AI Can’t Feel Water Pressure

Ruben Hassid recently ran an experiment - he asked Claude to write a 150-word summary explaining the origins of his name. Taking the draft, he re-prompted the engine with a simple directive: “Anti-AI rewrite this.” When he fed the resulting text into Pangram, an AI detection tool boasting near-perfect accuracy, it flagged the passage as 100% AI-generated.

Then Hassid made a small, seemingly trivial adjustment and replaced an em dash with a colon.

He ran the text through Pangram again, and the score instantly plummeted to 0% AI.

The ease with which we can trick AI detection software is almost comical. Swapping punctuation marks or tweaking stylistic quirks can outmaneuver tools built to catch them. Yet, the obsession with these detectors reveals something far more important, which is that an internet drowning in synthetic noise (aka AI slop) is desperately seeking a mechanism to verify human-led content, a proxy for trust. Consider travel as a vertical.  If a consumer is going to spend hard-earned money and time on a trip, they want to make sure they know what they’re getting. For now, a person who has been to a hotel, felt the water pressure in the shower, then wrote about their experience with that water pressure beats AI any time. Those human signals that ultimately feed the synthesized answer that a platform like ChatGPT provides are critical in establishing trust with the consumer.

Things are coming to a head now because automated bot and AI agent traffic account for more than half of internet activity. For the first time, bot activity has eclipsed human presence online, with 57.4% automated versus 42.6% human, and humans can tell. On LinkedIn, estimates suggest up to 40% of posts are entirely AI-generated, produced with annoying emojis and in seconds without any original thought. It’s quickly losing consumer trust. 

According to a Harris Poll, 73% of consumers say they have less trust in content they suspect was generated by AI (frankly, I am shocked that percentage isn’t higher).  It fills their feeds with dead giveaways like negative parallelism and awkward vocabulary ("unleash," "elucidate"...humans don’t speak this way). They’re fatigued by the slop.

Major publishers and social platforms are responding with human verification features. Not out of corporate benevolence, but to preserve audience retention. 

Substack’s latest feature, powered by Pangram, allows readers to scan posts to determine whether content was AI-generated or human-led. It’s an initiative that Chris Best pitched directly against "Claudefishing." Meta has mandated "Made with AI" labels across its feeds while tweaking algorithms to downrank unoriginal synthetic posts. Pinterest deployed anti-slop filters that give users explicit controls to cap the volume of AI-generated pins in their feeds. Real human thought is rapidly becoming the primary currency of platform authority.

Skeptics argue that this concern is overblown. Low-quality content, clickbait, and spam have existed since the dawn of the public web, and readers have always had to curate their own media consumption to find the best content.

But, we aren’t dealing with legacy spam, which was often obvious and lived on the margins of the internet. AI slop scales fast, at nearly zero marginal cost and mimics authoritative structure and tone. It offers a pretty convincing (and getting better) illusion of expertise without a single drop of original analysis or lived experience, which is especially tough for hospitality and travel, where human stories and experiences are what drive the highest-value stays.

To be clear, I’m an enormous AI advocate. I use it to draft emails, synthesize data and evaluate job candidates. I build agents to manage armies of sub-agents. Occasionally, AI has better ideas than mine. I use Claude to push my thinking further, condense reams of data, and help me write faster.  If an AI can help, I figure out how to integrate it into my workflow. But I draw the line where AI becomes the sole ideator; where a human with a unique point of view comes completely out of the loop. I want LLMs to generate recommendations based on real expertise and authority - we all should.

That’s why I think AI models and search engines will be forced to adapt to growing consumer demand for quality human content. Within a year, I’m confident that AI answers will have to do more than just cite a source. I envision something like an explicit “Authority Ranking” based on verified human content.

What does that look like? That remains an open question. Is content authoritative at 50% AI assistance? 15%? I think a percentage score ultimately matters less than the reader’s perception and trust. Consumers have a visceral intuition for good, trusted content. When the human element is there, and it speaks to them, they engage; when it’s not, they don’t. If an audience remains unconvinced that a real human (a trusted human) stands behind a neatly synthesized AI output, the outcome is clear. They won’t buy the product, the service, or the idea.