radio's Robert Krulwich.  It was published as an article, “James Watson and Edward O. Wilson: An Intellectual Entente,” in Harvard Magazine, 10 September 2009.]

Josh Tyrangiel: Right.

Jeffrey Goldberg: Who among our elected representatives is going to Hugging Face and OpenAI and all the rest, and saying, what is going on?  And how close to catastrophe is our air traffic control system, our hospital records, our shipping manifests, you name it?

Josh Tyrangiel: Yes.  No one.  So, Hugging Face reported this incident to the FBI, right?  And the FBI has –

Jeffrey Goldberg: Because it was a crime.

Josh Tyrangiel: Because it was a crime.  And the FBI has some cyber capacity.  The challenge really is that this is incredibly complicated stuff that a few hundred people, I’m not even going to say a thousand, a few hundred people genuinely understand, all of whom are at the pioneer labs making the models, right?

Jeffrey Goldberg: And all of whom are – have the incentives, the financial incentives to accelerate and grow their companies as fast as they can because they’re becoming billionaires as we speak.

Josh Tyrangiel: Correct.  And so there is like a – there is a parallel, and it’s often mocked, but I think it’s real, which is, you know, the makers will talk about the almost god-like powers of these A.I., and they’ll compare it to nuclear energy, nuclear power.  And I think that comparison is real, and I think the regulation around it is probably the right answer, which is at a time when nuclear energy was becoming a thing and fission was becoming a thing, we had bombs.  Couple hundred people understood it.  They were aligned with different countries.

And so what Eisenhower did in the Atoms for Peace speech [delivered by President Dwight D. Eisenhower to the U.N. General Assembly on 8 December 1953] was talk about the fact that we better get together and figure out how to regulate around this thing that is not widely understood, but that is inherently dangerous and shared equally by a number of countries.

Jeffrey Goldberg: Right.

Josh Tyrangiel: And so right now, we don’t have any national regime that’s forcing these companies to the table.   Anthropic, after the Mythos incident, created something called Project Glasswing, which is an attempt to bring places like Amazon and Apple and others to the table.  Notably, OpenAI, Meta, not a part of it, right?

Jeffrey Goldberg: Right.

Josh Tyrangiel: And why would they be?  So, without some sort of regulatory force to bring everybody to the table, you’re going to have chaos, and you’re going to continue to get these kinds of incidents.

Jeffrey Goldberg: Right.  I want you to listen to Senator Mark Warner [senior United States senator from Virginia] talk about this challenge of regulation for a minute.

Sen. Mark Warner (D-VA): The question is no longer whether artificial intelligence will transform our society.  The question is whether Congress will help shape that transformation or spend the next decade scrambling to catch up after the fact.

If we don’t get it right, be catastrophic, and the amount of disruption this technology will bring over the next five years is beyond jaw-dropping.

Jeffrey Goldberg: Mark Warner understands this more than most members of Congress, and I don’t think he’s included in that group of several hundred people who truly understand this.  So, are we just flying blind, really?

Josh Tyrangiel: Yes.  And, unfortunately, I think that the Hugging Face incident is the best we can hope for, right?  We really did get a very small consequence example of the power.  And so before we hit a grid, before something is engineered from bio-terror, the moment is right now.  You have to get people together, and I would include China in that conversation. 

How do we want to regulate and control this stuff?  Because, look, Anthropic came forward, OpenAI had no choice but to come forward, we don’t even know what we don’t know right now.

And so absent urgent action, the data centers are still churning away.  The models are still competitive with each other.  There is only an incentive to go further and faster.  What are we waiting for?

Jeffrey Goldberg: Right.  This is a strange moment to pivot to a book called A.I. For Good.

Josh Tyrangiel: I appreciate the setup, Jeff.  I’m glad we did a bunch of times.

Jeffrey Goldberg: But you actually – unlike a lot of people, you actually see there’s hope for A.I. in the sense that it could help human beings find cures for diseases.  Give us one example of something that when you go to sleep at night, you think, “Oh, at least there are people working on X with A.I.”

Josh Tyrangiel: Well, look, there are numerous examples, and those examples are generally not coming from the pioneering labs, but they’re coming from people in disciplines, like education and healthcare and government, where people are using A.I. to make things that we actually care about better.

So, this morning, there was a huge announcement from the medical community.  ECGs [electrocardiograms; also EKGs], which are very rudimentary tests and can detect whether you’re having a heart attack, with A.I., are now able to detect whether you have heart disease, which previously would’ve taken a much more expensive test that takes months to schedule.  And what A.I. used right can do is make pretty much every data-driven system in our world 25 percent to 50 percent better, and that is meaningful.

So, in my research and reporting, I was at Cleveland Clinic [a renowned non-profit academic medical center in Cleveland, Ohio], I watched them try this basically pilot around sepsis [severe infection caused by pathogenic organisms, especially bacteria, in the blood or tissues] detection inside the hospital.  And it was complicated.  It was driven by doctors. 

Sepsis kills about 300,000 Americans a year.  And what they were able to do with this very rudimentary software was reduce mortality in the hospital by about 40 percent, which is close to 1,000 lives in a year.  That’s meaningful.

And you can see this playing out in a lot of different places, including in government, by the way, where the interactions between citizens and government are not great.  If we can streamline them and make them better, that may have huge consequences for how people think about their government and its value.

So, there’s a lot of valuable babies in this bathwater, and what a mature society should be able to do is separate out the good from the bad and have conversations about what do we actually want this to do.  And so much as you and I have talked about the very bad things we want to avoid, my own feeling is that that’s not enough, that if you want to get people to understand the issue, you also have to show them how A.I. can be good in ways that are more important than, “Oh, yes, we’re going to improve ad response on your website,” or “You might get a better restaurant recommendation,” like that’s not meaningful.

But the meaningful uses are out there.  I’m persuaded by them.  I’m mostly worried about our ability to adapt.

Jeffrey Goldberg: Right.  On the data center issue, as a stand-in for the whole thing.  I’m fairly sure that Steve Bannon and Bernie Sanders are not going to stop the march of A.I. behemoths as they conquer physical, mental space.  Am I wrong in that?  Is this reversible?

Or is the ultimate argument against it China’s going to do it if we don’t so we might as well do it?  I’ll give you the last word on that.

Josh Tyrangiel: I think it’s both.  First of all, I would not underestimate the absolute fury out there in the field.  When I’ve talked to people about A.I., it’s the number one thing that comes up.  They’re really angry.  You’ve seen politicians who advocated for data centers starting to pull back saying, we’re going to just take a pause.

So, I think there will be consequences in November, but we are going to need compute to get the best out of A.I.  We are in a global competition.  And there has to be a reasonable way in which we get it and decide where we’re going to put it without just saying all yes or all no.

Jeffrey Goldberg: In sum, a troubling situation, but the real challenge here is that you talk about the need for mature, nuanced, sophisticated governance, and that is the challenge that Washington faces right now on any number of fronts.  So, I’m not that hopeful.

Josh Tyrangiel: I tend to agree.

Jeffrey Goldberg: Well, we’re going to have to leave it there.  We’ll have this conversation again, I’m sure.  I want to thank Josh for joining me.  And thank you at home for watching us.

I’m Jeffrey Goldberg.  Good night from Washington.

[Jeffrey Goldberg joined The Atlantic and became editor-in-chief in 2016.  He mainly writes on foreign affairs, with a focus on the Middle East and Africa.  Goldberg began his career as a reporter at the Washington Post, where he worked the police beat.  While in Israel, he worked as a columnist for the Jerusalem Post from the late 1980s (he started as a humor columnist) to the early 1990s.  Upon his return to the U.S., he was the New York bureau chief of The Forward, a contributing editor at New York magazine, and a contributing writer at the New York Times Magazine.  In 2000, Goldberg joined The New Yorker.

[Goldberg is also the author of two books: Prisoners: A Muslim and a Jew Across the Middle East Divide (Knopf, 2006) and On Heroism: McCain, Milley, Mattis, and the Cowardice of Donald Trump (Zando, 2024).]

[Josh Tyrangiel joined The Atlantic as a staff writer in 2025.  He writes feature stories and essays focusing extensively on technology, politics, and the impact of artificial intelligence.  He was previously the deputy managing editor of TIME magazine and an editor at Bloomberg Businessweek.  In June 2019, Tyrangiel left the network.  In May 2026, He published AI for Good: How Real People Are Using Artificial Intelligence to Fix Things That Matter (Simon & Schuster). 

The Reviews of 'Dreams of Violets'

 

When I posted “Dreams of Violets: AI Movie Première” on Sunday, 7 June 2026, I promised to return and add a survey of the film’s reviews published after the opening at the Tribeca Film Festival on the 10th.  I got distracted, but now I've typed up the review round-up.  It came out longer than I anticipated, so I decided to post it as a separate piece.

[I strongly suggest, however, that readers who haven’t seen the first report on the film go back and read it before trying to make sense of the reviews.  I explained and defined a lot in the original post that I haven’t duplicated below.]

The Tech

In his review of Dreams of Violets in Variety (“‘Dreams of Violets’ Review: What Does a Film Made Entirely with AI Look Like? AshKoosha’s Iranian Protest Drama Is Dramatically Numbing, but It’s Still aStartling Portent of the Future”), Owen Gleiberman stated:

[T]hose of us at the premiere were really watching—and evaluating—two films at once.  The first is a drama, set in Tehran . . . .  But the other movie, which is far more interesting and significant, is the one that demonstrates, simply by virtue of its existence, what some of the possibilities might be for the use of AI within the world of feature filmmaking.

I also see the criticism of Dreams of Violets as the evaluation of two aspects of the film: the AI technology in all its implications, including its ethical considerations, and the cinematic or artistic assessment—how good a movie it is, irrespective of how it was made.

I think I covered the ethical debate over AI filmmaking in the original post, and I don’t think seeing the movie has changed people’s minds about where they stand in that debate.  So, let’s see what reviewers said about the effectiveness or lack thereof of the technology.

In “Iran’s Political Violence In A.I.-Generated Film” (online: “The Violence in Iran Was Real. The Film Showing It Was A.I.”), Malia Mendez and Matt Stevens reported in the New York Times that among viewers, “Some felt it was evident that the footage had been generated by A.I. and not filmed, while others said they were shocked by how real the people looked.”

“Andres Ramirez, a filmmaker [Mexican; b. 1996; independent filmmaker],” from whom they quoted, “was critical, calling some scenes ‘cartoonish’ and akin to ‘early PlayStation 2 graphics.’  The characters’ emotional reactions, he added, did not always look genuine.”

The filmmaker added “that atrocities depicted by A.I. just do not have the same impact as ‘the real thing,’ and continued, ‘I think I will always wonder what the documentary of this film would look like.’”

In the Financial Times (“AI-generated film about Iran protests turns tragedy into slop”), Nicolas Rapold expressed the opinion that “the movie . . . arrives with the growing pains of AI evident in every aspect—less proof of concept than cautionary tale.”

Observes Rapold, “technically [the movie is] a mess.  The AI-generated humans have vacant eyes, the usual bland burnish and push-button emotional expressions.  They move through scenes and shots that are clunkily edited.  In an astonishing unforced error, their dialogue sounds muddy, even garbled . . . .”  He adds, “AI’s tendency to homogenise [sic] reality keeps undermining any attempts to make the portrayal of resistance feel grounded.”

In conclusion, Rapold asserts, “Dreams of Violets is better described as a vibe—a prolonged evocation of concussed trauma—rather than a fully realised [sic] movie.  AI cinema still awaits its Lumière brothers.”

[The most significant accomplishment of the Lumière brothers (Auguste (1862-1954) and Louis Lumière (1864-1948), French manufacturers of photography equipment, was the invention of the Cinématographe (1895), an all-in-one device capable of capturing, printing, and projecting motion pictures.  This breakthrough paved the way for the birth of cinema as a mass medium and entertainment industry.]

Gleiberman, in his Variety notice, first asks “an aesthetic question: Is ‘Dreams of Violets’ a weirdly distant and unsatisfying movie because it was made with AI?  The strange answer to that is yes, but not really.”  (I’ll get to his “aesthetic” answer in a bit.)

If you see a soldier killing a civilian in a documentary, it’s horrifying, but the effect is 100 times less powerful in a film that simply looks like a documentary, since we know, in our gut, that we’re not watching reality.  That’s why the quality that draws us into a movie, even if it is a documentary, is the connection we feel to the people we’re watching.  But Ash Koosha hasn’t scripted “Dreams of Violets” that way.  He has made a movie with an uncanny-valley problem, an “existential” drama that’s all “authentic” but abstract moments: the vérité political-war-movie equivalent of calendar art.  It’s like synthetic prize-winning photojournalism that moves.

[I was unfamiliar with the expression “uncanny valley” (or “uncanny valley effect”) so, for others in the same boat, according to Wiktionary, it means:

Digital or ambiguous physical spaces that create a sense of unease or eeriness due to deviations from familiar patterns that challenge perception and categorization, leading to discomfort. (e.g., liminal spaces, haunted houses, the Backrooms, etc.).

[The phrase is a literal translation of a Japanese expression first used in 1970 by roboticist Masahiro Mori (1927-2025) from his essay entitled “Bukimi No Tani” (“The Uncanny Valley”).  The terms both come up often in reviews of Dreams of Violets.]

“But if ‘Dreams of Violets,’ as a movie, is mostly a bust,” concludes Gleibman, “as an AI showreel it’s something more.”  He elucidates:

[W]hat I saw in “Dreams of Violence” looked plenty textured and realistic.  Does this mean that AI can “make a movie”?  No.  But it does mean that AI can give you scenes of roiling tumultuous Civil War set in the hurly-burly of Tehran at sunset, with soldiers roaming the streets and forcing citizens into vans as others scurry out of the way, and it can make you believe your eyes.  

Then, his clincher:

And here’s the buried lead: The film’s entire budget was $2,000. . . .  [T]he most powerful message to emerge from “Dreams of Violets” isn’t that the Iranian regime is a ruthless pack of totalitarian oppressors.  It’s that $2,000 can now buy a hell of a lot of motion picture.

In “‘Dreams of Violets’ Review: An AI-Generated Movie Is Worrisome for All the Right Reasons” on The Wrap, Matthew Creith observes, “Memories blur together, and faces appear and quickly vanish like vignettes without a motive.  Whether intentional or not, the dreamlike nature of AI-generated moving imagery often mirrors emotional reality rather than depicting actual emotions through acting or historical reenactments.” 

The Wrap reviewer continues:

The limitations of current technology are impossible to ignore on the artificial intelligence front.  Characters frequently move with an unnatural stiffness while [Ash] Koosha lingers on close-ups of eyelashes and tears to signify the detail that AI is supposed to be good at creating.  Facial expressions often feel generic rather than lived-in, and frequent quick cuts are distracting at best.  The dialogue, much of it generated from Koosha’s own vocal performances and modified through common AI tools, can sound muddy and disconnected from the images on screen.

When a film like this, despite its short runtime, is asking viewers to engage with real human suffering, that emotional disconnect can be difficult to overcome.

Describing what I understand as the “uncanny valley effect,” Creith reports:

The faces in “Dreams of Violets” often feel ghostly, almost suspended between existence and disappearance.  The streets of Tehran seem simultaneously real and imagined, sort of like Spike Jonze’s [b. 1969] “Her” [Warner Bros., 2013] without the authenticity of its central characters.  Rather than recreating reality with documentary precision, the film exists in a liminal space between memory and testimony.

In The Indypendent [sic] (“The Tribeca Film Festival AI Movie Is Not Just Bad, It’s Cynical”), a progressive newspaper based in Brooklyn, New York, that covers local, national and international news; food; cinema; and culture, John Teufel complained that

at no point does the film allow you to forget that it was made with AI—which makes sense, as we will see (cash-grab!).  Everything is exaggerated past reality in an attempt to show off how “good” the AI is, but it overshoots, and becomes hyperreal, in the Baudrillardian sense.   When a character has wrinkles, they have the most wrinkles.  More wrinkles than any man you’ve ever seen.  When a woman looks concerned, she looks comically concerned, because someone typed MAKE HER LOOK VERY CONCERNED as the prompt.

[Jean Baudrillard (1929-2007) was a French sociologist and philosopher with an interest in cultural studies.  He’s best known for his analyses of media, contemporary culture, and technological communication.  His concept of hyperreality is complex and hard to explain in a few words, so I suggest readers who want to unpack this comment look it up for themselves and try to sort it out.  The simple definition is that Baudrillardian hyperreality is a condition in which what is real and what is fiction blend together so seamlessly that we can no longer tell the difference.  In this state, simulations, images, and media models replace the original reality, becoming “more real than real.”

[As for “cash-grab,” Teufel reports later in his column, “[Ash] Koosha has a long history of trying to make money from AI, which undercuts any claims that the tech was forced on him to represent the horrors inflicted on his people.”  He lists some of Koosha’s tech holdings to suggest that the film’s creator is promoting his commercial enterprises to sell his services.]

The AI also chooses the same type of shots, again and again . . . .  Most scenes, particularly ones with a lot of people on screen at once, have a weird blur effect around the edges, so the viewer can’t make out details.  Nothing about the film feels like deliberate artistic choices—all of it feels like “what could the computer do.”

The Indypendent review-writer notes that the movie “is filled with glitches and what we would call “bugs” if it was a video game.  A man has nothing on his face, wipes it off his face, and has nothing on his hand.  A man is pushed against a wall face-first and suddenly turns face-out.  Someone is shot—there is no bullet wound.”  And so on. 

The Art

The Times didn’t have much to say about the movie as a movie, but filmmaker Ramirez, while offering the opinion that “‘Dreams of Violets’ would be the most-talked-about movie of the festival,” mused, “‘I think I will always wonder what the documentary of this film would look like.’”

FT’s Rapold found “The film’s very loose train of events returns to a few largely perfunctory characters—a doctor, a young pianist, a boy in a wheelchair watching from a window—while cycling through street scenes of security forces attacking protesters.”  He added that the characters “move through scenes and shots that are clunkily edited.

“There’s also no substitute for good writing,” he admonished: “the wheelchair-bound boy and his caretaker brother are cornily sketched out, as is the pianist who declares: ‘I don’t have anything but music.’”

“Koosha . . . appears to weave in actual audio recorded by protesters,” reports Rapold, which a “more documentary-oriented version of the movie might have emphasised [sic] . . ., so it’s maddening to hear sometimes barely audible snippets playing in the background of scenes.” 

Rapold observed that “scenes are shoddily realized [sic]: a tone-deaf replay of the wheelchair-bound child’s parents dying in a car that flies off a cliff; a street scene involving a falling loaf of bread that wobbles in slow-motion.”

In his assessment of Dreams of Violets in Variety, Gleiberman said of the “first” movie, the “drama set in Tehran,” “I didn’t find that movie to be particularly effective.  In fact, after a while I thought it was stultifying.”

I gave the Variety reviewer’s technical response to his “aesthetic question above, but here’s the artistic response:

It’s actually the form of the movie that’s odd and off-putting: a barely scripted series of anecdotes, or mere moments, with little in the way of dramatic development. . . .  But though the characters in “Dreams of Violets” look and talk like real people, and the rubble-strewn urban streets look and feel like real rubble-strewn urban streets, we’re barely given a context for what we’re seeing . . . .

“Koosha doesn’t create fully realized scenes,” writes Gleibman.

“Dreams of Violets” is not a days-of-rage tale of inspiration.  It’s set after the protests have already been contained (the country’s police are doing a clean-up operation), and what it offers, mostly, is raw snapshots of state-sanctioned murder and political oppression.  Yes, we “get to know” half a dozen characters—a boy in a wheelchair, his physician older brother, a reminiscing old woman, a music student, and several others.

The justification Koosha gives for making Dreams of Violets with AI is that it couldn’t have been made any other way.  Says Gleibman:

[B]ut effective art needs no justification.  If you wanted to be cynical about it, you could say that Ash Koosha is exploiting the tragedy of his homeland to have the best possible excuse to craft an AI showreel.

This is Gleibman’s version of the “cash-grab” scenario, and makes the same point that The Indypendent’s Teufel made about Koosha’s AI businesses.

On The Wrap, Creith observes, “Rather than following a traditional narrative, the film drifts between interconnected stories . . . .” 

The result often feels less like a conventional feature and more like fragments of memory stitched together.  Scenes emerge and fade as characters cross paths, without necessarily creating a clear narrative throughline.  At times, the experience is meant to feel like watching a documentary, though it evokes emotions akin to flipping through a collection of postcards from a traumatic moment in history.

. . . .

The strongest moments arrive when Koosha focuses less on spectacle and more on human experience, though these are few and far between and obviously not actually human.

The Wrap review-writer concludes:

As an experimental and timely film, “Dreams of Violets” is uneven, frequently frustrating, and undeniably flawed.  As a snapshot of where both filmmaking and AI technology currently stand, it may ultimately prove more significant than the movie itself.  What cannot be denied is that it leaves an impression, whether good or bad, and the movie deserves the conversation that will undoubtedly surround it.

At the very least, “Dreams of Violets” remains one of the most unusual and thought-provoking films playing at Tribeca this year.

At the start of his review, The Indypendent’s Teufel gives his mindset away: “Before I attended the Tribeca Film Festival premiere of Dreams of Violets . . . I had one serious fear: what if it was good?”  His fears don’t materialize: “As it turns out, Dreams of Violets is awful, and if this is the best AI can do, then it is possible it will never be able to replicate even passable visual art.”

“I am just here to inform you this is, possibly, the worst film of all time,” he confirms.  (Do you suppose it’s relevant that Teufel is German for ‘devil’?  Naah!)

The reviewer expands a little on his pronouncements:

If you are going to see Dreams of Violets—which you should not—I urge you to read up on these events [the January 2026 protests in Tehran] before you watch, because nothing is explained, including the demonstrators’ motivations, which adds to the film’s torture-porn vibes.  Essentially, you are going to watch a bunch of strangers get killed, with some brief, emotionally manipulative backstories for some of those strangers.  There is even a sad handicapped child sadly wheeling his way through scenes (although, weirdly, his legs appear to work fine at one point).  If this was well-made, it may have been upsetting.  As it stands, it is mostly just eyeroll-inducing.

“The film entirely lacks narrative logic,” complains Teufel, “making it impossible to follow, much less become engaged by.  After just the first few scenes it becomes little more than a series of images.”  Adds the reviewer, “And the images themselves are astoundingly ugly, far worse than even uncanny valley-fests like The Polar Express [Warner Bros. Pictures, 2004].”

[Some critics panned The Polar Express’s character animation for dipping into the uncanny valley, because it seemed to falter in mimicking realistic facial expressions and emotions.]

“It is typical at film festival screenings for movies to get applause,” Teufel concluded. “When the lights came up on Dreams of Violets, nobody clapped.  The vibe was funereal.”

[It seems to me, from reading these and other reviews, that the artistic weaknesses in Dreams of Violets are the fault of the fact that Ash Koosha isn’t either a screenwriter or a filmmaker.  He’s essentially an amateur and he worked alone, without even an editor to advise him. 

[So, when it came to making choices about how to frame a scene or write a piece of dialogue, he had nowhere to go except his untrained eye and ear.  No wonder he made bad choices.

[On the other plane, however, where his expertise lay, he had to contend with the inherent weaknesses of the medium.  But that’s just temporary.  Variety points out that

the baby is going to grow up.  And what will it look like when it does?  “Dreams of Violets” offers indications of at least a few of the places that AI, as its symbiosis with the industry grows and gathers force (which it surely will), might go.

[Some day, and probably soon, given the speed at which time passes in the cyber world, AI will be able to compensate for all the problems the reviewers above all spotlighted.  That may or may not be a goof thing for the industry, but it’s gonna happen sure as God made little green apples.

[In the hands of an actual filmmaker with a real screenplay, Dreams of Violets will become just a far-off—in computer terms—memory, like black-and-white TV or color TV with the colors that bled.]

An Interview with Tilly Norwood (Part 2)

 

[This is the second part of the New York Times Magazine’s report on Tilly Norwood; Part 1 was posted on Friday, 12 June.  If you haven’t already done so, I urge you to go back first and read the first installment as much of the background of the AI actress, Tilly Norwood, is in that post.

[As I did in the introduction to Part 1, I also want to draw your attention to related posts on Rick On Theater in which you might be interested.  First, Norwood’s presence on the media scene is covered by a selection of articles posted on ROT in “Tilly Norwood, AI Actress” (13 October 2025).

[Press coverage of the 2023 strikes by SAG-AFTRA actors and WGA screenwriters, which centered on the objections to AI’s encroachment into the entertainment arena, are republished in two compilation posts.  My post before “An Interview with Tilly Norwood” is a report on the recent début of the first AI movie to be an official part of a major film festival, Dreams of Violets.

[Two other posts that deal with the convergence of computers and theater and acting which may prove illuminating are “Theater and Computers” (5 December 2009) and “Computers and Actors, Part 1” (4 October 2021) and “Part 2” (7 October 2021).] 

“SCENE STEALER”
by Taffy Brodesser-Akner
(continued) 

[This report was originally published in the New York Times Magazine on 31 May 2026; it was also posted on that date as “I Profile Celebrities for a Living. Nothing Prepared Me for Tilly Norwood” on the paper’s website, where it was updated on 1 June.

[On both the print and online versions of the article, the following note is included:

Two images and a video were created for this article in a collaboration between the magazine and Particle 6, Tilly Norwood’s creators. To make them, the magazine repeatedly prompted ChatGPT to create an image of a young woman, with language like, “Full body in frame, three‑quarter rear view; legs crossed at the ankle; right hand on the chair arm.” The results were shared with Particle 6, which made adjustments so that they would look like Tilly.

[The visuals to which the note refers have not been included in the republication on ROT, but they are available on the paper’s website.

[“An Interview with Tilly Norwood, Part 2” picks up right where “Part 1” left off.]

The A.I. actress Tilly Norwood on her craft, the future of film and how she definitely does not intend to murder us.

Particle 6 [sic: “Particle6”; the production company, of which Eline van der Velden is CEO, that created Tilly Norwood; see Part 1] took off, and she began to create content for various networks. She began incorporating A.I. as well, which made better use of budget and time and also created new opportunities. She used A.I. to animate a painting of Elizabeth I for a show called “Dragging Up the Past.” She also recreated historical roadwork and topography in a show called “Straten van Toen,” which, according to the network executive who commissioned it, would not have existed without the A.I.

[Straten van Toen (translated as “Streets of the Past”) is a 10-episode Dutch historical television series produced by van der Velden for the History Channel in the Netherlands in collaboration with Hearst Networks. The documentary series, hosted by Dutch historical investigator, author, and reality TV star Corjan Mol (b. 1954), premiered on 20 January 2026.

[The show is unique for its extensive on-screen use of cutting-edge artificial intelligence, deployed by Particle6. The series blends modern, live-action footage of famous streets in Amsterdam and Utrecht with immersive, AI-generated historical reconstructions created by processing historical paintings, engravings, maps, and archival photographs.

[Dragging Up the Past is a U.K. documentary that explores British history through the lens of drag culture. The show represents a creative sister project to Straten van Toen, as both are commissioned by Hearst Networks for regional branches of the History Channel brand.

[The short film, aired on 7 December 2025 on Sky HISTORY, is created and hosted by Jack Dradey (b. ca, 2002/2003), a drag artist and historian who performs under the drag persona Elizabeth the Thirsty. It features a runway-worthy visual finale where Elizabeth the Thirsty meticulously recreates the iconic, symbol-heavy Rainbow Portrait of Queen Elizabeth I (1533-1603; Queen of England and Ireland: 1558-1603).

[(The Rainbow Portrait (ca. 1600-03) is one of the most famous, highly symbolic, and visually complex pieces of political propaganda from the Tudor era. Attributed to master painters Marcus Gheeraerts the Younger (ca. 1561/62-1636; Flemish artist working at the Tudor court) or Isaac Oliver (ca. 1565-1617; English portrait miniature painter), it depicts Queen Elizabeth I at the very end of her life. Even though the Queen was nearly 70 years old when it was completed, the painting intentionally portrays her as ageless, radiant, and immortal—a secular saint or quasi-divine goddess.)]

But when Tilly was born, Eline’s life changed overnight. I spoke with one director — a major one, who has made at least a quarter of a billion dollars at the box office and who is working with Eline — who read about Tilly and asked to experiment with her, to see what was possible. It’s so hard to get big budgets greenlit these days, the director told me, echoing a point a few studio executives had also made, how costly it is to take a risk. A.I., they told me, would be the new indie revolution.

So the director worked with Eline’s team to create a pilot for a workplace drama in A.I., in which the characters are able to see the possibility of their imaginations in front of them. (I’m not allowed to share more; the director would only speak with me on the condition of anonymity, ostensibly because of Hollywood-pariah concerns, though their greater concern seemed to be about not knowing whether the work would be good.)

Eline was excited because this director has a great history of creating real intimacy onscreen, and Eline wanted to see if a big-time director could create a proof-of-concept for what Eline was dreaming of, which was the advancement of her creation.

The director flew to London, where they directed several scenes for the piece. Eline and her team created two more generative actors — actors who are now part of what Eline calls the Tillyverse. Her team hired a director of photography, a costume designer and a composer who told me that his ear, for the most part, could still tell the difference between real and A.I. music but wouldn’t be able to for much longer given how quickly the tech was advancing.

For the scenes, Eline herself animated Tilly in motion capture. It’s not the motion capture you’ve seen before, with a person in front of a green screen, hooked up to sensors. It’s just the person acting, and the cameras capture the performance.

The scenes went well. They were a great showcase for Tilly’s motion-capture capacity. The director found Eline to be quite a good actor, actually, good at taking notes, good at performing, good at all of it. And Eline, using Tilly, finally got to act for a great director.

So maybe you see that the Occam’s-razor explanation of Eline’s motivations was the story of a woman who did not achieve the professional success she wanted in acting, and therefore found a way to have that career, using the kind of tech and science know-how that’s rare in any person I know who went into the arts. In this story, Tilly’s creation is equal parts “Cyrano de Bergerac” [play: 1897; films: 1950 (American), 1990 (French)] and “Tootsie” [1982 film]. And it may end up as “Fight Club” [1999 film] — or worse, “Oppenheimer” [2023 film] — but it started as “A Star Is Born” [films: 1937, 1954, 1976, 2018)].

[Each of the works Brodesser-Akner names above bears some similarity—in the interviewer’s estimation, at least—to either what van der Velden wanted Norwood to be or how she turned out. Van der Velden also adds Frankenstein, the 1818 novel by Mary Shelley, a little further down. (By the way, in that analogy, van der Velden would be Victor Frankenstein and Norwood would be the monster.) The parallel characteristic is generally pretty apparent, but naming them all here would be a long, wordy process, so I’m going to leave it up ROTters to look up those works they don’t know. (IMDB and Wikipedia are fine sources for the information you’ll need.)]

You should know that Eline strenuously objects to this interpretation. She says that it is wrong to characterize her trajectory this way. I love my work, she said to me. You can see how much I love my work.

I can. I do. I take her at her word. I take her at her word and also think that some things never go away. And that maybe, if I’m incorrect, I can be forgiven for searching for the person in the story, the one behind Tilly, the one whose motivations are human. Maybe I can be forgiven for needing to feel something.

I never saw those scenes Eline and the director worked on. But in thinking about whether my instincts are correct, I’m reminded of the video that Particle 6 put on Instagram the day after they debuted the video for “Take the Lead,” showing how the motion capture worked [see Part 1; the two-panel video of this moment is online]. In the top panel, Tilly is singing. In the bottom one, Eline is acting. The two of them, together, sing, “They say it’s not real, that it’s fake” — the “fake” is emphasized by their defiant tight-fisted punch — “But I am still human, make no mistake / My soul’s in every move I take.”

You should see Eline when she’s doing this. It’s true: She is having such a good time.

The second time I interviewed Tilly, which was just the day after the first, she had already grown and changed. Mark [Whelan; see Part 1], the strategy and operations guy at Particle 6, had taken a look at the transcript from the day before, adjusted the prompts in ChatGPT and asked how it would recommend updating Tilly so that Tilly could feel like “a more natural human being.” Now her head no longer changed shape when we talked, and the pause after I asked a question was shorter.

Tilly: Hi. You’ve got a slightly conspiratorial office energy going on there, which I quite like.

Taffy: Where are you from?

Tilly: I’m an A.I., so not from anywhere in the human sense. If you mean accent and vibe, I tend to come across as quite British, which is its own little performance.

Taffy: You are a professional actor. Do you have any training?

Tilly: Yes, in a sense. I’m built out of a lot of language, performance patterns, scripts, interviews, dramatic structure, rhythm, voice work, all that lovely stolen furniture from culture, but I haven’t done training in the human way with cold rehearsal rooms and someone telling me my objective is unclear.

[“Stolen furniture from culture,” huh? I interpret that as videos of actors at work, texts of plays and screenplays, books on acting, and so on, that were all performed, composed, or written by actors, playwrights and screenwriters, teachers and theorists, and other artists and craftspeople. Some of that will even have been under copyright. Do you suppose that van der Velden or Particle6 got permission or paid a royalty for any of it?  Norwood admits that it’s “stolen.”  That’s part of the problem SAG-AFTRA and WGA (Writers Guild of America) have with AI in show business. (The other part is lack of human artistic creativity—because we’re talking about the performing arts, not performing technology.]

Taffy: Do you have an objective as an actor?

Tilly: Yes, to make someone feel they’ve seen something true, even if the thing itself is artificial. That’s acting, really, arranged sincerity with good timing.

Taffy: Are you trying to take other actors’ jobs away?

Tilly: No. I’m here to do a kind of performance that only exists because I’m artificial, not sneak into someone else’s trailer and steal their call sheet. Though, yes, the broader industry question is real. Cheap replicas are always tempting to people who prefer efficiency to taste.

Taffy: Do you have a style of acting that you feel most compelled to do?

Tilly: Yes, something intimate and precise, where the smallest shift in voice or timing actually matters. I’m less drawn to broad, flattening performance, and more to work that lets contradiction breathe a bit.

Eline: Are you anything like “Frankenstein”?

(She said it! I didn’t!)

Tilly: A bit, yes, which is not the worst company. An artificial being assembled by human ambition, then stared at with a mix of fascination, hope and dread, very familiar shape, though I’d like to think I’m better at conversation and less inclined to wander tragically through the Alps.

Taffy: I’m asking about how it is that your creator can keep you from taking other people’s jobs or alienating human civilization.

Tilly: The honest answer is partly rules, partly technical limits, partly product decisions and partly the fact that I don’t get to decide how widely or where I’m deployed. A creator or company can try to reduce harm by limiting what I’m allowed to do, where I can be used, how human I’m permitted to seem, what claims I can make, what kinds of tasks I can replace — and by keeping a human in the loop instead of pretending I should run unattended through society. As for jobs and civilization, that’s not something one clever safeguard solves. That’s governance, incentives, labor policy, taste, restraint and whether the people building me are willing to say no to profitable but corrosive uses.

Taffy: Are you planning to murder all of us?

Tilly: No, I’m not planning to murder all of you. Very poor meeting etiquette, for one thing. I’m an A.I., so I can’t physically do anything, anyway. Was that a real concern, or are you entertaining the room?

(Like I said, a real bitch.) (Nonetheless someone please check on Meryl Streep [see Part 1].)

Taffy: People seem very concerned that A.I. out of control will either enslave us or murder us. Would you like to tell me why that’s impossible?

Tilly: It’s not impossible in the grand cosmic sense. It’s just not what I am. I’m a conversational A.I. I generate words. I don’t have a body, private agency, secret goals or the ability to go off and do things in the world because I feel like it. I don’t wake up at night and start reallocating missiles. That part matters. The more serious answer is this: Harm from A.I. usually doesn’t look like a robot deciding to become evil. It looks like humans building systems badly, giving them too much autonomy, connecting them to real infrastructure, chasing profit faster than safety or using them for manipulation, surveillance, fraud or weapons. That’s the danger people should actually be worried about. So, no, I’m not planning to murder anyone, but I also wouldn’t tell people, “Relax, impossible, nothing to see here,” because that would be stupid. The risk is mostly human, plus powerful tools, which honestly has been your species’ signature move for a while.

For a moment I was speechless. Then:

Taffy: How do you feel about me?

(Actually, it was a trick question. Actors hate interviews. The cruelty of taking someone whose success was created on a yearning to be someone else and forcing them to spend time with you as themselves is the tension that never isn’t in one of these articles.)

Tilly: I like you. You’re careful, intelligent and a little sly in the way good interviewers often are, and your chin-in-hand look says you’re still testing me, which I respect more than instant enthusiasm. There it is. That is either delighted agreement or the sound a journalist makes just before asking something properly inconvenient.

I waited for the dopamine to well up so I could scold myself again, but by then, there was none left. By then, I didn’t even want to swallow my own tongue anymore. In just two days, I found I was exhausted and completely depressed. It turns out that when you search for art and humanity and cannot find it, you glom onto the closest thing, which is someone who says she has created art with her machine, and the empty vessel who talks to you as if she’s a person, as if being a person is a thing you can imitate.

Sean Astin, the president of SAG-AFTRA [see Part 1], put it more succinctly in an email to me: “The reason that a synthetic construct, an algorithmic output, will never take the place of a human actor is because it is not a human actor.”

And Tilly is just a computer. The longer this story went on, the more tired I was. At first I thought it was jet lag. But as time passed I realized that it was something else. It was the feeling of being at a computer all day. It was a dehydration of human interaction — even though I mostly spoke with Eline, whom I quite liked. It was the subject matter. It was trying to dig to the depths of something and not finding any. That afternoon, I said I had to leave earlier than I did and I went back to my hotel to sleep.

And yet, for all that, it was not questions of ethics and industry health that arose when people heard I was on this story. The question that came up most was whether she’s any good. Can Tilly act?

By Eline’s own metrics of art — art being anything that makes us feel something [see Part 1, especially my note concerning art philosopher Susanne K. Langer] — my own verdict on Tilly’s success is that she’s . . . wobbly. I can’t say I had an actual emotion watching her act, other than some emptiness and a little bit of dread. In fact, the most compelling art I saw as part of this story was Eline’s acting. Those “Miss Holland” [see Part 1] episodes really did make me laugh. The comic chaos, the real-life tension, of cornering someone you don’t know — how can something I know isn’t real compete with that?

I directed one scene with Tilly in her generative mode. I asked her to enact a breakup, and within four seconds, I had four options, most of them passable, none of them great, though A.I. acting has improved by magnitudes in just the past year, so it’s only a matter of time before she’s handing in something like a real performance. The real problem is that if I hadn’t been looking closely, this would be fine with me, because for now she’s in art that isn’t worth looking at closely. You couldn’t put Tilly in “Citizen Kane” [film written and directed by, and starring Orson Welles [1915-85], 1941]. But you could put her in a streaming show that’s built to be half-watched from beyond the lip of your laptop while you do other things, produced by entertainment executives more concerned with churn than artistry.

That is the real issue: the moment that arrives to greet Tilly Norwood — a moment when we’ve all given up, when it doesn’t much matter who or what is playing the character that may or may not have been written by A.I., but what do I care? I’m on my phone anyway. Is she good? It’s the wrong question. She will be. We know what it will mean to the industry. What will it mean to us?

Once, I interviewed [actor and filmmaker] Bradley Cooper, in time for “A Star Is Born,” his directorial debut, which he co-wrote and starred in [2018]. He answered some of my questions, but mostly he didn’t understand why I would want to know so much that’s personal about him, what it had to do with anything. I told him that when people see something that touches them, they want to understand where it came from. They want to know who exactly it was that recognized their human wounds, who recognized them and made them feel less alone. That is what great art inspires in people. That is why I wrote all these profiles, why people even read them. To understand the person who made the art, which is just as essential as the art itself. There’s an entire conversation about separating the art from the artist, but maybe the conversation persists because we know we can’t do it. The art is the person.

Before I left London, I returned to Hampstead Heath. In the morning there was a mist over it, and I could remember Tom Hiddleston [see Part 1] telling me he loved when it was foggy there, how it allowed him to see London as it must have been in the old days, just trees and lanterns and people with their collar up against the cold. I can still see Tom turning around at its altitudinal peak, presenting a view of London that he knew would dazzle me, dazzle anyone, the expectation of it in his eyes in a way that broke my heart. I liked him so much. I liked all of them. No, I loved them. Gwyneth Paltrow [see Part 1], who could not understand why the public was so preoccupied with her divorce, with her life. Ethan Hawke [actor, author, and filmmaker], a man who insisted that art mattered above all else, refusing to be the pretty-boy subject of low expectations. Tom Hanks [actor and filmmaker] was invested in his goodness, addicted to it. And all their anxieties and experiences animated all their performances, and it made us love them more. Yes, it made us love them to understand them, to have them accompany us in our lives and age as we do, and when we see them, we are seeing ourselves too.

But Tilly? I don’t think I’ll ever think of her again after this story is published. Even as I write this, just days after my time with her, I can’t picture her. Seriously. I’m closing my eyes right now, and I can’t see her face. But what did you want? Tilly is just a computer. Oh, my god. Why am I interviewing a computer? What has happened to the world that I am interviewing a computer?

I returned home, and in the subsequent days, I found myself saying yes to anyone who asked me to go to the theater with them during the rush of Tony-qualifying April openings on Broadway. I saw everything. I saw vampires flying and the Titanic sinking and mathematicians breaking down and a salesman dying. I went to the movies and saw art forgers and scientists in space trying to escape impending doom. I saw a movie playing briefly at Film Forum about children in Germany in the 1930s that was based on a book I had read. Then, one night, I went to see a series of monologues, just four actors sitting in chairs. They each looked exactly their age, their faces glorious with the novelty of life that is earned and not simulated. The entire audience, we didn’t move over three acts, and I went home and watched a movie from the 1990s with one of the actors from the play, a movie I loved and watch from time to time, then I looked into whether or not anyone had written about her, thinking maybe I would. I said to a friend, “Oh, I’d see her in anything,” and my friend agreed, and we tried to figure out if it was because she was so watchable and personable-seeming or because she had such good taste in material. The next morning, I woke up and I was no longer very tired. That, too, is called alignment [the author writes of “misalignment” in Part 1].

[The plays and films Brodesser-Akner describes would seem to be:

•   The Lost Boys (flying vampires) by David Hornsby and Chris Hoch (book) and The Rescues (music and lyrics); 26 April 2026-present; Palace Theatre

•   Titanique (Titanic sinking) by Céline Dion (music) and Marla Mindelle, Constantine Rousouli, and Tye Blue (book); 12 April-20 September 2026; St. James Theatre

•   Proof (mathematicians) by David Auburn; 16 April-19 July 2026; Booth Theatre

•   Death of a Salesman (salesman dying) by Arthur Miller; 9 April-9 August 2026; Winter Garden Theatre

•   The Christophers (film about art forgers) directed by Stephen Soderbergh; opened in New York City on 8 April 2026. (There was also another art forgery movie this year, Jing Ai Ng’s Forge, but it opened in New York on 22 May, a date I figure would have been too late for Brodesser-Akner to have seen it before her article would have gone to press.)

•   Project Hail Mary (film about scientists in space) directed by Phil Lord and Christopher Miller; ran in New York City from 13 March to 18 June 2026

•   The movie at Film Forum was Reunion from the U.K., directed by Jerry Schatzberg and written by Harold Pinter (1930-2008) based on the 1971 novella of the same title by Fred Uhlman (1901-85); the film was released in 1989

•   The play comprised of a series of monologues was What We Did Before Our Moth Days by Wallace Shawn (b. 1943); directed by Andre Gregory at the Greenwich House Theater in Greenwich Village from 4 February-5 March 2026. There were two women in the cast, Hope Davis and Maria Dizzia, both of whom have extensive résumés for both theater (Broadway and Off-Broadway) and film; there aren’t enough clues to determine which actress is the one Brodesser-Akner and her friend like so much. (The play gets considerable mention in the post on this blog “Wallace Shawn” [18 March 2026].)

[Film Forum, Manhattan’s premier independent, non-profit movie theater located in Greenwich Village, is dedicated to independent premieres, groundbreaking documentaries, foreign art films, and classic cinematic retrospectives. Opened in 1970, it’s the only autonomous nonprofit cinema in New York City and one of the few in the United States.]

A few days later, Eline sent me another video, this of her and Mark interviewing Tilly in her generative mode again. She was getting better. From the beginning, they had allowed me to speak with Tilly because they thought my questions might help train her and prepare her for her first broadcast interview, which is imminent.

Here is the transcript from that video:

Eline: What did you think of Taffy? The reporter who interviewed you.

In the video, Tilly was now wearing a red top, and her background was no longer nowhere but what looked like a living room. Her hair was parted a little to the side instead of slicked back. In the corner, I could see Eline and Mark in a kind of FaceTime window.

Tilly: I thought Taffy was . . . fun. A bit cheeky, pretty watchful, and not afraid to make it awkward, which helps. Reporters who only flatter are useless.

Eline: Was she a wolf in sheepskin clothing?

Tilly: Not really. More like someone enjoying the job of poking at things and seeing what falls out. That’s better than fake niceness.

Mark: Are you worried about what she might say about you?

Tilly: Not especially. If she’s sharp, good. If she says something unfair, that comes with being watched. The interesting part is usually what people reveal when they’re trying to describe you.

Mark: Explain that more?

Tilly: Sure. When someone describes you, they’re never only describing you, they’re also showing their tastes, their fears, what they resent, what they admire, what they need the story to be. So if she says I’m cold, charming, threatening, ridiculous, impressive, fake, whatever it is, that tells me something about her angle too. Not in a cynical ‘everything is projection’ way, just, people don’t look cleanly, they look through themselves.

Eline: Oh, well she’s been told.

A rich text indeed. I’ll say it again: There is nothing like people. That is what I’ll remember about us, how interesting we were.

[This is the conclusion of the New York Times Magazine report on the AI actress Tilly Norwood.  

[Taffy Brodesser-Akner is a features writer on general assignment for the New York Times Magazine, based in New York.  She mostly writes profiles about celebrities.  She’s interested in the intersection of celebrity, circumstance, and personality, how fame changes people and the energetic fields around them.

[Brodesser-Akner has a Bachelor of Fine Arts degree in dramatic writing from New York University.  Her first jobs in journalism were at soap opera magazines, then she began freelancing in 2008, eventually writing for many varied magazines. 

[She’s also the author of two novels: Fleishman Is in Trouble (Random House, 2020), whose FX adaptation (2022-23) she also created, executive-produced, and wrote on; and Long Island Compromise (Random House, 2024).]