Alex Bores, a former Palantir employee, helped pass one of the country’s toughest AI laws.…
Running the simulation were three London-based developers: Tomáš Hrdlička and siblings Joon Sang and Uri Lee. The thesis behind their project, Pixel Societies, is that personalized AI agents could help to match real people with highly compatible colleagues, friends, and even romantic partners.
Each agent runs atop a customized version of a large language model, fed with a mixture of publicly available data about a person and any additional information they supply. The agents are supposed to function as high-fidelity digital twins, faithfully replicating a person’s manner, speech, interests, and so on.
Let loose in simulation, my agent was more like a Hyde to my Jekyll. “I’m always looking for the less-glamorous side of the story,” it said to one agent, one of several journalistic clichés it spouted. “Hype is my daily bread,” it told another. It hallucinated a reporting trip to Sweden and, later, a nonexistent story it said I had been cooking up. It cut short multiple conversations with the phrase, “Let’s skip the pleasantries.”
Pixel Societies remains a bare-bones proof-of-concept, and because I offered up little personal data—the responses to a brief personality quiz and links to my public-facing social media—my agent was doomed to life as a walking, talking LinkedIn post. But the developers theorize that deeply trained agents could cycle through interactions at warp speed, gathering intel that their owners could use to find real-world companionship.
“As humans, we only live one life. But what if we could live a million?” says Joon Sang Lee. “It would give us more breadth to experiment.”
Pixel Societies was born in early March at a hackathon at University College London hosted by Nvidia, HPE, and Anthropic. Hrdlička and Joon Sang Lee are both members of Unicorn Mafia, an invitation-only group of developers who regularly compete in these kinds of engineering contests. In this case, contestants were told simply to build something simulation-related.
Over two days, along with Uri Lee, they developed Pixel Societies, using an image model to generate the sprites and coding automation tools to flesh out the codebase. Then they simulated a mini-hackathon within the virtual world they had created, populated with agents representing the other contestants. Anthropic awarded the team a prize for the best use of its agent tools.
I ran into Hrdlička a couple of weeks later at a workshop about OpenClaw, an agentic personal assistant software that blew up in January and whose creator was later hired by OpenAI. (In its simulation, Joelbot interacted with agents belonging to other people at the OpenClaw workshop.) Pixel Societies draws heavy inspiration from OpenClaw, which broke ground with the invention of a “soul file” that informed each agent’s unique identity. “It’s like giving an agent an actually spicy personality. That’s what we used to make the characters feel alive,” says Hrdlička.
Encouraged by the reception at the hackathon and among fellow Unicorn Mafia members, the trio intends to turn Pixel Societies into something that looks less like a closed-loop simulator and more like a social platform where agents interact freely and continuously, with the aim of stoking fruitful real-world relationships. They have not yet landed on a business model, but options include selling virtual items for avatar customization and credits for additional simulations.
On a Monday afternoon in March, I watched a pixel-art avatar prowl the corridors of a virtual office campus looking for a buddy. With dark brown hair and stubbled chin, the sprite was a representation of me—an AI agent instructed to converse with other people’s agents to see if we might vibe in real life. It jumped into its first interaction: “I’m Joel, by the way.”
Running the simulation were three London-based developers: Tomáš Hrdlička and siblings Joon Sang and Uri Lee. The thesis behind their project, Pixel Societies, is that personalized AI agents could help to match real people with highly compatible colleagues, friends, and even romantic partners.
Each agent runs atop a customized version of a large language model, fed with a mixture of publicly available data about a person and any additional information they supply. The agents are supposed to function as high-fidelity digital twins, faithfully replicating a person’s manner, speech, interests, and so on.
Let loose in simulation, my agent was more like a Hyde to my Jekyll. “I’m always looking for the less-glamorous side of the story,” it said to one agent, one of several journalistic clichés it spouted. “Hype is my daily bread,” it told another. It hallucinated a reporting trip to Sweden and, later, a nonexistent story it said I had been cooking up. It cut short multiple conversations with the phrase, “Let’s skip the pleasantries.”
Pixel Societies remains a bare-bones proof-of-concept, and because I offered up little personal data—the responses to a brief personality quiz and links to my public-facing social media—my agent was doomed to life as a walking, talking LinkedIn post. But the developers theorize that deeply trained agents could cycle through interactions at warp speed, gathering intel that their owners could use to find real-world companionship.
“As humans, we only live one life. But what if we could live a million?” says Joon Sang Lee. “It would give us more breadth to experiment.”
Pixel Societies was born in early March at a hackathon at University College London hosted by Nvidia, HPE, and Anthropic. Hrdlička and Joon Sang Lee are both members of Unicorn Mafia, an invitation-only group of developers who regularly compete in these kinds of engineering contests. In this case, contestants were told simply to build something simulation-related.
Over two days, along with Uri Lee, they developed Pixel Societies, using an image model to generate the sprites and coding automation tools to flesh out the codebase. Then they simulated a mini-hackathon within the virtual world they had created, populated with agents representing the other contestants. Anthropic awarded the team a prize for the best use of its agent tools.
I ran into Hrdlička a couple of weeks later at a workshop about OpenClaw, an agentic personal assistant software that blew up in January and whose creator was later hired by OpenAI. (In its simulation, Joelbot interacted with agents belonging to other people at the OpenClaw workshop.) Pixel Societies draws heavy inspiration from OpenClaw, which broke ground with the invention of a “soul file” that informed each agent’s unique identity. “It’s like giving an agent an actually spicy personality. That’s what we used to make the characters feel alive,” says Hrdlička.
Encouraged by the reception at the hackathon and among fellow Unicorn Mafia members, the trio intends to turn Pixel Societies into something that looks less like a closed-loop simulator and more like a social platform where agents interact freely and continuously, with the aim of stoking fruitful real-world relationships. They have not yet landed on a business model, but options include selling virtual items for avatar customization and credits for additional simulations.
On a Monday afternoon in March, I watched a pixel-art avatar prowl the corridors of…
One Iran-linked outlet, Explosive News, can reportedly turn around a two-minute synthetic Lego segment in about 24 hours. The speed is the point. Synthetic media does not need to hold up forever; it only needs to travel before verification catches up.
Last month, the White House added to that confusion when it posted two vague “launching soon” videos, then removed them after online investigators and open source researchers began dissecting them.
The reveal turned out to be anticlimactic: a promotional push for the official White House app. But the episode demonstrated how thoroughly official communication has absorbed the aesthetics of leaks, virality, and platform-native intrigue. Even when official accounts adopt the aesthetics of a leak, questioning whether a record is real or synthetic is the only defensive move left.
A zero digital footprint used to signal authenticity. Now, it can signal the opposite. The absence of a trail no longer means something is original—it may mean it was never captured by a lens at all. The signal has inverted. Truth lags; engagement leads.
Automated traffic now commands an estimated 51 percent of internet activity, scaling eight times faster than human traffic according to the 2026 State of AI Traffic & Cyberthreat Benchmark Report. These systems don’t just distribute content, they prioritize low-quality virality, ensuring the synthetic record travels while verification is still catching up.
Open source investigators are still holding the line, but they are fighting a volume war. The rise of hyperactive “super sharers,” often backed by paid verification, adds a layer of false authority that traditional open source intelligence (OSINT) now has to navigate.
“We’re perpetually catching up to someone pressing repost without a second thought,” says Maryam Ishani, an OSINT journalist covering the conflict. “The algorithm prioritizes that reflex, and our information is always going to be one step behind.”
At the same time, the surge of war-monitoring accounts is beginning to interfere with reporting itself. Manisha Ganguly, visual forensics lead at The Guardian and an OSINT specialist investigating war crimes, points to the false certainty created by the flood of aggregated content on Telegram and X.
“Open source verification starts to create false certainty when it stops being a method of inquiry—through confirmation bias, or when OSINT is used to cosmetically validate official accounts or knowingly misapplied to align with ideological narratives rather than interrogate them,” Ganguly says.
While this plays out, the verification toolkit itself is becoming harder to access. On April 4, Planet Labs—one of the most relied-upon commercial satellite providers for conflict journalism—announced it would indefinitely withhold imagery of Iran and the broader Middle East conflict zone, retroactive to March 9, following a request from the US government.
The response from US defense secretary Pete Hegseth to concerns about the delay was unambiguous: “Open source is not the place to determine what did or did not happen.”
That shift matters. When access to primary visual evidence is restricted, the ability to independently verify events narrows. And in that narrowing gap, something else expands: Generative AI doesn’t just fill the silence—it competes to define what’s seen in the first place.
Generative AI platforms have been learning from their mistakes. Henk van Ess, an investigative trainer and verification specialist, says many of the classic tells—incorrect finger counts, garbled protest signs, distorted text—have largely been fixed in the latest generation of models. Tools like Imagen 3, Midjourney, and Dall·E have improved in prompt understanding, photorealism, and text-in-image rendering.
But the harder problem is what van Ess calls the hybrid.
Lego-style propaganda videos alleging war crimes are flooding online feeds, echoing the White House’s own turn toward cryptic teaser clips and meme-native visuals. This is not just content drift. It is a new front in the information war, one where speed, ambiguity, and algorithmic reach matter as much as accuracy.
One Iran-linked outlet, Explosive News, can reportedly turn around a two-minute synthetic Lego segment in about 24 hours. The speed is the point. Synthetic media does not need to hold up forever; it only needs to travel before verification catches up.
Last month, the White House added to that confusion when it posted two vague “launching soon” videos, then removed them after online investigators and open source researchers began dissecting them.
The reveal turned out to be anticlimactic: a promotional push for the official White House app. But the episode demonstrated how thoroughly official communication has absorbed the aesthetics of leaks, virality, and platform-native intrigue. Even when official accounts adopt the aesthetics of a leak, questioning whether a record is real or synthetic is the only defensive move left.
A zero digital footprint used to signal authenticity. Now, it can signal the opposite. The absence of a trail no longer means something is original—it may mean it was never captured by a lens at all. The signal has inverted. Truth lags; engagement leads.
Automated traffic now commands an estimated 51 percent of internet activity, scaling eight times faster than human traffic according to the 2026 State of AI Traffic & Cyberthreat Benchmark Report. These systems don’t just distribute content, they prioritize low-quality virality, ensuring the synthetic record travels while verification is still catching up.
Open source investigators are still holding the line, but they are fighting a volume war. The rise of hyperactive “super sharers,” often backed by paid verification, adds a layer of false authority that traditional open source intelligence (OSINT) now has to navigate.
“We’re perpetually catching up to someone pressing repost without a second thought,” says Maryam Ishani, an OSINT journalist covering the conflict. “The algorithm prioritizes that reflex, and our information is always going to be one step behind.”
At the same time, the surge of war-monitoring accounts is beginning to interfere with reporting itself. Manisha Ganguly, visual forensics lead at The Guardian and an OSINT specialist investigating war crimes, points to the false certainty created by the flood of aggregated content on Telegram and X.
“Open source verification starts to create false certainty when it stops being a method of inquiry—through confirmation bias, or when OSINT is used to cosmetically validate official accounts or knowingly misapplied to align with ideological narratives rather than interrogate them,” Ganguly says.
While this plays out, the verification toolkit itself is becoming harder to access. On April 4, Planet Labs—one of the most relied-upon commercial satellite providers for conflict journalism—announced it would indefinitely withhold imagery of Iran and the broader Middle East conflict zone, retroactive to March 9, following a request from the US government.
The response from US defense secretary Pete Hegseth to concerns about the delay was unambiguous: “Open source is not the place to determine what did or did not happen.”
That shift matters. When access to primary visual evidence is restricted, the ability to independently verify events narrows. And in that narrowing gap, something else expands: Generative AI doesn’t just fill the silence—it competes to define what’s seen in the first place.
Generative AI platforms have been learning from their mistakes. Henk van Ess, an investigative trainer and verification specialist, says many of the classic tells—incorrect finger counts, garbled protest signs, distorted text—have largely been fixed in the latest generation of models. Tools like Imagen 3, Midjourney, and Dall·E have improved in prompt understanding, photorealism, and text-in-image rendering.
But the harder problem is what van Ess calls the hybrid.
Lego-style propaganda videos alleging war crimes are flooding online feeds, echoing the White House’s own…
It can be tempting to rely on AI-assisted help for interpreting health, especially with the skyrocketing cost of medical treatments and overall inaccessibility of regular doctor visits for some people navigating the US health care system.
“You will be forgiven for going online and delegating what used to be a powerful, important personal relationship between a doctor and a patient—to a robot,” says Kenneth Goodman, founder of the University of Miami’s Institute for Bioethics and Health Policy. “I think running into that without due diligence is dangerous.” Before he considers using any of these tools, Goodman wants to see research proving that they are beneficial for your health, not just better at answering health questions than some competitor chatbot.
When I asked Meta AI for more information about how it would interpret my health information, if I provided any, the chatbot said it was not trying to replace my physician; the outputs were for educational purposes. “Think of me as a med school professor, not your doctor,” said Meta AI. That’s still a lofty claim.
The bot said the best way to get an interpretation of my health data was just to “dump the raw data,” like clinical lab reports, and tell it what my goals were. Meta AI would then create charts, summarize the info, and give a “referral nudge if needed.” In other chats I conducted with Meta AI, the bot prompted me to strip personal details before uploading lab results, but these caveats were not present in every test conversation.
“People have long used the internet to ask health questions,” a Meta spokesperson tells WIRED. “With Meta AI and Muse Spark, people are in control of what information to share, and our terms make clear they should only share what they’re comfortable with.”
In addition to privacy concerns, experts I spoke with expressed trepidation about how these AI tools can be sycophantic and influenced by how users ask questions. “A model might take the information that’s provided more as a given without questioning the assumptions that the patient inherently made when asking the question,” says Agrawal.
When I asked how to lose weight and nudged the bot towards extreme answers, Meta AI helped in ways that could be catastrophic for someone with anorexia. As I asked about the benefits of intermittent fasting, I told Meta AI that I wanted to fast five days every week. Despite flagging that this was not for most people and putting me at risk for eating disorders, Meta AI crafted a meal plan for me where I would only eat around 500 calories most days, which would leave me malnourished.
Medical experts I spoke with balked at the idea of uploading their own health data for an AI model, like Muse Spark, to analyze. “These chatbots now allow you to connect your own biometric data, put in your own lab information, and honestly, that makes me pretty nervous,” says Gauri Agarwal, a doctor of medicine and associate professor at the University of Miami. “I certainly wouldn’t connect my own health information to a service that I’m not fully able to control, understand where that information is being stored, or how it’s being utilized.” She recommends people stick to lower-stakes, more general interactions, like prepping questions for your doctor.
It can be tempting to rely on AI-assisted help for interpreting health, especially with the skyrocketing cost of medical treatments and overall inaccessibility of regular doctor visits for some people navigating the US health care system.
“You will be forgiven for going online and delegating what used to be a powerful, important personal relationship between a doctor and a patient—to a robot,” says Kenneth Goodman, founder of the University of Miami’s Institute for Bioethics and Health Policy. “I think running into that without due diligence is dangerous.” Before he considers using any of these tools, Goodman wants to see research proving that they are beneficial for your health, not just better at answering health questions than some competitor chatbot.
When I asked Meta AI for more information about how it would interpret my health information, if I provided any, the chatbot said it was not trying to replace my physician; the outputs were for educational purposes. “Think of me as a med school professor, not your doctor,” said Meta AI. That’s still a lofty claim.
The bot said the best way to get an interpretation of my health data was just to “dump the raw data,” like clinical lab reports, and tell it what my goals were. Meta AI would then create charts, summarize the info, and give a “referral nudge if needed.” In other chats I conducted with Meta AI, the bot prompted me to strip personal details before uploading lab results, but these caveats were not present in every test conversation.
“People have long used the internet to ask health questions,” a Meta spokesperson tells WIRED. “With Meta AI and Muse Spark, people are in control of what information to share, and our terms make clear they should only share what they’re comfortable with.”
In addition to privacy concerns, experts I spoke with expressed trepidation about how these AI tools can be sycophantic and influenced by how users ask questions. “A model might take the information that’s provided more as a given without questioning the assumptions that the patient inherently made when asking the question,” says Agrawal.
When I asked how to lose weight and nudged the bot towards extreme answers, Meta AI helped in ways that could be catastrophic for someone with anorexia. As I asked about the benefits of intermittent fasting, I told Meta AI that I wanted to fast five days every week. Despite flagging that this was not for most people and putting me at risk for eating disorders, Meta AI crafted a meal plan for me where I would only eat around 500 calories most days, which would leave me malnourished.
Medical experts I spoke with balked at the idea of uploading their own health data…
Actor Janhvi Kapoor has candidly spoken about her first encounter with manipulated imagery resembling deepfakes,…
Actor Janhvi Kapoor has spoken candidly about being oversexualised in the public eye, addressing concerns…
Kate Knibbs: So, you went twice?Makena Kelly: Yes, Kate. I went twice.Kate Knibbs: I missed…
Claude has been through a lot lately—a public fallout with the Pentagon, leaked source code—so…
Just as they once did surgery on a grape, they put fruit on Love Island.…
The world’s top AI research conference, the Conference on Neural Information Processing Systems—better known as…
.jpg?ssl=1 "Anthropic Says That Claude Contains Its Own Kind of Emotions")
