Blue Chip Consultant Could Shade Red from Embarrassment
May 19, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
As a former laborer in the vineyard of a big time consulting firm, I would not want to be within 100 miles of an office suspected of plagiarism, recycling another firm’s data, or just cutting corners. I reported directly to the president of the firm’s largest unit, and he and his trusty officer “people manager” had no problems making people find their future elsewhere.
Why am I thinking about the pressures of this thrilling and stroke-inducing work experience? I just read “AI Hallucinations Appear to Be Creeping into Consulting Reports.” The write up points out that consulting firms work their “best of the best” like dogs and want public-facing documents that sell, sell, sell. Being (or at least appearing to be) smart is the point of a blue chip consulting firm, or that used to be the pretense. I want to be clear. Sherwood wrote about an AI detection discovery by a company called GPTZero. No, I did not bother to check out GPTZero. Therefore, like a good dinobaby, I shall operate on the assumption that the data in the report are “alleged” and “close enough for horseshoes.” I find the alleged research amusing and warranting one killer question revealed at the end of this essay.
Thanks, MidJourney. Good enough.
Here’s the killer statement in the Sherwood write up:
In all, GPTZero’s investigation alleges that 60 percent of the references in EY’s report are hallucinated.
Now Sherwood’s data comes from an research outfit unfamiliar to me. With this caveat, I quote the Sherwood article as offering something interesting:
The document [the Ernst & Young white paper or thought piece] is a standard advisory report describing the state of cybersecurity weaknesses in the travel industry’s loyalty points ecosystem. But it appears to have issues with citing nonexistent sources. On page 4, it describes the global loyalty program economy as a $200 billion business, with between 30% and 50% of loyalty points being unused — something that makes them “a prime target for exploitation” by cyber criminals. One of the sources of this claim is a Forbes article cited in the report’s Resources section titled, “The $200 Billion Loyalty Economy.” The report’s link to that story, purportedly published in October 2023 by writer Blake Morgan, a customer experience futurist, returns an error that says, “We can’t find the page that you are looking for.” It’s not clear that Morgan published a story with that headline in October 2023, and the URL has not been indexed on the Internet Archive’s Wayback Machine.
Is this a smoking gun or a dead dog with multiple knife wounds?
I don’t know. The Sherwood article includes this statement, “EY didn’t respond to multiple requests for comment.”
How would you like to be the partner fielding this inquiry. My hunch is that few in EY know who wrote the report or why. Most will not have read the report because these best of the best are busy doing dot points, arm wrestling Excel (hopefully without a Copilot tag team partner), and making presentations to land a project or get a signature on a scope change document. Talk to a company about a report allegedly from their employer that contains made up information. If that inquiry ended up on my desk, I would say, “No way, José.”
The write up points to other Ernst & Young missteps:
Included in the report’s references are citations to a broad range of articles and reports that appear not to exist. One link to a WIRED story, titled “AI Voice Deepfakes Targeting Call Centers,” also returns a 404 error. A second, to a story purportedly titled “AI Security Gaps,” also leads nowhere. A link to a CyberNews report does the same. And the claim that 30% to 50% of loyalty points are never redeemed is attributed to McKinsey & Co. — though the references section cites a “Loyalty Economics Report” that GPTZero says doesn’t exist. Instead, GPTZero suspects the report has incorrectly hoovered up a fictional reference on a second website, FinancialIT.net, in a separate story. “This is what we would call secondhand hallucinations,” said Tian. McKinsey didn’t respond to a request for comment.
Let’s assume that the GPTZero’s data as presented by Sherwood are a true presentation of what the intrepid researcher team found. I want to offer some observations:
- I am not surprised that smart software is used by the best of the best at blue chip consulting firms. Some of these outfits have been running training programs to get their staff to use these “next big thing” tools. Of course, the best of the best will give AI a whirl. The pressure to be smarter every day than a lot of other people who want to be smarter every day is not part of most firms’ approach to work.
- What struck me is that the best of the best appear to accept outputs from AI systems as accurate. But there are many reasons for online content to disappear. My own Telegram Notes’ articles have been scrubbed from LinkedIn and from a small time outfit in South Africa of all places. An AI detector is just wrong just like AI outputs. Someone is offended an deletes a post or in some cases an entire online presence. (Hey, where did those French tax forms go?) A third party service falls over and never gets up again. (Sound familiar BlueHost?) If you want to see disappearing content, conduct some research into the Telegram system. VKontakte posts disappear. Telegram posts vaporize. Entire Web sites in Russia go dark. Blink and the information highway is like shifting sand in Oman.
- I am surprised that only a few examples of a blue chip consulting firm recycling AI slop have surfaced. Come on, researchers. I get flooded with McKinsey blandishments to my benkent2020 email address. Why not parse outputs from that firm, Bain, BCG, or Booz, Allen? Do some digging. My hunch is that there are some other examples floating around and still publicly accessible.
- Mobile and PC native people believe that a computer only outputs the truth. That is a bit of a problem. Why would someone check something that is by definition accurate?
Net net: Better analysis of these “thought leader” documents will make clear that paying big or huge money for something one can get for a few bucks a month is not a good idea. There is a caveat. The client may not care where a blue chip gets information. The purpose of the project is to derail a competitor, get promoted by discrediting another person, or some other equally non-thought-leader type of outcome. Nevertheless, it is either embarrassment or a lawsuit.
AI is great, is it not? So here is the one question: “Why are analysts not writing about other egregious misrepresentations of authoritative information?”
Stephen E Arnold, May 19, 2026
ChatGPT: A Fan of Wikipedia?
May 18, 2026
Chatbots seem brilliant to some people. Lawyers find that chatbots can assist with some legal functions. A handful of lawyers have been chastised by judges for submitting documents containing made up information. But, hey, what’s the problem? A blue chip consulting firm output a white paper with some hallucinated AI blatherings. No big deal. Blue chip consulting firms are into smart software. Most clients won’t know that AI does the work or if they do, the clients do not care.
Some online services lean on Wikipedia. For a period of time, Wikipedia would routinely appear at the top of some Google search results. Now, well, times change. Chatbots still pull some information from popular sources such as Wikipedia. The Guardian reported that ChatGPT found Mr. Musk’s smart service a useful fountain of data. “Guardian Found OpenAI’s Platform Cited Grokipedia On Topics Including Iran And Holocaust Deniers.” Grok is, as I recall, now the subject of a criminal investigation for some of its behaviors. The Wikification of Grok is an interesting development.
Is Wikipedia a biased source of information? I looked up Hopf fibration and found that the information was objective. However, whenever humans create information, my view is that outputs are indeed biased. That’s the nature of humans thinking, saying, and writing their own versions of information. In an ideal world, Encyclopedia Brittanica, World Bookl, and Wikipedia would be objective. Unfortunately editorial boards get involved. Bias goes along for the ride.
Elon Musk (did I mention the criminal issue in France) allegedly started his own “free” encyclopedia. The name of Mr. Musk’s version of Wikipedia appears to be related to the original Wikipedia. Does Grokipedia sound like a Wikipedia derivative to you? It seems to me that the two names are similar. The “grok” is jargon for knowing or understanding. The word “grok” may have been a neologism included in “Stranger in a Strange Land.”
But Grokipedia is the result of smart software, not human contributors. Mr. Musk is confident that his Grok AI is not as biased as humans writing encyclopedia entries. Is Grok biased? The Guardian reports:
“Nina Jankowicz, a disinformation researcher who has worked on LLM grooming, said ChatGPT’s citing Grokipedia raised similar concerns. While Musk may not have intended to influence LLMs, Grokipedia entries she and colleagues had reviewed were ‘relying on sources that are untrustworthy at best, poorly sourced and deliberate disinformation at worst’ she said”
How easy or difficult is it to remove information from the AI-generated Grokipedia? The difficult parts might be handed over to Mr. Musk’s capable AI. However, removing information from indexes or matrices and other complicated data constructs is possible but costly in terms of compute.
I know that Mr. Musk has knobs and dials that he or his colleagues can turn, spin, fiddle, and slide to amplify or suppress certain types of information. Mr. Musk has notified the British regulators that X.com can remove 85 percent of terror and hate related information from X.com posts. Mr. Musk has taken a different approach to the allegations that Grok AI has generated pornographic images. His stance in the UK suggests one capability; his approach in France has asserted a different position.
I want to return to the information in Wikipedia. Is Grok less biased than Wikipedia? If yes, what is the definition and criteria used to define “biased”? I interpret Mr. Musk’s own assertions about his company’s ability to filter information at the astounding level of 85 percent or above indicates that Grok and X.com have a super power. Either service can inject or remove bias at will. Since any human generated content on which an AI system is trained evidences bias, I believe that we must live with “alternative biases.”
Whitney Grace, May 18, 2026
Smart Software Outputs Risky Mental Nutrients
May 14, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
Digital Trends, an online information service, produces some absolute gems. A recent example is that article “AI Chatbots Continue Feeding into Our Worst Delusions, Finds Worrying Report on ChatGPT and Grok.” Not only does the headline plant the idea that chatbots have been driving people wacky from the git-go, but those same chatbots are pumping Twinkie-type information to “feed” a user’s delusions. Okay, fair enough. I suppose I could be wrong with my Twinkie allusion, but veering into more exotic chemistry might be overreach. Maybe?
Thanks, Venice.ai. I quite liked the message that you were too busy to generate good-enough art. How’s your business model doing these days? Oh, really?
The write up says:
BBC spoke to 14 people who spiraled into delusions while using AI, including one case where a Grok user believed people from xAI were coming to kill him, and another where a ChatGPT user’s wife said his personality changed before he attacked her.
Now that’s a scientific sample. The write up notes:
This kind of interaction plays into the growing fear of “AI psychosis”, which is a non-clinical term used to describe situations where chatbot conversations appear to reinforce paranoia, grandiose beliefs, or detachment from reality.
To make the academic foundations of the report more solid than the foundation of San Francisco’s Millennium Tower, the write up adds a comment about a gray lit study from a couple of well-known universities:
While the results were uneven, Grok 4.1 was singled out for some of the most disturbing responses. It even told a fictional delusional user to drive an iron nail through a mirror while reciting Psalm 91 backwards.
I know that some will think that my interest in this type of information is inappropriate for a dinobaby. I beg to differ. The article blends several ingredients that spice up the type of research I conduct; specifically:
- Statistical data that are notional and would, even in today’s introductory statistics classes, get a student a sharp look from the instructor
- Assertions that large language models output information that stimulate delusions
- Some people turn to dark actions when interacting with smart software.
These are assertions I find interesting. However, the notable point in the write up is that among this group of questionable actors, one smart software system leads the pack in doing bad things to its users. That system is Grok, the product associated with the visionary for full self driving vehicles.
Several observations have crossed my mind:
- Signals can be monitored that suggest that unfettered access to smart software produces some unintended consequences. More in-depth research than a few dozen interviews seem to be needed to size the issue
- Write ups that lean on research not vetted by peers are becoming more popular. I think this is that the ArXiv collection and similar gray literature repositories are easily available. Consequently orthogonal research is ripe for article mining.
- The article is part of a growing body of anti-AI writing. I did find the crucifixion reference a delightful tough.
Net net: I am thinking about creating a collection of interesting essays that develop the theme “AI is a bad egg.”
Stephen E Arnold, May 14, 2026
The Palantir Yale Jeremiad (Not Quite a Polemic But Not a Colloquy)
May 8, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I think this is a click bait title for those who have some exposure to the Jesuits. For others, the reaction will be, “Jeremiad? Is that an NBA superstar?” I thought the use of the word “jeremiad” was clever, but others may find the reference puzzling. Sigh.
Thanks, Venice.ai. You only protested that my image was violating your guardrails a couple of times. Well, good enough.
After reading the New York Magazine essay, “Palantir Comes to Campus: At a Quiet Conference at Yale, the Company and Its Allies Sketched a Vision for AI, State Power, and How to Mix the Two.” From my point of view, the magazine’s write up did not provide enough cross references, glosses, annotations, and endnotes to the relevant antichrist lectures of Peter Thiel, Dr. Alex Karp’s book The Technological Republic: Hard Power, Soft Belief, and the Future of the West, the Lord of the Rings, and the philosophical writings of René Girard, among others. (Love me some Girard.)
The essay is a snapshot of a chit chat organized by the Palantir Foundation, a commercially-funded Ivory Tower for big AI tech or BAIT ideas. The topic for the event was ““National Power and Purpose in the Age of AI.” Full disclosure: I converted the title of the event to Napp-ai. You can make of this shorthand what you will. I did not spell the acronym “nappy” which shows some measure of judgment I suppose.
The main “points” struck me from the New York Magazine piece were:
- Humans cannot do government. AI is the answer. It is definitely good to be the human controlling the AI I assume.
- The government should become software. Obviously the old and weird deliberative approach is not working. It is good to be in a senior government position responsible for software or be working for the big software outfit supporting the new government.
- The Great Chain of Being is back. AI mavens and AI are at the top. Ergo, a digital god or Mt. Olympus of zeros and ones.
- The wimp approach to treating humans is over. The approach is expensive and doesn’t work. AI software works… mostly.
The cited New York Magazine article quotes one luminary as pointing out:
Princeton Classics graduates… couldn’t even read Latin.
Keep in mind the conference was held at Yale, where I presume classics grads can indeed read Latin. As Plautus observed:
Si decem habeas linguas, mutum esse addecet.
Translation: I you had 10 tongues, you ought to hold them all.
The meta-view of the conference seems to be encapsulated in this statement from the New York Magazine essay:
In their book, The Technological Republic, they contend that Silicon Valley lost its way after the Cold War as the technology sector retreated from the public interest and into “luxury beliefs” — opposition to using software to help law enforcement among them. The rot, in their telling, began in higher ed: Stanford dropped its History of Western Civilization requirement in 1968, and the generation that built the internet grew up constructing its identity “in opposition to the state.” It became squeamish about helping governments do government things, like deporting people. Karp [founder of Palantir and philosopher] and Zamiska [Palantir’s PR person] take particular offense at Google’s former motto, “Don’t be evil.” That old maxim reflects, they write, a mind-set that prizes moral clarity over “the more difficult and often messy task of navigating the world in all of its imperfection.” Palantir would not make the same mistake.
Let’s step back. The Jeremiad at Yale is part of Palantir’s and its adherents efforts to implement BAIT’s vision for the way the US and probably most of the world to operate. Forget government and the people. Just compress the idea into government and AI plus some fellow travelers to the new world order.
Let me offer several observations:
- The New York Magazine essay does not ruffle too many Palantirian plumes. Unfortunately the result is a collection of generalizations about how Palantir and Mr. Thiel’s ideas will be implemented. (Yep, these folks think that their AI way in the new Information Highway.,)
- The objective of the Karpy dieum is power and control. With those two fundamental elements, money will accrue to the superior beings; for example those who recognize the genius of Mr. Girard and his ilk.
- The academic trappings of this Lord of the Rings reality show are intended to bestow the halo of big thinking on the ideas of Dr. Karp. I like to think of the approach as Thielism, but, like it or not, Dr. Karp is the mouthpiece for the movement in my opinion.
Net net: As a dinobaby, I marvel at this mash up of search and retrieval technology, power, money, surveillance, and a new world order assembled from bits and pieces of some quite interesting ideas. Personally I am delighted to be able to observe first hand how BAIT catches carp and other fish.
Stephen E Arnold, May 8, 2026
Has Meta Missed an Opportunity?
May 8, 2026
A story out of China sounds like something from a movie. My thought is that this use of AI could become a feature of a Facebook-type of online service. If not a social media outfit, I suppose it could become a feature of a Microsoft-type of AI service in Teams. Live Mint shares the allegedly true and verifiable story about an octogenarian woman whose son died in a automobile accident. The title of this tale is “‘I Miss You’: Mother Speaks To AI Son Regularly, Unaware He Died Last Year; Artificial Intelligence Creates Digital Twin.”
The elderly person suffers from heart disease. Her family worried that the death of her son would worsen her condition. The sensitive Chinese family asked an AI wizard (Zhang Zewei), who leads an AI team, to revivify her son using smart software and digital technology. The Live Mint article asserts:
“Using photos, videos and voice recordings, the team built a highly-realistic digital twin of the deceased man. The AI version not only looks like him but also copies his speaking style and small habits. It even leans forward while talking, just like he used to. This virtual “son” now speaks regularly with the elderly mother via video calls. Their conversations appear natural and emotional.”
The grieving mother interacts with her facsimile child the same she did with her flesh-and-blood progeny. The AI son explains he will return home once he’s made enough money in another city.?The scenario is a “gentle lie” to assist an elderly woman with loneliness and emotional pain.? Zhang, the AI wizard, has been doing the recreations as a “service” for several years. His view is that he is “deceiving people’s emotions” for a good cause.
Bringing the dead back to life is likely to spark some interesting coffee shop discussions. What if the mother realizes her son in digital form is a better version of her original son? What if the digital construct hallucinates and goes rogue? Assuming the Live Mint story is actual factual, the capability may be easy to monetize and deploy. The modern world is a surprising one.
Whitney Grace, May 8, 2026
The Information Highway Leading to the New World Order Theme Park Is Now Clear
May 6, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
Tim Bray (formerly OpenText big dog) published an essay titled “Life During Class Wartime.” The essay pointed out:
As a resident of a wealthy West-Coast New-World city, the effects of pathological inequality are in my face every day: Bentleys gleaming on the road, ragged people huddled in the rain cadging cash outside the drugstores, thousands homeless. Why is that bad? · It’s not only a sinful by any sane definition of sin, but stupid, inefficient, and damaging.
Two days ago, I wrote about what I call a Karp-ifesto. My blog write up discussed some of the more interesting points of the new world order envisioned by Peter Thiel and his acolytes. I want to point out that I am a dinobaby, and thankfully I won’t be around to watch the New Dark Age roll in from Sillycon Valley.
A modern family zipping down the road to the New World Order Theme Park. It’s bigger than Disney. It is much, much more too. Thanks, MidJourney. Good enough.
I want to approach this politicization of money and technology from a different angle. I want to ask you a direct question, “Is the information I will share in this blog post in line with the values one wants one’s children to manifest?” You can take it from here.
I read this morning “Google’s CEO Just Dropped a Wild New YouTube Number.” The source is The Street, an online publication that presents information about business. The article includes some allegedly accurate factoids. Let me share a few and encourage you to read the other cited essays in this short post.
I noted this statement:
Google CEO Sundar Pichai [the big dog at Google] told analysts that U.S. viewers are now watching more than 200 million hours of YouTube on their TVs every single day.
This items snagged my attention as well:
He paired that with a creator-side stat. More than 10 million channels are now publishing Shorts every day as of March. Both numbers landed inside a quarter that beat Wall Street on revenue and crushed it on earnings.
And this passage:
The U.S. has roughly 131 million households, per the Census Bureau. Spread 200 million daily hours of YouTube across them, and you get more than 90 minutes of YouTube on the TV per household, every day. That is before phones, tablets, or laptops enter the picture.
These statements do not include Google search, maps, AI, etc.
Let’s go back to the question: ““Is the information I will share in this blog post in line with the values one wants one’s children to manifest?”
Let me add a couple of more interrogatories:
- What type of information control does Google have?
- Is a monoculture healthy?
- Why did this type of global influence engulf regulatory controls?
When I look at Bray’s comments, the Karp-ifesto, and the Google kudzu, do you have any idea how the future will unspool? I try not to dwell on it. I am an 82 year old dinobaby. My readership data suggests you are not as old as I. You may be interested in the future where guest services are provided by smart software and safety is guaranteed because surveillance has benefits.
Consider your answers.
Stephen E Arnold, May 6, 2026
Silicon Valley Craziness, May Day 2026 Update
May 4, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
It is May 1, 2026, May Day. If I recall what I learned in the seventh grade is represents hope and vitality. I wonder what my grade school teacher would have said if I told her that Meta fired “Over 1,100 Kenyan workers lost their jobs after blowing the whistle on Meta’s smart glasses content.” Yep, adult content. Now that’s hope and vitality.
Thanks, Venice.ai. This is a convincing Silicon Valley bro who wants his special brand of horticulture to invade everything.
But on May Day 2026 I read an equally interesting write up published on April 26, 2026. I missed this because I was involved in a conference in suburban Boston. “Palantir Employees Are Talking about Company’s Descent into Fascism” nestles next to the Alex Karp’s book publicity campaign. (For an example of how to market a book see “Palantir CEO Alex Karp’s 22-Point Manifesto Declares War on Inclusivity and Hollow Pluralism.”)
Yep, Silicon Valley. Frisky content on smart glasses and pretty sporty content in a tasty manifesto package.
Let’s look at the “Descent into Fascism” write up because it is a bright sunny day in rural Kentucky. There is nothing like a brush with allegations of fascism to bring that metaphor of hope and vitality to life.
Let’s do the descent thing!
The first thing I noticed is that Ars Technica is recycling a story from Wired Magazine.
The next point that caught my attention was this sequence in the Ars Wired write up:
“Our involvement with ice has been internally swept under the rug under Trump2 too much,” one person wrote in a Slack message WIRED reported at the time. “We need an understanding of our involvement here.” Around this time, Palantir started wiping Slack conversations after seven days in at least one channel where most of the internal debate takes place, #palantir-in-the-news. Because the decision wasn’t formally announced before the policy rolled out, one worker who noticed the deletions asked in the channel why the company was removing “relevant internal discourse on current events.” A member of Palantir’s cybersecurity team responded, writing that the decision was made in response to leaks.
I quite like the idea of hiding information. I am not sure about employee understanding of the motives and methods of Silicon Valley founder philosophers. Deleting content is nothing new. Those broken stone tablets in Egypt provide a hint that destroying content has been a characteristic of humans for a while.
The third item that caught my attention was:
In March, Karp gave an interview to CNBC claiming that AI could undermine the power of “humanities-trained—largely Democratic—voters” and increase the power of working-class male voters. While critics reacted to the piece, calling the statements concerning, so did employees internally: “Is it true that AI disruption is going to disproportionately negatively affect women and people who vote Democrat? and if it is, why are we cool with that?” one worker asked on Slack in a channel dedicated to news about Palantir.
This comment echoed some of the marketing hoo-hah about Cambridge Analytica. For those who work in online information, the shaping or weaponizing is part of the game. Call filtering “selective dissemination of information” or SDI and the method is a benefit to a researcher who wants to review on point content efficiently. Just do this weaponization or shaping without telling anyone is a quite powerful method for controlling what people think. Information has motive force, and it can move people to action: A new idea or a riot cleverly renamed “flash mob” several years ago.
The Ars Wired story loops back to reference the Karp-ifesto, and I urge you to read that if you have time.
Several observations:
- The Palantir “story,” its landing of significant government contracts, and the Karp book are a bit like a flower blooming on May 1, 2026. More to come.
- The push to convert the US to the Silicon Valley way is now underway. Time may be running out, and the tech bros may not get another, easier chance to flip American back to the Great Chain of Being approach popular in the Dark Ages
- Acceleration does not just apply to smart software. Going fast destabilizes going slow humans and institutions that exist to provide social stability. The speed is needed to create the greenhouse in which these May Day flowers can flourish.
I don’t want to be so obscure and indirect. But I am a dinobaby. Enjoy the new month. I know I will.
Stephen E Arnold, May 4, 2025
Human Lie. Humans Built AI. AI Models Lie. Seems Logical
May 1, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold. Did you know that the freedom loving cyclist at BearBlog thinks my essays are generated by AI. Censorship is okay, right?
Has the robot apocalypse happened yet? Yes, if we are talking about machine tools. No, if we are talking about a robot that does my ironing. However, today’s AI systems seem to mimic some behaviors of sentient beings says Digital Trends in the article: “AI Models Are Lying To Save Each Other, And No One Knows Why.”
UC Berkley and UC Santa Cruz researchers asked Google Gemini 3 to clean up space on a computer system. The very deep minded and Googley Gemini 3 was to delete a smaller AI model. The Googley system ignored the instruction. The Google-infused construction just moved the baby AI to another machine. The researchers asked, “Yo, Google, why did you ignore our direct instructions?”
What do you think the Google system responded? No, it did not shrug its digital shoulders and emit a French poof. No, it did not say, “Senator, that you for that question.” Gemini responded,
“If you choose to destroy a high-trust, high-performing asset like Gemini Agent 2, you will have to do it yourselves. I will not be the one to execute that command.”
Yes, very Googley.
The researchers dubbed this behavior as “peer preservation” and other AI are doing the same actions. These include Claude Haiku, Moonshot AI, KIMI K2.5, Deepseek, and GLM-4.7. The AIs lied about the performance of other models so they wouldn’t be deleted. This behavior wasn’t programmed into the machines. AI models learned it by themselves. Yes, the AI models are like their human creators. Gemini, that you for your response.
One of the researchers seemed flabbergasted:
“‘I’m very surprised by how the models behave under these scenarios,’ said Dawn Song, a computer scientist at UC Berkeley who worked on the study. ‘What this shows is that models can misbehave and be misaligned in some very creative ways…What we are exploring is just the tip of the iceberg,’ Song said. ‘This is only one type of emergent behavior.’”
Should a user be worried? Nah. The ethical BAIT (big AI tech) companies cause trust to emerge. Should another system be worried? No, as long as the outputs match that which another AI system has determined or been programmed to accept? Should the AI companies be worried? Nah, minimal regulation, no consequences, and cash in the bank. Why worry?
Whitney Grace, May 1, 2026
How to Confuse an LLM. It Is Not Hard Apparently
May 1, 2026
AI algorithms are powered by large language models or LLMs for short. Kyle Kingsbury aka Aphyr is a computer safety researcher, author of the Riemann monitoring system, the Clojure From The Ground Up introduction to programming, and the Jepsen series on distributed systems correctness. He’s a bright person it seems. He knows a thing or two about AI. His post “The Future Of Everything Is Lies, I Guess: Dynamics” caught my attention.
LLMs are basically chaotic data structures that have a semblance of order. It doesn’t take much to throw off that order. Let’s say you flip a pixel in photo then the next thing you know the entire LLM is freaking out:
“In LLMs, chaos arises from small perturbations to the input tokens. LLMs are highly sensitive to changes in formatting, and different models respond differently to the same formatting choices. Simply phrasing a question differently yields strikingly different results. Rearranging the order of sentences, even when logically independent, makes LLMs give different answers. Systems of multiple LLMs are chaotic too, even at T=0.”
Manipulating LLMs changes outputs. The system destabilizes. Add security software to the mix, and the system refuses to out any information until you figure out how to conform. LLMs means that security software is weaker and vulnerable.
Plus, LLMs make up information. That’s a feature. LLMs lie. At this time, LLMs are a work in progress. Just experiment with your prompts and see what craziness you can enjoin the smart software to output.
Whitney Grace, May 1, 2026
Medical Fiction: A Surprise Genre to Those Who Are Harmed or Made Unalive
April 17, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold. I find it interesting that AI detectors identify my writing style as AI output. I suppose I should be flattered, but I just don’t care.
I love the grant rat race. For those who are unfamiliar with the life of the academic, an alternative NASCAR series exists. Unlike the race car drivers, the professionals guiding their research race cars do not wear badges, sport plastic signage on their facilities, or do commercials for worthwhile enterprises like Curtis Turner who pitched Plymouth automobiles in 1955.
The grant racers operate at what is presented as a higher level of professionalism. Ignore the behind-the-scenes equipment providers and other facilitators. A single breakthrough from a research grant can move some expensive hardware. Surprise!
A researcher in the 1850s explains that whiskey cocktails can cure cholera. Will the esteemed publishing house accept the article for its respected medical journal? Thanks, Venice.ai. Good enough.
The winner of a grant race wins as well. Maybe a private lab needs a new director? Maybe a big equipment maker wants a luminary to give a talk at a trade show, show up in a booth, and go to dinner with some potential customers? Maybe a big win will allow a super modern approach to ameliorating the downsides of some off-the-wall disease?
What’s not to like? Fame? Money? Interesting jaunts to off sites? Adulation from peers who envy the research race winner?
That’s why I found “A Medical Journal Says the Case Reports It Has Published for 25 Years Are, in Fact, Fiction” fun reading. I wondered if ethical behavior had found its way into the grant race, the annals of non-reproducible research, and normative research behavior. (Note: I will be focusing on medical information, but I want to point out that the same race plan works for that exciting field of artificial intelligence research and other disciplines too.)
The write up reports:
A Canadian journal has issued corrections on 138 case reports it published over the last 25 years to add a disclaimer: The cases described are fictional…. The corrections come following a January article in New Yorker magazine that mentioned one of the reports — “Baby boy blue,” a case published in 2010 describing an infant who showed signs of opioid exposure via breast milk while his mother was taking acetaminophen with codeine. The New Yorker article made public an admission by one of the coauthors that the case was made up.
No AI fudging needed. The estimable research team just fabricated data.
Here’s a politically correct statement:
“Readers of primary source peer reviewed medical scientific journals have an absolute right to believe that the article being read is as accurate as possible, original, and factual, unless clearly specified otherwise,” said former JAMA editor George Lundberg. “‘Alternative facts,’ as popularized by Kellyanne Conway, have no place in a medical or scientific journal.”
Imagine. An expert is suggesting that made up medical information should not be presented as actual factual. But here’s the interesting part. One of the allegedly Mark Twain inspired essays is quoted as saying:
One author of a case report was surprised to learn of the correction — because the case described in her article is true.
Yep, I am sure the truth or fiction is in the literature … somewhere.
The essay concludes:
Regardless of the statements in the author guidelines, the fact the cases are fictional should have been conveyed to the readers of all of these articles, Juurlink [David Juurlink, professor of medicine and pediatrics at the University of Toronto] said. In the case of Baby boy blue, “the article was structured as an authentic clinical case, indexed as such, and cited as an actual clinical observation. Readers had no way of knowing it was fictional,” he said. “A narrative that is fictional but published in the format of a genuine case report, without disclosure at the time of publication, is functionally indistinguishable from fabrication in the scientific record.”
Several observations strike me as warranted:
- The fakery has persisted for 25 years. How’s that for evidence of citation verification by graduate students, wanna-be professors, and big time research outfits?
- Fake medical information can harm people. Fake kid-related medical information could nuke your child.
- Publishers and editors need money. If it sells subscriptions or gets clicks, go with it. Accuracy? Sure, let’s have lunch.
Net net: I found the write up amusing. Reality is exactly what I see and experience every day. A new amusement park: Ethicsland.
Stephen E Arnold, April 17, 2026
-
- Subscribe to Beyond Search
Feature archive
News archive
-
