Reliable Signals of Honest Intent

The story of Microsoft's Windows NT 3.1 launch in 1994 offers a surprising lesson for our current moment. Faced with convincing system administrators to adopt a new 32-bit server operating system, Microsoft didn't rely on technical specifications or feature lists. Instead, they hired an advertising agency that produced an elaborate box containing a free mouse-mat and a pen, all packaged in gratuitously expensive materials. This wasn't just marketing fluff—it was a deliberate signal. In an attention economy where thousands of stimuli compete for our focus, the elaborate packaging served as a reliable indicator that what lay inside was worth the recipient's time and consideration. The strategy worked remarkably well: nearly all boxes were opened, and about 10% of the technically sophisticated server administrators actually tried the new operating system—a conversion rate that would be impressive even today.

This example from Rory Sutherland's Alchemy illustrates a fundamental principle that extends far beyond software updates. When we encounter any piece of communication—whether a software update, a job application, or an article—we're not just processing the objective information. Beneath our conscious analysis runs a rapid, subconscious assessment of intent. We're asking: Is this worth my time? Is the author honest? Do they know what they're talking about? This evaluation happens faster than conscious thought, operating as a survival mechanism repurposed for navigating the modern information landscape.

The human brain has evolved sophisticated pattern recognition systems for this purpose. Consider the difference between a hiker identifying a bird species in the Alps and a rabbit determining whether a bird is a predator. The hiker consults a bird book, carefully examining multiple features to make a precise identification. The rabbit, however, needs only to answer a binary question: predator or not? This rapid categorization happens pre-intellectually, what researcher R. Horsey calls a "gestalt phenomenon." When we read online content, we're often more like the rabbit than the hiker—making snap judgments about quality and authenticity before we've even processed the specific arguments.

This instinctive evaluation has become particularly acute with the rise of large language models. We've developed a kind of paranoia about AI-generated content, and for good reason. The tells are often subtle: awkward lists of three, parallel sentence structures, certain predictable phrases. But even when these obvious markers are absent, we sense something off. The text feels generic, lacking the friction of genuine human thought. This isn't mere prejudice—it's pattern recognition at work. When we've consumed enough AI-generated content, we begin to associate certain forms and rhythms with artificiality, even when we can't immediately articulate why.

The tragedy lies in what happens when humans attempt to "improve" their authentic writing through AI assistance. Consider the job applicant who crafts a heartfelt, slightly clumsy cover letter, then runs it through an AI to "polish" it. What they've done is replace the very thing that would have made them memorable—their unique voice, their imperfect but genuine expression—with something that sounds like everyone else. The elaborate packaging worked for Microsoft because it signaled investment and care. Your own words work for the opposite reason: they're unmistakably yours. Imperfect prose, idiosyncratic turns of phrase, slightly awkward sentence structures that nevertheless express exactly what you mean—these are proofs that someone sat with a problem and wrestled with it personally.

When you launder your thought through AI, you remove proof instead of adding polish. The result is fluent, structured, professional text that's indistinguishable from thousands of other posts. The reliable signal of honest intent disappears. You might as well have sent just the prompt.

Technologists often object that this problem is temporary. "When the technology progresses further," they argue, "the rough edges that signal humanity will be reproducible by the models." But this argument misunderstands both the nature of progress and the nature of human perception. While the gap between early GPT-3 and ChatGPT was indeed vast, the improvement from GPT-4 to what comes next is likely to be incremental rather than revolutionary. AI research faces diminishing returns and staggering costs. Training new models requires eye-watering amounts of money, and there's a theoretical limit to how many parameters can be tweaked before improvements plateau.

Meanwhile, human pattern recognition accumulates linearly with exposure. As we consume more AI-generated content, we get better at detecting it. The gap between model capability and human detection might narrow, but it's unlikely to close completely. More importantly, the signal we're detecting isn't just technical imperfection—it's the presence of a thinking, caring human being on the other side.

This brings us to the core of what makes writing valuable. You cannot fake having been there. You might produce more content by prompting a language model, but the author who sits at the desk, who cares enough to pour over every sentence and choose every word with deliberate purpose—that author will be the one who gets read. Because they understand that the reader is worth the trouble. Because the text has friction. Because it has opinions that cost something to hold. Because it couldn't have been written by anyone else, for anyone else, about anything else.

The elaborate box Microsoft sent to system administrators wasn't just about the physical object—it was about the signal it sent: "We value you enough to invest in this interaction." In the same way, when we write with our own words, we're sending a signal to our readers: "You are worth the effort of my genuine attention." This signal cannot be automated because it's fundamentally about the investment of time, care, and authentic presence.

Slop predates AI. Before ChatGPT, we had ghostwritten engagement bait on LinkedIn. Before LinkedIn, we had SEO-optimized keyword listicles. We learned to detect those too. The question isn't whether AI will eventually become indistinguishable from human writing. The question is whether mimicry was ever the thing we were detecting in the first place. The signal was never really about technical perfection or even detectability. It was about whether you showed up. And you cannot fake putting in the time.