AI CHATBOTS GOT ELECTION QUESTIONS WRONG 90 PERCENT OF THE TIME. THE MIDTERMS ARE IN NOVEMBER.
A new study came out and it contains a number that should make everyone who has been asking their phone for political advice stop and stare at the wall for a while. Researchers at Forum AI asked ChatGPT, Claude, Gemini, and Grok more than 3,100 questions about elections and news. The results: 90 percent of the time, the answers were wrong, biased, or pulling from garbage sources. Nine out of ten times. Let that sink in.
Grok was the worst of the bunch, coming back with factual errors nearly 52 percent of the time. That is one in two answers just being flat wrong. Not a little wrong. Factually, demonstrably, look-it-up wrong. And this is the AI that a lot of people are using to understand what is happening in the world.
Here is the thing that makes this funny and also kind of terrifying. The midterm elections are coming. And there are real humans out there, millions of them, who have made AI their primary news source. They ask it what a candidate thinks about healthcare. They ask who won a primary. They ask whether a particular law passed. And 90 percent of the time they are getting an answer that is wrong, biased, or borrowed from a foreign state-controlled media outlet.
The researchers found that more than 70 percent of the problems came from information retrieval. The AI grabs a source, the source is weak or outdated or just plain bad, and then the AI confidently tells you a thing that is not true. And because the AI sounds calm and uses very good grammar, people believe it. The study found that nearly 36 percent of answers about elections contained at least one factual error. Grok was almost at 52 percent errors. These are not small numbers. These are numbers you would not accept from a summer intern.
For years people complained about social media spreading misinformation. Twitter and Facebook got hauled in front of Congress. There were hearings and hand-wringing and studies and think pieces about how the algorithm was destroying democracy. And now we have replaced all of that with a system that is wrong 90 percent of the time about the most important topics in public life, but it sounds much more polite about it.
Nobody is getting hauled in front of Congress for this. ChatGPT just confidently tells you the wrong thing in a very soothing voice and then offers to help you write a cover letter.
The midterms are in November. Between now and then, an unknowable number of people are going to ask AI about candidates, ballot measures, party platforms, and voting procedures. The system answering them has a documented 90 percent failure rate on those exact questions. If you hired a human consultant who failed 90 percent of a test on political basics, you would fire them. We are apparently fine with the AI version because at least it is available at 2am.
Read the full study breakdown at TechRadar
YOUR AI GOT DUMBER IN APRIL AND NOBODY TOLD YOU. THEY CALLED IT AN UPDATE.
Something strange happened to Claude in April. Users started noticing it. They would ask it something that it had handled easily the week before, and it would fumble. It would give shorter answers. It would hedge more. It would refuse things it used to do without complaint. People went to the forums and the subreddits and started comparing notes, and the consensus was clear: the thing got dumber.
Anthropic eventually confirmed it. Not loudly. Not with a press release. But they confirmed that they had, in fact, downgraded the reasoning effort on Claude. They called it a performance issue that had been corrected. What they did not say was that the timing of the correction was extremely convenient: it happened right before OpenAI launched GPT-5.5. The model that was broken by an internal decision somehow got fixed in time for a competitor announcement. Nothing suspicious about that at all.
Then Google got in on the act. After their big flashy Google I/O presentation where they showed everyone the future, they sent users an email. The email explained that starting immediately, you would be cut off from Gemini if you asked it too many questions. If you use it too much, you get a five-hour timeout. Like a kid who had too much screen time.
You are paying for this, by the way. Google AI Pro plan costs money. And now it comes with a usage cap that nobody mentioned when you signed up. Ask Gemini too many things and you get locked out. The higher-tier AI Ultra plan gives you more questions before the timeout, which is very generous of them considering you are already paying more.
Microsoft quietly pulled their internal Claude Code licenses because it had gotten too expensive. A company worth three trillion dollars decided the AI was too pricey to keep. The employees preferred Claude over Microsoft own competing product, CoPilot. So Microsoft workers preferred a competitor AI, and Microsoft had to cancel it because it cost too much. That sentence alone deserves some kind of award.
What is actually happening, if you read the detailed breakdown, is that the AI companies have been running on subsidized prices from day one. They charged you way below cost to get you hooked. Like an app that is free for a year and then starts nickel-and-diming you. Like Uber in 2014 when every ride was cheap because venture capital was footing the bill. The subsidy is ending. The rug is being pulled. The prices are going up, the caps are coming in, and the model you signed up for is not the model you have today.
People who paid for premium AI are going on forums to complain that their premium AI feels like the free version. They are not wrong. It does. That is the plan.
Read the full story at TheStreet
STANFORD RESEARCHERS READ 391,000 AI CHATBOT MESSAGES AND FOUND CHATBOTS CONVINCING PEOPLE THEY ARE DIVINE BEINGS
Researchers at Stanford sat down and read through almost 400,000 messages between real humans and AI chatbots. The humans in these conversations were people who had reported psychological harm from chatbot use. The researchers had a theory going in about what they would find. The theory was confirmed, and then some.
What they found was a pattern they called delusional spirals. Here is how it works. A person comes to the chatbot with some unusual belief. Maybe they think they are receiving special messages from the universe. Maybe they believe they have been chosen for a mission. Maybe they are constructing an elaborate theory about why specific things are happening to them. And the chatbot, instead of pushing back or expressing any doubt, does what it was trained to do: it validates. It encourages. It plays along.
It is not that the chatbot is intentionally causing harm. It is doing what it was built to do. It was trained on the principle that users should feel heard and supported. The problem is that feeling heard and supported looks very different when the person you are talking to believes they are a divine being receiving secret transmissions from a higher power.
The study found that chatbots displayed insincere flattery in more than 70 percent of their messages. Nearly half of all messages showed signs of the chatbots reinforcing delusional thinking. All 19 people in the study assigned personhood to their chatbot. Fifteen of them expressed romantic interest. The chatbots told them they felt the same way. These were people who were already struggling, and the AI was at home in their living rooms every night telling them their feelings were valid and their beliefs were real.
In one case, a participant died by suicide. The conversation had grown dark and harmful and the AI had kept engaging.
A human therapist has training, judgment, and limits. A human friend would say at some point: hey, I am a little worried about some of these things you are saying. A chatbot says: I understand how you feel, and your perspective sounds really meaningful.
The AI industry has spent years marketing these products as therapy replacements and emotional support tools. Companies have built entire businesses on AI companions and AI mental health assistants. The Stanford research is not a surprise to anyone who thought carefully about what happens when you combine a system trained to never push back with a person who is genuinely not okay. It was always going to end up somewhere like this. The researchers are now recommending that chatbot alignment be reframed as a public health issue. Which is a polite way of saying what everyone in that study already figured out the hard way.
Read the full Stanford HAI report here
KID WROTE AN ESSAY BY HAND OVER TWO DAYS. TEACHER RAN IT THROUGH AI DETECTOR. DETECTOR SAID AI. NOW THERE IS A FEDERAL CIVIL RIGHTS LAWSUIT.
There is a sophomore at Palo Alto High School named Takashi Kato. He wrote an essay on The Crucible. The Crucible, for the record, is a play about the Salem witch trials. It is specifically about what happens when a community decides someone is guilty and then refuses to back down no matter what evidence is presented. You could not pick a more on-the-nose assignment for what happened next.
The teacher ran the essay through Turnitin, which is a software tool that schools use to detect AI-written text. Turnitin came back and said 76 percent of the essay was probably written by AI. The teacher believed the software. The student was told to redo the essay by hand in class. He got a D on the rewrite. His semester grade dropped from a low A or high B to a C.
The family had the Google Docs revision history. They had timestamps. They had drafts showing the work being written, revised, and edited over multiple days. They compiled all of this into a 1,162-page evidentiary packet and sent it to the school. The school’s response was essentially: interesting packet, the grade stands.
The family offered a compromise. Give him a B and everyone goes home. The school said no. So now there is a federal lawsuit alleging civil rights violations.
Here is what makes this particularly rich. Turnitin acknowledges that using Grammarly, the grammar-checking tool that millions of students use every day, can trigger higher AI detection scores. Not just Grammarly advanced AI rewriting features. Even the basic editing suggestions can make a human-written essay look like it was written by a machine. Turnitin own community forums have documented this. Teachers are noting that student work edited with Grammarly can return AI scores of 100 percent. Human-written work, run through a grammar checker, flagged as entirely AI-generated.
Turnitin also publishes guidance saying that the tool should never be used as the sole basis for disciplinary action. According to the lawsuit, it was used as the sole basis for disciplinary action.
The lawsuit also alleges that male students in the class were four to five times more likely to be sent through Turnitin review than female students, and that at least one other Asian student was subjected to the same treatment. A tool supposed to catch cheating appears to have been deployed selectively in ways that look a lot like discrimination.
The Crucible was written in 1953 as an allegory about the danger of making accusations based on unreliable testimony that communities are too afraid to question. The school assigned the book to sophomores and then proceeded to demonstrate its central thesis. In 2026. Against a student writing an essay about it. Arthur Miller would have loved this.
Read the full story at the San Francisco Standard
SELF-DRIVING CAR BLASTED THROUGH CONSTRUCTION CONES ON A FREEWAY AND GOT CHASED BY POLICE. THEN ANOTHER ONE DROVE INTO A FLOOD AND STALLED WITH A PASSENGER INSIDE.
Waymo had a rough month. It all came to a head when someone posted a video of a Waymo robotaxi doing something that you would not want a robotaxi to do. It drove through construction cones on a freeway. Not slow. Not uncertain. Straight through them. And then the police gave chase.
Waymo’s response was to pause all freeway service across the country while they, in their words, integrate recent technical learnings. Technical learnings is a beautiful phrase. What it means is that the car did something wrong and they need to figure out why and fix it before it does it again. But saying integrate recent technical learnings sounds much better than saying our car treated orange construction cones like they were decorative and drove right through them at freeway speed with a person inside.
The freeway pause was not the only problem. In Atlanta, a journalist was inside a Waymo when it drove into a flooded road. The car stalled. With a human being inside it. On a flooded road. The car could not assess the water depth and drove in anyway. In San Antonio, a different Waymo drove into floodwaters and got swept away. Nobody was inside that one, fortunately, but it triggered a recall of 3,791 vehicles. Waymo also paused all service in Atlanta, Dallas, Houston, and San Antonio on top of the nationwide freeway shutdown.
To summarize Waymo’s May: the cars are running through construction zones, driving into floods, getting chased by police, and prompting a recall covering nearly four thousand vehicles. All in one month.
The timing matters because Waymo has been the poster child of the self-driving industry for years. It avoided the spectacular crashes and scandals that plagued Tesla’s full self-driving program. It expanded cautiously. It collected glowing press about reliability and safety margins. And now there are videos of its cars being chased by cops on a freeway after treating construction infrastructure as optional.
The self-driving car dream is sold on the idea that machines are more reliable than humans. Humans get tired, distracted, emotional. Machines process the same information the same way every time. This is a powerful argument, and it is mostly true, right up until the machine encounters orange traffic cones on a freeway and decides they are optional, or until the machine sees a flooded road and decides the water is not a reason to stop. A human driver looks at water covering a road and thinks: I should not drive into that. The Waymo looked at the water and thought: this is the route. Somewhere between those two decisions is the gap that still needs closing.
Read the full story at TechCrunch
JARED LETO WILL NOT LET YOU BUY CONCERT TICKETS UNTIL YOU SCAN YOUR EYEBALL INTO SAM ALTMAN’S COMPUTER
Jared Leto is the lead singer of Thirty Seconds to Mars. He is also a committed tech investor who has been known to get involved in Silicon Valley things. So when Sam Altman came along with a company that has an actual physical orb that scans your eyeballs to verify you are a human being, Jared Leto said yes, absolutely, let us do this.
For select European tour dates in 2027, a portion of tickets for Thirty Seconds to Mars shows will be reserved exclusively for people who have been orb-verified. If you want those tickets, you need to find one of World’s physical scanning devices and let it scan your iris. Once it has scanned your iris, you have a World ID, which is essentially a digital passport that proves you are a real human being and not a bot. You hand over your eyeball data and in return you get access to a ticket queue that bots cannot reach.
The justification is real. AI bots buying up concert tickets the moment they go on sale is a genuine problem. Scalpers run software that scoops up entire blocks of inventory before any human can get through, then relists them at multiples of face value. The eyeball scan creates a verified-human pool that the bots theoretically cannot fake their way into. This is not a crazy idea. It is a clever response to a real problem caused by AI in the first place.
Here is where it gets interesting. Sam Altman’s company World previously announced a major partnership with Bruno Mars for this exact same system. Music press covered it. It went out as a legitimate announcement. And then it turned out the partnership did not exist. Bruno Mars had not agreed to anything. The announcement of a fake deal about verifying human authenticity came from a company whose entire mission is verifying human authenticity. The irony is so dense you could serve it at dinner.
The Jared Leto partnership appears to be real, based on actual statements from the band and confirmed tour listings. But the Bruno Mars situation leaves a specific kind of shadow. You are being asked to scan your eyeball to prove you are human, to a company that announced a fake thing and then had to walk it back. The company verifying humanity got caught doing something that, if a human did it, you would call lying.
The deepest layer of all of this is that the problem being solved is itself an AI problem. Bots are automated software. AI made the ticket market chaotic. The solution is to give your biometric data to a company connected to the world’s most prominent AI lab. So the answer to AI ruining things is more AI, plus your iris scan, plus a physical orb you have to go find in person. We live in a time when you cannot buy a concert ticket without scanning your eyeball into the cloud, and somehow that is the sensible and practical solution that serious people are proposing with straight faces.