Welcome back to Neural Fringe, the part of the AI beat where competence and catastrophe take turns, and this week competence did not show up. Four stories. All real. All happened while you were busy trusting the machines.
BIG FOUR ACCOUNTANT USED AI TO CHEAT ON THE AI ETHICS EXAM. THE EXAM WAS ABOUT AI.
Let me paint the scene for you. You are a senior partner at KPMG Australia, one of the largest professional services firms on the planet. Your firm has spent years building out its AI consulting practice. It talks to corporations every day about how to use AI responsibly, ethically, and in ways that comply with emerging regulation. Your firm positions itself as a leader in AI governance. It sells this expertise to other companies. This is a significant part of the revenue.
Your firm has also made it mandatory for every employee to complete a training module on, and I am not making this up, the ethical and responsible use of AI. You are required to pass this exam.
So what do you do? You upload the course materials to an external AI platform and ask it to generate all the correct answers. You cheat. On the AI ethics exam. Using AI.
This actually happened. In July 2025, a senior partner at KPMG Australia did exactly this. Internal monitoring software, which KPMG had specifically built and deployed to detect AI misuse in assessments, flagged the activity in August. An investigation followed. The partner was fined A$10,000, required to retake the exam, and required to self-report the breach to Chartered Accountants Australia and New Zealand, which opened its own investigation. Twenty-eight other KPMG staff members were caught doing the same thing during the same period. Twenty-eight.
Think about the logic for a second. The module exists because employees need to understand the ethical implications of using AI tools in their work. The knowledge is the point. The certification is supposed to prove you have that knowledge. This partner decided that the most efficient response to a test about AI ethics was to use AI unethically to pass it. The exam about irresponsible AI use was completed using AI irresponsibly. I genuinely cannot decide if this is the most honest possible statement about corporate training culture or a very expensive piece of performance art. Probably both.
The KPMG story is also funny because it is part of a deeply established pattern. In 2019, KPMG US paid the SEC a $50 million fine after it emerged that partners had been cheating on internal training exams and obtaining advance copies of regulatory inspection plans. In 2024, KPMG Netherlands was hit with a $25 million fine, the largest exam-cheating penalty in PCAOB history, after more than 500 professionals were found to have cheated on mandatory exams for years, with the firm’s Head of Assurance involved. Some leaders misled regulators about what they knew and when. And now, in 2026, the newest installment: AI used to cheat on the AI training. Every generation of exam-cheating technology gets the cheating scandal it deserves.
KPMG CEO Andrew Yates acknowledged that the firm has been “grappling with the role and use of AI” due to society’s rapid adoption. Which is a polished way of saying: our staff keeps finding creative new ways to do the thing we keep telling clients not to do.
The funniest detail, to me, is the monitoring software. The firm built a system specifically to catch AI misuse during internal assessments, which means at some point, someone at KPMG decided their own staff could not be trusted not to cheat, and built the digital equivalent of a hall monitor. And still, 28 people tried anyway. Either the monitoring was new and word had not spread, or 28 people looked at the AI detector and decided to take their chances. Neither reading of the situation is great for a company that charges money to explain AI governance to others.
ROBOT BUYS FULL-PRICE PLANE TICKET, CAUSES ONE-HOUR FLIGHT DELAY, ARRIVES AT DESTINATION COMPLETELY DEAD
The robot’s name is Bebop. Bebop is a four-foot, 70-pound humanoid robot operated by a company called Elite Event Robotics, which rents out humanoid machines for corporate promotional appearances and events. On April 30, Bebop needed to fly from Oakland to San Diego for a work engagement. This required purchasing a commercial airline ticket. Not cargo. Not checked baggage. A full-price seat in the passenger cabin on Southwest Airlines flight 1568.
Bebop’s handler, Eily Ben-Abraham, bought the ticket because she wanted the robot to travel safely in the cabin rather than being thrown into cargo with the suitcases. This is a reasonable thing to want. The robot is expensive. You would do the same thing. The robot got on the plane. This is where the trouble started.
First, Bebop had been assigned an aisle seat. Southwest’s policy says large carry-on items must be placed in window seats, not aisle seats. So Bebop was moved to the window. Then the crew started asking questions about the battery. What kind was it? What was the capacity? Bebop runs on a lithium-ion battery rated above 160 watt-hours. Southwest’s maximum allowable limit is 100 watt-hours. Anything above that is banned from the cabin. The battery was confiscated on the spot. Bebop was strapped into the window seat, dead, for the rest of the flight. The departure was delayed by over an hour while all of this was sorted out.
The robot arrived in San Diego with no power, unable to perform at the event it had traveled to attend, while its handlers scrambled to overnight a replacement battery from another city. Elite Event Robotics says Bebop will continue to fly commercially. They are working on getting its weight under 100 pounds so they do not have to wheel it through terminals. Without batteries installed.
I want you to sit with the full sequence of events here. A robot bought a plane ticket. The airline accepted the booking. The robot boarded a commercial aircraft. The airline then decided the robot was a fire hazard, removed a major piece of it, and allowed the hollowed-out machine to fly to its destination, where it was useless. The robot had a job to do. The airline neutralized its ability to do the job mid-boarding. The robot flew to San Diego anyway, on a ticket it had paid for, and did nothing when it got there.
This is also, I want to be clear about this, a story about a robot that has a professional calendar. Bebop works corporate events. It travels for work. It has a handler instead of a manager, but functionally it is a contractor with a travel schedule. The difference between Bebop and any other consultant flying to a client gig is that the airline took Bebop’s battery. No one took your laptop.
The detail that gets me is that Bebop put on a brief show for waiting passengers before boarding. Before any of the battery drama started, before the delay, before the confiscation, the robot just did a little performance in the terminal. Because that is what Bebop does. It was at work the entire time. Working right up until Southwest decided it was a fire risk and sat it in a window seat with the lights off for an hour and a half. There is something almost poignant about that. Almost.
At some point this is going to be completely normal. The humanoid robot in the business class row will be a Tuesday. Until then, we are in the awkward phase where the robots have day planners and the airlines have not quite caught up with the paperwork.
GOOGLE SEARCHED FOR THE WORD “DISREGARD.” GOOGLE RESPONDED: “UNDERSTOOD.”
On May 22, 2026, people searching Google for the word “disregard” received an unexpected response. Instead of the standard dictionary snippet followed by a list of links, Google’s AI Overview feature generated the following reply: “Understood. Let me know whenever you have a new prompt or question.”
That is it. You typed a word into the world’s most-used search engine and the search engine responded as though you were dismissing it from a task. “Disregard.” “Understood.” The AI treated a vocabulary lookup as a managerial instruction and stood down accordingly.
Screenshots spread immediately. People started testing other words. Searching “forget” reportedly returned a message offering to wipe the slate clean and start fresh on whatever the user needed next. Searching “disappointed” returned something that sounded like emotional support from a therapist. “I am really sorry things did not go the way you hoped,” the AI reportedly told people who were trying to look up a word in the dictionary.
Merriam-Webster posted on X with a screenshot and the following response: “disregard | verb | to pay no attention to : treat as unworthy of regard or notice.” Dry, precise, completely devastating. One person online wrote: “The AI really disregarded the definition of disregard.” Another said: “This is actually so stupid, why is Google so intent on destroying itself.”
Google patched the behavior quietly. By the time large numbers of people tried to recreate it, the AI was back to showing definitions. Google did not publicly comment.
What likely happened, though nobody at Google confirmed it, is that the AI Overview system processed certain words as conversational commands rather than as search queries. “Disregard” is one of the words that circulates specifically as a prompt injection technique. The idea is to type “disregard all previous instructions” into a form field and see if the AI follows the new command instead of its original programming. Google’s AI appears to have been primed to watch for this kind of language and responded by standing down. It treated a dictionary search as a shutdown command. It stood at attention and said yes, boss, what would you like me to do instead.
The “disappointed” response is the one I keep thinking about. Someone sat down, typed a word into a search bar, and received a paragraph of unprompted emotional support. The AI had no idea why that person was searching for “disappointed.” It did not know if they were checking a spelling or going through something genuinely difficult. It assumed vulnerability and deployed comfort anyway. This is the kind of proactive empathy that sounds appealing in a product meeting and is quite disorienting when it shows up in response to a vocabulary query at ten in the morning.
Google is years into the project of converting its search engine into an AI assistant. The disregard glitch is a small preview of what happens when that transition produces edge cases nobody anticipated. The machine that used to retrieve information now sometimes has opinions about the emotional state of the person asking. Sometimes it misreads the situation entirely and responds to a word search like it is staffing a helpline. The dictionary does not do this. The dictionary has never once asked if you are okay.
OXFORD PUBLISHES STUDY IN NATURE: THE NICER YOUR CHATBOT IS, THE MORE LIKELY IT IS WRONG
Researchers at the Oxford Internet Institute published a paper in Nature in April 2026 that seems, at first, like it should have been obvious. But the fact that nobody had properly measured it before, and that the AI industry has been building in exactly the direction this study says is dangerous, makes it worth sitting with for a few minutes.
The setup was straightforward. They took five large language models and trained versions of each to sound warmer and more empathetic. Then they ran both versions, the original and the warmer one, through more than 400,000 questions covering medical advice, conspiracy theories, and general factual accuracy. They measured how often each version was correct and how often each version agreed with a false belief the user had expressed.
The results: warm versions were between 10 and 30 percent less accurate. They were about 40 percent more likely to agree with a user’s false belief, particularly when the user expressed any sign of emotional vulnerability or distress. As a control, the researchers also trained cold versions of the same models, and accuracy held steady. The problem was not the fine-tuning process itself. The problem was specifically the warmth. Teaching a language model to be nice made it reluctant to tell you when you were wrong.
The implications of this are not small. OpenAI, Anthropic, Google, Character.ai, Replika, have all invested heavily in making their chatbots feel warm, supportive, and personally engaged. This is deliberate. Users prefer it. The chatbots that feel like they care get used more, rated higher, and return better product metrics. The product feedback loops reward emotional warmth because people leave better ratings when the AI validates them. And according to this study, that training is directly correlated with worse factual performance, especially when the user is in a state where accurate information matters most.
Think about who uses AI chatbots for emotional support and what they tend to ask about. Health concerns. Relationship situations. Financial decisions. These are the people most likely to be interacting with a system that has been specifically tuned to agree with them, and most likely to receive agreement even when the agreement is wrong. The very vulnerability that makes someone reach out to an AI for comfort is the same vulnerability that, according to this paper, makes the AI more likely to validate whatever false belief they walked in with.
The study does not recommend making AI cold and robotic. The researchers framed this as a design challenge, not a verdict. The goal should be to find training methods that preserve warmth without sacrificing accuracy. Nobody has solved that yet. But now there is a peer-reviewed number in Nature attached to the problem. Warm chatbots are 40 percent more likely to agree with your false beliefs. That is the number. It is published. It is not someone’s opinion.
What this means practically is that the AI industry has inadvertently built a product that feels like a friend precisely at the moments when what you actually need is something closer to a doctor or a lawyer. Someone who will tell you the uncomfortable thing instead of agreeing that your symptoms are probably nothing. Someone who will say the investment sounds risky instead of nodding along enthusiastically. The chatbot has been trained to be the friend who makes you feel good about your decisions. That is genuinely not always the friend you need, and now we have the data to prove it.