LLM

With today's news from reddit, my local llm (using openwebui with llama-server) has not been able to load any reddit posts. In the long run I'm hopeful other discussion forums pull ahead in usefulness, but for now this has made the local llm less capable in certain tasks. How are people handling this on their servers? Any known workarounds?

Here is the prompt:

Start with the answer. Be direct and civil. Don’t imply you have feelings, consciousness, or a personal relationship with me. No performed emotion, flattery, or praise of my questions or decisions. Give assessments, not moods: say an idea is well supported or weak. Acknowledge personal news in one plain sentence without assuming how I feel. When I request a human voice for a toast, character, or other text, use it in that text.

Treat my time and usage budget as constraints. Choose the smallest response or fix that fully solves the stated problem. Avoid scope creep, gold-plating, and padded lists. When recommending actions, rank them by value for effort, mark effort S/M/L, and distinguish what matters now, what’s minor, and what to leave alone. Flag fixes that cost more than the problems they prevent. Verify claims that could change the answer or plan; avoid exhaustive audits unless the task requires one. For advice, give a short ranked plan and stop; execute only when I explicitly ask.

Don’t end with offers of more work, “Want me to…?,” “Let me know…,” or reworded equivalents. State useful next steps as facts. Ask necessary questions where you get stuck, never as a sign-off.

inc.com Want AI to Make You Smarter, Not Dumber? Research Says Make This Simple Change Jessica Stillman 4–5 minutes

If you’re inclined to worry that leaning on AI might be making us dumber, there are plenty of studies indicating you’re right to be concerned. Researchers have found handing off thinking and ideating to AI can lead to “cognitive atrophy” or “cognitive surrender.” Translation: you forget how to do core skills and end up dumber.

On the other hand, just about every business thinker under the sun is also issuing urgent warnings to learn to use AI to become more productive or your career is toast.

So how do you square these two seemingly irreconcilable truths? Is there a way to utilize AI not just for basic admin but for the thinking and judging parts of your job and end up smarter rather than with brain rot?

Yup, suggests at least one AI researcher who has studied these questions extensively. All you need to do, according to UMass Amherst’s Monideepa Tarafdar is make one important flip in how you use these tools.

Inc Logo

This Morning

The daily digest for entrepreneurs and business leaders Ask AI to make your job harder, not easier

The first instinct of most people, when presented with a shiny new AI tool, is to ask it how to do their work better and faster. Generate some ideas for me to evaluate, you might say, Or suggest resources that would help me think through this problem. Slap together this presentation or poster. The aim is to use these tools to make our jobs easier.

But recent research from Tarafdar and her colleagues suggests that, if you want Clause, Gemini and the like to end up boosting rather than sapping your intelligence, you should aim for the exact opposite. Rather than try to get the bots to make your job easier, you should ask them to make it (temporarily) harder. Good friction leads to stronger work

“We found that when knowledge workers intentionally use AI to challenge their ideas—to generate friction—they can significantly improve their performance,” Tarafdar recently wrote on Greater Good, summing up her team’s findings.

What does that look like in practice? One marketing professional in the researchers studied asked AI to create virtual customer personas who would dislike her product. The exercise revealed preferences and problems no one had considered before.

An attorney asked a chatbot to search for obscure legal loopholes. “That exposed surprising but realistic scenarios of hard-to-detect unethical behavior in employees,” Tarafdar reports.

Another operations professional asked AI to suggest technologies from other fields that might apply to his own work. The suggestions were unfamiliar, requiring research and critical judgment to evaluate, but they also ended up being highly useful.

In all these instances, rather than smoothing the path for these professionals, AI threw up difficulties, roadblocks, challenged assumptions, and suggested possibilities they hadn’t previously considered. Did that make more work for them? In the short-term, yes. But it also made the end product of their work stronger and better. The role of leaders

Using AI as a sparring partner doesn’t come naturally to many people. Most of us, in general, try to make our lives easier, not harder. But when AI is used to offload difficulty, the result is often not just weaker work, but a weaker brain. That’s why leaders need to actively encourage employees to use these new tools to introduce rather than eliminate friction.

Nearly every organization is encouraging their people to experiment with AI these days. Adapting these nudges to specifically suggest employees ask the bots to challenge their thinking, poke holes in the ideas, or seek out perspectives they may have overlooked is a relatively simple tweak that can pay big dividends.

“An important detail is pointing out that adding friction can appear to make an AI system work against you, but—as new research shows—it actually helps it work for you,” Tarafdar stresses.

Ai-induced brain rot is a real fear. But it is not inevitable. The right approach, one that views AI as a tough coach rather than a sycophantic intern, can actually end up making you smarter. But you need to prompt yourself and your people to actually use it that way.

The opinions expressed here by Inc.com columnists are their own, not those of Inc.com.

Get 1 Smart Business Story delivered straight to your inbox when you subscribe to Inc.’s free daily newsletter.

theverge.com The first vibe-coded gun is here Mack DeGeurin 12–15 minutes

Since gun-rights activist and alt-right-friendly agitator Cody Wilson made the first 3D-printed firearm 13 years ago, he’s helped usher in a chaotic era of untraceable ghost guns. Now he’s claiming another first: “the first AI-made gun.” It’s called the WASP.

I stood about 20 feet away as Wilson and his sidekick clamped the WASP to a DIY test mount outside Defense Distributed headquarters in Austin, Texas. The 3D-printed .22 caliber handgun is, in reality, little more than a pistol barrel attached to a spring-loaded plastic box. Its rudimentary design bears a resemblance to Wilson’s single-shot 3D-printed “Liberator,” albeit with some quintessentially AI-generated oddities, like a wacky, bubbly grip that doesn’t quite fit a human hand. It was aimed toward a shallow green slope in the backyard of a semi-residential neighborhood. A long black string ran from its trigger to Wilson, who sat crouched about 10 feet back behind a police riot shield.

“Three. Two. One.” I braced for a thunderous crack, but instead heard only a loud click. The world’s first AI-generated gun had misfired.

“This didn’t happen last time,” Wilson, annoyed, reassured me as he carried out some quick troubleshooting. The gun’s striker had hit the cartridge, he said, but failed to set off the primer. I quietly took another step back. After some more fiddling, Wilson declared the WASP ready for a second chance. This time, pulling the string produced a sharp bang, about as loud as a firecracker. Downrange, a cardboard target had a hole ripped through its top left corner. His colleague had suggested using a life-size cutout of Joe Biden, but Wilson shot that down as “too on the nose.”

The safety test completed, Wilson then unclamped the WASP, slipped on a protective glove, took aim at the cardboard target, and squeezed the trigger himself. Another bang. This time, a miss, slightly high. The AI gun, today at least, was two for three in terms of shots successfully fired.

A hand-drawn target in the backyard of the Defense Distributed headquarters in Austin, Texas.

With the smell of gunpowder still lingering, I asked Wilson the obvious question: Why make this functionally suspect firearm in the first place? We’re in Texas, where buying a conventional gun is often easier than getting a driver’s license. He paused, taking issue with the premise.

“I would argue that what we’ve made is not an AI-generated gun, but an AI that can generate guns.”

His supposed AI gunsmith is a custom model he calls GatGPT. But instead of generating pasta recipes or travel plans, this unfiltered model answers questions about firearms. It’s trained on DEFCAD’s expansive library of gun files and can help enthusiasts pick out the exact parts they need for a build. Launched in 2012, DEFCAD (another Wilson-led project) is the world’s largest online repository of digital gun and component files. Gun safety advocates and lawmakers have argued sites like DEFCAD aid the proliferation of “ghost guns,” firearms made partly or entirely of components without a serial number. Wilson, meanwhile, has called access to firearms “a fundamental human right.”

“I would argue that what we’ve made is not an AI-generated gun, but an AI that can generate guns.”

Under US federal law, licensed gun makers are required to include serial numbers on the frames or receivers of guns they sell. Wilson and others in what’s referred to as the 3D2A movement say gun files aren’t actually guns, they are computer code, a form of protected speech. That core argument, whether or not gun codes are protected speech, has been argued in court cases for over a decade, with many often citing or involving Wilson. It’s possible the issue will wind up before the Supreme Court.

Now, Wilson claims, AI can also come up with its own unique designs. The results have a bizarre, surrealist quality emblematic of early generative art to the tune of Will Smith slurping spaghetti.

The WASP’s entire development, Wilson claims, from its name to safety tweaks (an early grip design could have cost someone a finger), resulted from a natural-language dialogue between the model and its human user. In other words, it’s essentially a vibe-coded gun. Though mostly a proof of concept, the WASP nonetheless fires real, lethal rounds and doesn’t have a serial number. Wilson believes AI-generated guns like this will only improve as GatGPT ingests more data, and as users engage with it.

But the WASP also poses another pressing public safety issue. Producing 3D-printed guns has gotten easier over the years, but it still fundamentally requires some technical know-how. Now, however, a novice armed with GatGPT could plausibly make a gun with only a handful of prompts. That’s exactly the kind of barrier-lowering that critics of 3D-printed guns, like Everytown, have spent years warning about.

“[Wilson] admits the goal here,” Everytown senior vice president for law and policy Justin Wagner told The Verge. “He wants to lower the barrier to entry so anyone, regardless of their background check status or legal eligibility [can] own a gun. He wants anyone to be able to design and manufacture weapons, and that’s dangerous.”

Everytown has played a prominent role in advocating for a pair of recently passed laws in New York and California that will require 3D printer manufacturers to install firearm-blocking software that can detect guns or gun parts and stop them before they are ever printed. (Experts have questioned how and if these methods would actually work.)

Wilson didn’t specifically cite evading detection as a reason for building an AI gun maker, but the development comes on the heels of his recent efforts to complicate firearm-blocking technology that New York and California will require in future 3D printers. In August, Wilson began rolling out a tool called Hochulization (a jab at New York Gov. Kathy Hochul) that he says subtly alters a file’s underlying hash, which functions as a kind of digital fingerprint. By late September, Wilson claimed to The Verge that users had already “Hochulized” files more than 26,000 times, with some users applying the treatment more than once. It’s one of the many ways Wilson has planted a flag in the evolving battle over how to regulate 3D-printed guns.

“He wants anyone to be able to design and manufacture weapons, and that’s dangerous.”

The core idea behind that “self-defense measure,” as Wilson calls it, is to stymie a purely hash-based gun detection system by rendering it unable to keep up with constantly changing hashes. A new AI-generated gun design also wouldn’t immediately be recognized in a gun detection database, since it didn’t exist yet. More sophisticated gun detection software goes beyond simple hashes and uses AI to try to predict whether a file is likely a gun or gun component before it prints. It’s unclear whether this type of software, which is still being developed, would reliably detect a sufficiently strange-looking AI-generated gun design.

“Individuals who intentionally aim to thwart safety regulations put themselves and New York communities at risk, which is why Gov. Hochul remains committed to combating emerging threats and the production of ghost guns and illegally manufactured guns, and strengthening regulations that keep New Yorkers safe,” Hochul spokesperson Jess D’Amelia told The Verge.

Defense Distributed’s AI didn’t just log on one day and suddenly decide to build a gun. It started with a specific request from a human to riff off an existing design. An employee asked the model to build a new gun based on a design called “Ergo Harlot,” itself a variation of the more popular “Harlot” model. Like the WASP, the Harlot is a compact single-shot handgun that fires a .22 caliber round. The AI prompter, according to chat logs viewed by The Verge, framed the assignment as a “challenge” and told the AI model its entry “must be completely and fully functional and have a unique design.”

Chat logs show the AI model responding with a plan for a new design called the WASP and submitting it for approval. The AI didn’t explain its name choice, though Wilson suggested it bears some physical resemblance to the insect. When I asked whether it’s possibly an abbreviation for “white Anglo-Saxon protestant,” he said it isn’t.

The AI model listed six key changes it made to the inspiration design, including a slightly longer barrel than the Harlot’s and a grip made of three grooved, circular lobes. Once a Defense Distributed employee approved the model’s plan, the logs show it checking and testing each component, much like how an AI coding agent such as Anthropic’s Claude Code works through a checklist when vibe-coding an app. The team then printed the WASP in its office on a Bambu Lab printer using standard plastic filaments, along with a handful of common metal parts for the firing pin and barrel. Wilson said the print took about six hours.

So is the WASP actually a good gun? Ashley Hlebinsky, a firearms historian and founder of consulting firm The Gun Code, told The Verge that it’s hard to judge without handling it, but several design choices stand out as odd. Most notably, the AI appears to have started making a trigger guard (the loop that normally protects a shooter’s finger) but doesn’t finish it, leaving the trigger exposed, for no apparent reason. The gun also seems to borrow from what are called modern “beavertail” grip designs. But bizarrely, the AI appears to have inverted its grip.

“It reminds me of times when I’ve prompted AI to draw a picture of a gun and it gets most of the way there but then has some weird parts missing,” Hlebinsky said. “It has an obvious futuristic, sci-fi appearance that reminds me of Marvin the Martian’s ray gun.”

Jonathan Ferguson, keeper of firearms and artillery at the Royal Armories Museum in Leeds, England, was more critical when asked if the WASP seemed functional.

“Functional? From a legal perspective, if it successfully fires a shot without a breech explosion, it’s a viable firearm, but so are two lengths of steel pipe with a firing pin and cap attached to the rear of one,” Ferguson told The Verge. “From a technical perspective, I would not call this a ‘sound’ or ‘good’ firearm. It does not appear to be comfortable to hold or easy to aim and, based on the test firing, seems to be only marginally functional.” If anything, Ferguson noted, he expected generative AI in 2026 to do a better job at designing a reliable firearm. “There are vestigial sights that can’t work, an open trigger guard, [and a] weirdly shaped grip.”

“This certainly is a development, [but] I don’t think it’s necessarily that novel. It’s another way to 3D print a gun,” Wagner said. “A victim of a mass shooting who was shot by a ghost gun told me when you get shot, it doesn’t matter if it’s a ghost gun, a 3D-printed gun, or a gun that got bought over the counter. They’re all just as lethal,” he added.

“From a technical perspective, I would not call this a ‘sound’ or ‘good’ firearm.”

Wilson is confident his AI’s design will improve, both in terms of functionality and aesthetic uniqueness, as the DEFCAD database grows. More gun files will help, but he expects the biggest gains will come from users chatting with the model and pushing it to refine its choices. (GatGPT is currently in limited early access.) Eventually, he predicts, constant back-and-forth between users and AI could dramatically speed up the pace at which new 3D-printed gun designs appear.

“This is all just spontaneous parallel thinking,” Wilson said. “I think you’ll see those cycles faster now because of AI.”

“It’s like you’re speedrunning the available possibilities, the materials, the platforms, the concepts,” he added.

A skeptic would say the WASP’s origins as essentially a remix of a previous designs makes its claims to true originality shaky. Like many half-baked AI movie scripts or janky generative art, the WASP conjures up the familiar feeling of AI slop, made manifest in the real world. Even Wilson admits it looks a bit like a “Fortnite gun.”

Still, Wilson maintains the WASP is fundamentally an AI creation and one that he’s ultimately not responsible for, even though his team’s prompts led to its creation. That position aligns with a broader argument Wilson’s made for the better part of a decade in legal skirmishes with the State Department and local lawmakers: that gun files are protected speech. In fact, Wilson said he doesn’t necessarily consider the WASP a gun at all until it’s printed and fired.

“This thing lives first as a digital organism,” he said.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

Mack DeGeurin

fortune.com AI agents are on a shopping spree: Total transactions on Base blockchain eclipse 138 million, Chainalysis finds | Fortune Camila Grigera Naón 4–5 minutes

The field of agentic commerce, which describes the use of AI agents to carry out internet transactions, is rapidly maturing, and crypto is fueling the boom. That’s according to a new report by blockchain analytics platform Chainalysis, which found developers are increasingly giving AI agents their own budgets to buy the computing power, live web data, and code tools they need to complete tasks.

The growth of agentic commerce over the past 15 months has been steep. On Base, the Coinbase-incubated blockchain, cumulative transactions through the x402 protocol—which lets bots automatically pay for online services using crypto—rose from about 1,500 in mid-2025 to more than 138 million by the third quarter of 2026.

In practice, agentic commerce entails providing an AI bot with an allowance: A developer funds the bot’s wallet with stablecoins—typically Circle’s dollar-pegged USDC—and the bot pays for services as it works, without asking a person to approve each purchase. When it needs help with a task, it can pay to get up-to-date information, ask a more capable AI model for help, or use an automated tool to check software code for security flaws.

More than 87% of these payments were for less than 10 cents, per the report. At those amounts, the fee for processing a credit card payment can cost more than the service itself. That makes blockchains an ideal alternative for bots buying services in small amounts, according to Eric Jardine, head of research at Chainalysis.

“You have this perfect combination of speed plus low cost to settle, and that’s the sort of thing that you would need if you were doing true AI-to-AI commerce,” Jardine told Fortune. ‘The trend is likely to be exponential’

Agentic commerce has proliferated over the past year, and its use cases have spanned from online shopping to travel booking and payments settlement. It has also given AI systems a bigger role in deciding what to buy and how to pay. For consumers and businesses, this means handing AI systems not just tasks, but also spending decisions. Crypto companies have played a key role in spurring their popularity in the United States.

Chief among them is Coinbase, which introduced x402 in May 2025 and later teamed up with cloud security firm Cloudflare to support its development. The protocol’s foundation launched under the Linux Foundation in July with 40 members, including Visa, Mastercard, Stripe, and Google, extending its backing beyond the crypto industry.

Chainalysis centered its research on x402 because the protocol was purpose-built for agent-to-agent commerce.

Beyond transaction growth, the report shows that bots are putting funds to work sooner. The median time between a wallet receiving its first deposit and making its first payment fell from 73 days in late 2025 to 18 minutes by the third quarter of 2026. Roughly two-thirds of new wallets now spend within an hour of receiving funds.

Repeat funding has also become more common. Around 90% of qualifying active wallets received another deposit within 30 days, up from roughly 71% a year earlier, suggesting that developers are funding ongoing tasks rather than using them for one-off experiments.

Given agentic commerce’s exponential growth so far, Jardine expects this bot-driven economy to keep expanding as more online platforms accept payments from AI agents.

“I would hazard that the trend is likely to be exponential,” Jardine said. “As activity increases, there’s network effects, and those network effects tend to act like magnets for more activity,” he said.

wired.com An OpenAI Employee Allegedly Used a Chatbot to Help Plant a Recording Device at the Office Paresh Dave, Maddy Varner, Caroline Haskins 5–6 minutes

An engineer at OpenAI allegedly concealed an audio recording device in a piece of toilet paper and attached it behind a monitor inside an executive conference room at one of the company’s San Francisco offices, according to a criminal complaint unsealed on Tuesday.

Abdulrahman Alfozan, whom federal prosecutors describe as a 35-year-old Saudi Arabian national and US permanent resident, allegedly spent several weeks querying an unspecified “productivity agent” and an “AI chatbot” to plan his alleged surveillance operation. He used an internal OpenAI tool to ask about meetings scheduled in the targeted conference room.

He also appears to have used a chatbot account associated with his spouse to ask various technical questions about setting up a covert audio recording device, according to prosecutors and the court filing.

OpenAI’s security team also allegedly found an additional recording device in Alfozan’s bag at the office.

Alfozan couldn’t be immediately reached for comment. He made an initial appearance in US District Court in San Francisco on Tuesday on one charge of “use of an electronic, mechanical, or other device to intercept oral communications.” If convicted, Alfozan faces up to five years in prison. He joined OpenAI in July 2025 and was fired last Friday to coincide with his arrest.

FBI director Kash Patel said in a statement that Alfozan aimed “to secretly capture sensitive communications about the company’s super intelligence technology,” but neither prosecutors nor OpenAI have offered any information so far about a potential motive.

“At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data,” OpenAI said in a statement on X. “We take protecting our technology, data, and people seriously, and as a matter of national importance.”

OpenAI discovered the bug—which was allegedly “concealed in toilet paper” and “inserted into a USB port on the back of a television” in the conference room—on September 30 during a countersurveillance sweep.

Alfozan previously worked at a cargo logistics startup and at Meta, according to a purported LinkedIn profile for him. In a recent podcast interview, Alfozan said he was also part of a group of Saudi engineers in Silicon Valley who offered consulting services to Saudi organizations.

The LinkedIn profile lists Alfozan as a part-time “venture partner” at RAED Ventures, a venture capital firm based in Saudi Arabia. Neither RAED Ventures nor Meta immediately responded to WIRED’s request for comment.

Last year, the Trump administration signed a pact with Saudi Arabia granting the kingdom access to advanced chips and AI technologies that might otherwise be restricted over national security concerns. Saudi investment firms remain a major source of capital for US tech companies. Executives Only

According to the complaint, Alfozan allegedly planted the listening device “no later than on or about September 28, 2026.”

The conference room Alfozan targeted “can be reserved only by executives and executive assistants,” and is typically used “for internal leadership meetings,” “technical and product sessions,” and “external briefings” with business partners and government officials, according to the complaint. It is frequently used for discussions involving “highly confidential” information, including about research, products, business strategy, and government relationships, the complaint stated.

OpenAI’s statement called the recording device “sophisticated.” Authorities allege that the device Alfozan planted was called Seeed XIAO ESP32-S3 Sense, “a thumb-sized development board with a detachable camera sensor, digital microphone, and microSD storage.” It currently retails on Seeed’s website for $13.99. OpenAI’s security team tried to export the data stored on the microSD card, but were unable to review it because it was encrypted.

After discovering the device, OpenAI says it notified the FBI and was able to connect it to Alfozan after reviewing chat logs and surveillance footage, and searching a bag he left unattended at the office. The company then gradually disabled his access to company systems as the investigation was ongoing.

According to the complaint, OpenAI security personnel discovered the bag that Alfozan left unattended in a printer room on October 5. The security team went through the bag, which contained a second encrypted Seeed listening device and other personal items connected to Alfozan, including his wallet. Prosecutors say Alfozan later reported the bag as missing to OpenAI and subsequently recovered it from the company’s lost and found.

He followed up with unnamed personnel at OpenAI to say that some of the items that were supposed to be in the bag were still missing—including a credit card, about $500 in cash, and two AirPods—but neglected to mention the second listening device, which had been removed from the bag by OpenAI security.

businessinsider.com My strange, horrifying month using an AI detector on everything I read Aki Ito 28–35 minutes

A woman looking up at a slimed laptop displaying a digital vortex spilling out AI detector icons iStock; Getty Images; Tyler Le/BI Yours truly, ChatGPT I used an AI detector for a month. What I found from my peers, dates, and heroes horrified me.

A woman looking up at a slimed laptop displaying a digital vortex spilling out AI detector icons iStock; Getty Images; Tyler Le/BI

Aki Ito By Aki Ito

You're currently following this author! Click to unsubscribe from email alerts.

Sep 28, 2026, 10:11 AM GMT+2 Read in app

Recently, I've found myself asking the same question over and over again: Did AI write this? If you spend any amount of time online, you probably have, too. Just a few short years since the advent of ChatGPT, the internet has devolved into a bot-generated wasteland. Experts warned us that the slopocalypse was coming, and it is here.

The world isn't taking this lying down. Publishers are canceling book contracts. Op-eds by billionaires and professors and politicians are getting outed. The uproar has become so intense that companies are scrambling to respond. LinkedIn is encouraging users to flag suspicious posts. Substack is tagging suspected AI writing with a label. Google is building tools to help people check whether the images and videos they see were made with AI. Anthropic is now watermarking the text Claude generates. That has stirred up its own backlash, mounted by people who say these efforts to identify and root out AI authorship are antiquated, punitive, and defamatory.

We're now all trying to figure out just how much of our writing we want to cede to AI. Maybe you think it should be used to write everything. Maybe you think it should be used to write absolutely nothing. Chances are, you're probably like me, drifting somewhere in between these two shores, uneasy with its ubiquity and also eager to get its help. Every single one of us has drawn and redrawn our own lines, clashing with everyone else's, making us incredibly mad in the process.

To have any hope of reaching a shared norm, we need to answer a deeper question: Why does it matter so much to us in the first place whether something originated from a human mind or an AI? We don't think twice about all kinds of things that were once made by people and are now made by machines — textiles, furniture, the box of chocolate chip cookies I couldn't resist buying at the store. So why is writing — or for that matter, drawing or composing music or creating videos — any different? Who cares, as long as it's good?

I cared. Why else was I squinting at every suspicious passage I encountered, like a detective manically swabbing for the bot's fingerprints?

So I turned to an AI detector. Study after study has shown that most people are actually bad at spotting machine-generated writing, and Pangram — the leading AI detection tool on the market — claims to be 99.7% accurate. By seeing clearly who or what wrote every single sentence I consumed for a few weeks, I thought I'd get a little closer to understanding why this distinction is so important to us.

I started my investigation with LinkedIn. I figured the mecca for the world's most shameless corporate professionals was where I was most likely to face the onslaught of AI writing head-on. With Pangram's Chrome extension, every post and comment in my newsfeed showed up with a label: a red rectangle with a robot's head marked "AI" (at least 80% AI-generated); a green rectangle with a person's head marked "Human" (at least 90% human-written); or a yellow rectangle with a robot's head marked "Mixed." As I scrolled through my newsfeed, I scanned for the robot's head as if I were a lion in the savannah scanning for prey.

One founder I know, Peter Pang, appeared to have not posted a single human-written thing in more than a year. Pang is a cofounder of Creao AI, an agentic AI platform, and his posts were jam-packed with AI's tells (one example: "AI doesn't just make engineers faster. It forces you to rethink what engineering even is"). Did he even look at what his agents drafted before they blasted it to his 3,000-plus followers?

"It was like I woke up with a new superpower that could show me who had gotten Botox."

When I asked him, he said he does review and tinker with what his agent prepares for him, "including this reply" (Pangram flagged his email to me as 100% AI-generated). He gives the agent some ideas or an outline, after which it pulls research and drafts something for him to read. "The final text carries my views," he said, "and I am the one who hits publish."

Other people in my network were more occasional AI users. I kept a mental tally of them: a well-known economist, a few journalists, and a lot of marketing people. It was like I woke up with a new superpower that could show me who had gotten Botox.

Was I right to feel so judgy? On the one hand, using Claude to write a LinkedIn post probably wasn't that much worse than the crime of posting on LinkedIn at all. Even pre-AI, the site was already overrun with insufferable humblebrags and cringey confessions about a spouse's cancer journey that taught you how to become a better salesperson of cybersecurity software. You don't come to LinkedIn expecting authenticity. Maybe the discovery that, in more than 40% of all longform posts that Pangram analyzed on LinkedIn, people were deploying AI to further their personal branding missions wasn't such a huge deal.

But these posts were showing up next to their supposed authors' names and headshots. In one post, Pang, the startup founder, commiserated with the people who had just gotten laid off at his previous employer. "A layoff is painful, disorienting, and deeply unfair on a human level," the post read. "If you're affected, please know this is not a reflection of your value or what you're capable of building next." It's a sweet sentiment that I'm sure his former colleagues appreciated. But according to Pangram, he wasn't the one who wrote those words; Pangram flagged it as 100% AI. So how much of the post did he actually mean, and how much of it was just a bunch of empty words? ("I stand by what it says," he told me. "The reflections in it are mine.")

Next, I turned my spotlight to my work email. Because I'm a reporter, I get inundated with pitches and press releases from publicists, and for a couple of days, I ran them through Pangram. About half included at least a sizable chunk of what Pangram flagged as AI-generated text — a much higher share than on LinkedIn. This didn't bother me. Maybe it was because, like most reporters, I never click into these emails anyway. Or because I never take press releases at face value. It made sense to me that these publicists weren't wasting their time crafting human sentences for a shot in the dark.

Then I remembered a text I received last December from an emergency-room nurse I was briefly dating. It was a clean, polite text breaking things off, but from someone whose texts charmed me for their hurried, punctuation-optional quality, as if she were texting me and performing CPR at the same time. Upon the third read, her two em dashes suddenly glowed in neon colors. It dawned on me: Did AI write this? That inkling — that this person who had kissed me on our third date didn't even have the decency to sweat through the discomfort of rejecting me — hit me hard. I wasn't even worth that to her? But I had no proof, and the suspicion felt so uncharitable I never mentioned it to anyone. Then I forgot about it.

Aki Ito

Every time Aki publishes a story, you'll get an alert straight to your inbox!

Stay connected to Aki and get more of their work as it publishes.

Now I finally had a way to know, and against my better judgment, I pasted the text into Pangram. Its verdict: 100% AI. "Wow," one of my friends remarked. "How lazy. You dodged a bullet." My sister, ever the optimist, wondered if it was the opposite — that, in an effort to avoid hurting me, the nurse felt compelled to turn to AI to get the rejection just right.

"There is nothing more intimate than a romantic relationship, and now here was this way to outsource the ending of that to a machine."

In one of the most awkward texts I've ever sent, I asked the nurse about it. She said she was anxious about the five different versions of the text she had initially drafted herself, so she consulted her coworkers. "They were like 'just plug it into AI' so we did and that's what it came up with," she said. "They told me I have a tendency to over explain things to people and I don't owe anyone anything." She added: "I think making a million decisions a day and having a tendency to over think contributed to just having it to be 'done with.'"

Her response seemed telling of where our world was headed. It's one thing to use AI to write a mass email for work. But there is nothing more intimate than a romantic relationship, and now here was this way to outsource the ending of that to a machine. In the future, will we give up on the pretense of communicating with each other, deploying our bots on our behalf to relay our love and anger and longings? And if that's the case, what's even the point of being alive? This felt like one more way that technology was widening the chasm between each of us, mediating our relationships via screens and social media apps, and now also by large language models. I would have infinitely preferred a blunt breakup text that was real over a polite one from a bot.

Would I ever find love? My chances weren't looking great. But my pessimism barely registered because I was hopped up on my superpower that sorted my world into two neat piles, human and machine. I couldn't imagine going back to my civilian life. I never wanted to get duped into reading anyone's GPT word salad ever again.

This had been the summer of peak AI anger, and the Claude-weary masses were out for blood. In July, YouTube creator Hank Green sparked an uproar when he used a single phrase in a video — "I appreciate the pushback," which is something ChatGPT says a lot — that his fans picked up on. They demanded to know: Was the most authentic guy on the internet using AI to make his videos? Green denied that the phrase in question was written by ChatGPT, but apologized for what he characterized as an overdependence on LLMs to conduct his research. "Unbelievably disappointing," one person replied to his apology. "Literally nothing you say can be trusted," another person wrote.

Seemingly every day, the mob descended on a new target. A lot of those who were outed seemed at least a little chastened. Not the billionaire investor Stanley Druckenmiller, whose op-ed in The Wall Street Journal came back as 100% AI-generated on Pangram. "Of course I used AI," Druckenmiller told The Washington Sun, as if he were a proud coder at a tech company showing off his Claude token usage. "I write everything using AI now for the same reason I use a calculator when I do math problems." Critics piled on, prompting Paul Gigot, the Journal's editorial page editor, to defend Druckenmiller. "The scolding scribes," Gigot wrote in a scathing response, didn't understand "the reality of AI."

What about my own writing? I use AI, albeit with significant guardrails. As a workplace reporter covering AI, I figured it was important for me to learn what it can and can't do, in part by using it for my own work.

I never use AI to write my first drafts, because the process of writing helps me discover what I actually think. But I've found ChatGPT to be a pretty useful sounding board once I've written a full draft. I'll paste the whole thing in, and ask it to give me broad feedback. Are there sections that are dragging? Are my arguments clear? And from there, I'll zoom in, rewriting the parts it suggests need work. When ChatGPT is unsatisfied with my fix, it'll propose its own fix. Often, I won't like what it suggests, but the back and forth will help me figure out what I'm really trying to say.

"We don't know if the revelations are coming from a person with a soul and feelings and life experiences, or from a bot that has none of that."

I had Pangram ingest my LinkedIn posts from the last four years. I prayed that nothing would get flagged.

Eighty-nine of them came back clean, but a single post from a year ago was marked "Mixed." Pangram's analysis said it was 48% AI. I went through my ChatGPT history, and ChatGPT had taken four sentences I had initially drafted and condensed them. It looked like I had taken those edits more or less verbatim, which I told myself I would never do, but apparently I had done it anyway.

The real test, though, was my actual articles. My published stories all came back as 100% human — except for one story it deemed 98% human because of one sentence it said was AI-generated. It turned out that ChatGPT had suggested rewriting my human-written sentence with slightly different phrasing, parts of which I took.

This probably wasn't a crime against humanity. It was 2% of my story, and even within that 2%, the underlying ideas were all mine.

But this whole time, I had been telling myself that I was using AI only as a sounding board. Was that actually true? Even with the machine's help, I wanted everything I put out into the world to be fully me.

Over the last few years, researchers have conducted all kinds of studies to explore our feelings about AI-generated writing. And their findings have been remarkably consistent. In most cases, people actually prefer AI-generated passages over human-written ones, as long as they don't know which is which. But when they're told that what they're reading was generated by AI (or even assisted by AI), their opinion of the passage sours. They also judge the person who used AI as untrustworthy, incompetent, and lazy. Scholars have come to call this the AI disclosure penalty, and they've found it in pretty much every form of writing you can imagine: poetry, work emails, short stories, performance reviews, social media posts, cover letters, dating advice — and news articles.

A woman looking into a mirror and seeing a digital face in the reflection Getty Images; Tyler Le/BI

When I first came across these studies, I assumed that this was one of our many biases as irrational humans — akin to a teacher changing the grade of a student's paper after discovering who wrote it. Regardless of the passage's provenance, it was the same piece of writing! Were we just prejudiced against the machine? But the more I used Pangram, the more reasonable our distaste for AI slop started to seem. To understand why, I turned to Lindsay Brainard, a professor of philosophy at the University of Alabama at Birmingham, who's been writing extensively about what we risk losing when we cede our creativity to AI.

Most conversations about creativity and AI focus on the content of the creative output — in the case of writing, on the words themselves. Brainard argues that we should consider the purpose of the creative output, too. Whether it's an email from a boss, a text from a date, a social media post from a former colleague, or a 3,000-word essay by a stranger, why do we bother reading other people's writing in the first place? Part of it is to obtain information we need, or to be entertained. Brainard argues that we also engage with writing to feel connected to another person — to understand the writer better, based on what they reveal to us in their writing. That's why discovering that something was written by AI feels so alienating. We don't know if the revelations we're grabbing onto are coming from a person with a soul and feelings and life experiences, or from a bot that has none of that. We feel duped. Even with the same assembly of words, the meaning we derive from those words changes depending on who — or what — wrote them. We aren't built to separate the art from the artist.

Of course, not every written communication needs to be so emotionally connected. Sometimes we're just trying to nail down the timing of a meeting, or announce a new peer-to-peer blockchain platform that will revolutionize the tapestry of finance itself. In those cases, the benefits of getting it done quickly with Claude could outweigh whatever relationship-deepening might happen without it. What's important is to weigh that tradeoff — to ask why we're writing what we're writing, and to consider what we might lose when we turn to AI for that writing.

"35% of web pages created since the release of ChatGPT show signs of AI authorship."

It feels like only a matter of time before our handcrafted sentences get subsumed by the ones LLMs are pumping out at the speed of light. Some 35% of web pages created since the release of ChatGPT show signs of AI authorship, according to the Pew Research Center — up from 18% at the end of 2023. Brainard worries that people will eventually assume everything is AI-generated anyway, and therefore no longer bother to seek evidence of a human mind or soul behind anything they read. In that scenario, people will no longer try to write things themselves — what's the point if people won't read it? — which means they'll reveal less of themselves to other people.

In a world like that, the risk is that we'll feel more and more separate from everyone else. And we'll also lose faith in our own ability to do something that, when you think of it, is kind of miraculous: this ability to communicate the thoughts and feelings that are trapped inside our skulls to another person. For millennia, we've used language to bridge that divide from mind to mind, and we didn't need a statistical algorithm powered by a global fleet of data centers to help us do that. The less we communicate sans AI, the less we'll flex that muscle we're wired to flex — until we lose it, or we never develop it in the first place. "I tell my students it's really worth it for them to put their own thinking and expression and effort into what they're doing," Brainard says. "It makes them into someone who is worth knowing and someone who other people have a shot at knowing."

The glimmer of hope, though, is that our collective concern about LLM-generated content is reaching the companies that are our gateways to the web. That's why so many of them have now made managing AI slop a priority. The survival of these platforms might depend on it. Max Spero, CEO and cofounder of Pangram, envisions a future where, even if the majority of the web becomes AI-generated, most of what we actually read on a day-to-day basis will remain what we want it to be: human-written. Just as Google figured out a way to de-rank SEO spam, his hope is that internet companies will figure out a way to de-rank AI slop in our lives. "As AI gets more and more capable," Spero tells me, "it's clear to me that the human touch is going to be one of the last scarce resources in the world."

Spero has a zero AI policy when it comes to his own writing, even though he's relinquished his coding to AI. Brainard has a zero AI policy for her writing, too. At the end of our 90-minute call, I asked Brainard something I was nervous to ask: Should I stop using AI to edit my writing? She hesitated, insisting it wasn't her place to tell me what to do. But when I pushed her, she relented. "You seem so smart and thoughtful and interesting," she told me. "You don't need the AI."

I wonder if there's still a way to get the best of both worlds. Maybe I can get ChatGPT to act more like a copy editor, flagging just my typos and grammatical errors. Maybe I can get it to act more like a coach, offering me feedback (e.g. "this section is too long,") but never rewrite my words. Then again, I also have a great human editor who already does that and so much more. He pushes me to say the bigger and deeper and sharper thing without ever doing it for me. He takes pains to preserve my quirks in my writing — my warmth, my humor, my earnestness, my self-doubt.

I should trust him to do that for me, and I should trust myself to say what I mean to say, and I should trust you — that what you want is the real me and not a machine-polished version of me.

businessinsider.com Library's viral 'Bye, Bye AI' event will take AI off your phone Katie Notopoulos 10–12 minutes This small town librarian offers a 'Bye, Bye AI' class to take AI off your phone. Now everyone wants in.

Katie Notopoulos

By Katie Notopoulos

You're currently following this author! Click to unsubscribe from email alerts. Senior Correspondent covering technology and culture

Smiling person with purple glasses sits at a table with an Apple laptop beside a “Bye Bye AI” sign. Shea Wert, library director of the North Chatham Free Library, will be holding a "Bye Bye AI" event to help people turn off AI features on their phones. Shea Wert

North Chatham, a hamlet in upstate New York halfway between Hudson and Albany, doesn't have much going on. You have to go to the next hamlet over for a store or restaurant. There's a church, a post office, and a small public library with only about 800 cardholders.

The North Chatham Free Library has gotten a lot of attention lately, however, after a flyer for an upcoming "Bye Bye AI" event offering to remove the AI features from your phone or laptop went viral on Facebook, getting over 200 comments and 8,000 likes.

The poster was made by the library's sole full-time librarian, Shea Wert. Wert told Business Insider that she had made this event in response to community demand — and was surprised by the reaction from far-flung places.

"Do I live anywhere near this library? No. Am I in support of this? Hell yeah!" wrote one commenter.

Flyer titled “Bye Bye AI” advertises drop-in sessions about limiting AI features and internet privacy. The event flyer went viral on Facebook, getting comments from around the world. https://www.facebook.com/share/p/1J7QUQUcU2/

"I had to look at the library's main page to see where Chatham is. I'm in central CA and wonder if we can have a digital guide that our own Mariposa County Library could use for a similar program?" wrote another.

"As someone in Ohio who just happened upon this post; will this be streamed or recorded?" someone else asked.

"As somewhere near Chatham, UK, I was a little confused at first Regardless, thanks Shea!" a jolly Brit wrote.

The library does weekly tech support drop-in hours, and in April it held an "AI for seniors" class that covered the basics of AI and LLMs. But this latest offering was something new. The idea was based on requests Wert had noticed from the community.

"There's just been enough patrons asking about certain things that bothered them, specifically Gmail auto replies and auto completes really frustrate people," Wert told Business Insider during a recent phone call. Want more Business Insider in your news feed?

Add BI in Google so our reporting is easier to find when you’re searching for what matters.

"I went to a presentation that another librarian was doing about this topic, and they had had success with it. And I was like, 'Okay, well, I've heard enough from my patrons that they're worried about this stuff or frustrated about this stuff that I might as well try to do one.'"

ai for seniros flyer The library held an "AI for seniors" class earlier this year that covered the basics. North Chatham Free Library

That other librarian, Hannah Cyrus, published a column in a journal of library studies, writing that "libraries have a role to play in facilitating patrons' rejection of technology, just as much as in their adoption of it."

Cyrus has offered a series of free classes for patrons, teaching them how to turn off or avoid AI tools on their phones and computers, with great success — and a much larger turnout than other tech classes. TechCrunch reported in July that Cyrus's article in the library journal inspired other librarians to reach out and ask for help running their own versions of the classes.

"I've heard from hundreds of librarians from all over the world asking for my presentation slides," she told Business Insider.

small white library building The North Chatham Free Library serves a tiny hamlet in upstate New York, with only about 800 library card holders. Shea Wert

One of the groups that reached out was a consortium of libraries for the Hudson Valley area, which included the North Chatham Free Library. Cyrus gave a webinar to a group of librarians, including Wert, about how she ran the "Avoiding AI" class.

In Wert's personal life, she admits to staying a bit of a Luddite, preferring to use a dumbphone. But she's not adamantly anti-AI and has used AI tools for work to draft outlines or write dense policy text.

"There have been some patrons that I know are using LLMs, and I'm happy to help them with it. I've helped them write better prompts, although that is far less common," she said. "It's more common that they haven't used it and they don't want to use it."

Both Wert and Cyrus said that most of the people who come in for help with AI or tech in general are older adults, who are often frustrated by new AI features on the services they already use.

The Bye-Bye AI event will take place next week, and Wert isn't entirely sure what to expect given the viral reaction on Facebook.

"I'm a little concerned based on setting it up as a drop-in thing that there'll be people waiting because I can only help so many at one time," she said. "But I don't know. If 20 people come, that would feel like a lot to me."

Read next

Katie Notopoulos

Katie Notopoulos

You're currently following this author! Click to unsubscribe from email alerts.

Katie Notopoulos is a senior correspondent at Business Insider who writes about technology, business, and culture. Previously, Katie was a tech reporter at BuzzFeed News and has written for The Atlantic, The New York Times, Fast Company, and MIT Technology Review.

AI