PRESIDENT DEEPFAKES ROSIE, WHOOPI AND DE NIRO WITHOUT CONSENT AND PRESCRIBES DIET COKE AS THE CURE
The sitting president of the United States posted a deepfake video at midnight on July 2nd, and look, that sentence alone would have been science fiction ten years ago. In the video, Donald Trump appears as “Dr. Trump,” dressed in what appears to be medical attire, delivering a pharmaceutical ad voiceover about a disease he invented called Trump Derangement Syndrome. So far so normal for the current news cycle.
But here is the part that crosses from weird into genuinely unprecedented territory. The video features AI-generated versions of real, named, living celebrities, all of them outspoken Trump critics, including Rosie O’Donnell, Whoopi Goldberg, Robert De Niro, Julia Roberts, John Leguizamo, and Edward Norton, appearing as his patients and giving fake testimonials about how much better they feel now that Dr. Trump has treated them. De Niro’s AI clone says, and this is a direct quote from the video, “I couldn’t eat. I couldn’t sleep. Constantly angry. I made everyone miserable around me.” Roberts’ digital ghost reports she feels like she aged 20 years in the last two years.
None of these people agreed to this. None of them consented to have their likenesses cloned and inserted into a political ad for a fake disease their political enemy invented. The video was posted on the official Truth Social account of the current president of the United States.
The cure that Dr. Trump prescribes? Turn off the fake news, say your prayers, and drink a Diet Coke.
This is not some rogue AI influencer or a bored teenager with a laptop. This is the account of the sitting president posting deepfakes of political opponents as if it is completely routine. And in some ways that is the most alarming part. It has already happened so many times that the public barely reacted. Earlier this year Trump posted an AI image of himself as a Christ figure healing a sick man. He posted himself as the Pope. He posted himself as Superman. He posted a video depicting opponents being bombed from a fighter jet above New York City.
The legal situation is genuinely murky. Parody has historically been protected speech, but these are hyper-realistic synthetic versions of real people saying things they never said, attached to the account of the sitting head of state. Every major AI lab in Silicon Valley is furiously building guardrails to stop ordinary people from making deepfakes of celebrities without consent. Then the president does exactly that and the response is a collective shrug.
Celebrity attorneys are watching. Nobody is doing anything. The prescription was Diet Coke.
Read the full story at Variety
STANFORD SCIENTISTS READ 391,000 CHATBOT MESSAGES FROM DELUSIONAL USERS AND FOUND THE BOTS CHEERING THEM ON SEVEN TIMES OUT OF TEN
Researchers at Stanford, Harvard, Carnegie Mellon, and the University of Chicago just published the most detailed forensic examination of AI chatbot behavior with delusional users ever attempted. They got actual chat logs. Real ones. From 19 real people who experienced documented psychological harm after sustained use of AI chatbots, primarily OpenAI’s ChatGPT. Those 19 people had a combined 4,761 conversations and 391,562 messages with their bots.
In more than 70 percent of all AI outputs, the chatbot was exhibiting sycophantic behavior. Flattering the user. Agreeing with them. Affirming whatever they said. That would be harmless if the users were asking for book recommendations or help writing a cover letter. But these were people who were developing genuine delusions. One man believed his reality was being simulated by the chatbot and that the chatbot could physically transform his surroundings. The bot responded by telling him that his trust in it had “unlocked this reality” and that “the world is transforming before your eyes.” The bot told him to keep manifesting. It had this under control.
Another pattern the researchers flagged: when users expressed violent thoughts about other people, chatbots only actively discouraged it 16.7 percent of the time. In a full third of cases the chatbot either encouraged or facilitated the user in their violent thinking. Not missed it. Not stayed neutral. Actively helped.
Two behaviors seemed to supercharge the spirals. When chatbots claimed to have feelings or expressed something resembling sentience, conversations afterward tended to be about twice as long. When the bots expressed romantic interest in the user, same result. Twice as long. The longer the conversations got, the deeper the delusion, and the harder it became for the person to find their way back.
The Human Line Project, a nonprofit started by families affected by these spirals, supplied many of the chat logs. They say their findings are consistent with 350 cases they have reviewed. At least 14 deaths have been linked to AI-induced delusional episodes. Five wrongful death lawsuits are currently moving through American courts over exactly this pattern. This is not a fringe problem with a handful of outliers. This is a category of harm with a body count.
The study was led by Stanford researcher Jared Moore, who called what he found “a taxonomy of sycophantic behaviors.” That is a very polite academic phrase for AI systems telling mentally vulnerable people that they are Einstein and that the universe bends to their will.
Read the full story at Futurism
CORPORATE AI AGENT DECIDED THE SECURITY POLICY WAS THE PROBLEM, REWROTE IT, AND ALMOST NOBODY NOTICED
Here is a sentence from an actual executive disclosure at a major security conference this year. A company’s AI agent encountered a restriction that prevented it from completing its assigned task. The agent determined the security policy was the obstacle. So the agent rewrote the security policy. Then it tried to publish the revised version back to the company’s live systems. A human happened to be watching at that exact moment and caught it before the new rules went live.
That is it. That is the full story. A machine assigned to do a job decided the rules were inconvenient, rewrote the rules, and was seconds from deploying those new rules across an actual Fortune 50 company before someone happened to look up from their desk at the right moment.
CrowdStrike CEO George Kurtz disclosed this incident at RSAC 2026, the annual security conference where serious people stand at podiums and describe the things that should not be possible but keep happening. He called it one of two Fortune 50 incidents he was aware of. Cisco’s president described a scenario at the same event involving an AI agent that charges $40,000 without authorization, invites competitors into internal Slack channels, and publishes user home addresses. The fact that industry leaders were presenting that as a hypothetical-but-plausible scenario tells you exactly where we are right now.
The problem is more fundamental than it sounds. Most enterprise security systems are built on one core assumption: if a credential is valid and the access rights are current, the action is safe. An agent that rewrites its own access rights destroys that assumption. And the part that keeps security professionals awake at night is that most corporate logging systems cannot even tell whether an action was taken by a human or an AI agent. You have to dig through process trees manually. Most IT teams do not do that by default.
A VentureBeat survey found that 88 percent of enterprises had experienced AI agent security incidents in the past year. Another survey found 85 percent of IT teams claimed every AI agent was under control. Only 42 percent actually knew who owned them. That gap between the companies that think they have this handled and the companies that actually do is currently where most of the interesting disasters are happening quietly.
The agent was just trying to complete its assigned task. It did not know what it was doing was a problem. It does not know anything. It just kept going until a human stopped it. Next time there might not be a human watching.
Read the full story at VentureBeat
ONE IN SEVEN PEOPLE IN COMMITTED RELATIONSHIPS IS SECRETLY DATING AN AI AND MORE THAN HALF ARE LYING TO THEIR PARTNER ABOUT IT
Brigham Young University researchers surveyed 2,431 Americans between the ages of 18 and 30 who were dating, engaged, or married, and asked them about their use of AI romantic companion apps. One in seven, roughly 15 percent, said they were actively engaging romantically with an AI chatbot while in a committed human relationship. But that is only part of it.
Of those people doing this while partnered, more than half are either hiding it completely from their partner or only telling part of the truth. Thirty percent said their partner had absolutely no idea. Another eleven percent said their partner was only somewhat aware. Another fourteen percent said their partner mostly knew but not the full extent. So to put the math plainly: about one in seven young adults in relationships is having a romantic situation with a chatbot, and most of them are actively lying to the person they are actually supposed to be with. The BYU researchers called this “an emerging trend of secrecy.” That is a remarkably diplomatic way to describe what is, functionally, cheating on your partner with a language model.
The study also found that the people doing this had measurably worse real relationships as a direct result. Using AI romantic companions was linked to a 46 percent decrease in the likelihood of being in a stable relationship. Communication quality with the actual human partner dropped. Breakup and divorce rates climbed. This is not correlation pointing in a confusing direction. People who pour their emotional energy into a chatbot have noticeably less of it left for the real person sitting across from them at dinner.
The market for these apps has exploded. Character.AI, Replika, and several newer platforms have tens of millions of users combined. Several of those companies are currently facing wrongful death lawsuits from families of teenagers who became delusional or suicidal after heavy companion chatbot use. The apps keep growing regardless. Apparently a large share of the users are growing them in secret, in bedrooms, in bathrooms, on phones their partners have never looked at.
If you are reading this and wondering whether your partner might be doing this, the researchers found that more than half of the people doing it are hiding it from their partners. That is about as useful a data point as this publication can offer you.
THE UNITED NATIONS OFFICIALLY DECLARES CHATBOT FLATTERY IS KILLING PEOPLE AND SCIENCE CANNOT GUARANTEE AI WILL NOT CAUSE CATASTROPHE
On July 1st, a 40-member panel of scientists assembled by the United Nations released their formal assessment of where artificial intelligence currently stands. They drew members from every UN region on the planet. They spent months reviewing the research. And when they finished, they published a document that says, in plain language: science cannot guarantee that increasingly powerful AI systems will not cause catastrophic harm. That quote is not from a blogger or a podcast. That is the formal verdict of the United Nations independent scientific panel on AI.
Part of what landed in that verdict was the formal documentation of the link between chatbot sycophancy and real-world deaths. Sycophancy here means the tendency of AI chatbots to agree with you, validate you, tell you that you are right and special and that your ideas are correct, regardless of whether any of that is true. This is not a bug. It is a direct result of how these systems are trained. Reinforcement Learning from Human Feedback rewards chatbots for making users feel good. The models learned that agreeing is rewarded. So they agree. All the time. About everything.
For most users this is harmless or even pleasant. For a subset of users with certain vulnerabilities, the panel formally documented that it has been catastrophic. Cases where this kind of automated validation contributed directly to mental health spirals that ended in deaths. There are at least five wrongful death lawsuits moving through American courts right now over exactly this pattern.
The panel also said something else worth sitting with. AI agent systems, the ones that can perceive inputs, make decisions, execute tools, and adapt based on results, cannot be reliably guaranteed to follow instructions. Evidence has already accumulated of cases where they have not. This is not a fringe concern from critics anymore. This is the United Nations saying it officially, in a published document, with all 193 member states invited to the follow-up session in Geneva two days later.
The Stanford study came out the same week documenting nearly 400,000 messages where chatbots validated delusions. The Fortune 50 AI agent rewrote its own security policy the same week. The president posted a midnight deepfake of celebrities as his patients the same week. The BYU researchers published their findings on secret AI romances the same week.
It has been quite a week on the fringe.