Pangram found AI in more than 40% of the longform LinkedIn posts it analyzed. What that means for trust in online writing

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?

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.

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.

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.

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.

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.

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.