The Signal You Can't Get Back

Hey there,

On a recent episode of the Acquired podcast, Ben Gilbert was trying to explain why their creative process can't be replicated even if they describe it in detail. He said something that's stuck with me ever since: "Language is a lossy compression of thought." He walked through the idea that when he communicates something to David, he's compressing a thought into a narrow bandwidth of speech. David uncompresses it in his own brain. And what David reconstructs might mean something meaningfully different than what Ben intended. Not wrong, necessarily. Just different. Some of the original signal is gone.

A few weeks later I was on a call with Eric Turkington, a CPO in healthcare AI, and that thought surfaced in a completely different context.

We were talking about how AI now mediates the way teams communicate, and I floated this idea: every layer of AI mediation is another round of lossy compression. An idea in your head is the purest version. Explaining it to someone, you lose something. Writing it down, you lose more. Running it through AI to polish it, more still. And when the recipient feeds your polished output through their AI to summarize it, you've passed through four or five layers of compression. What arrives might be coherent. It might even be elegant. But you can't reconstruct the original idea from it anymore.

Eric simplified it: you can't reverse engineer the signal.

Lossy compression in communication isn't new. It's how language has always worked. Every time you translate a thought into words, something gets lost. Every time someone reads your words and reconstructs meaning in their own mind, something shifts. This has been true since humans started writing things down.

What's new is that we've outsourced the compression itself.

Think about how a typical strategic communication used to move through an organization. A product leader had a rough, half-formed insight about a market shift. She sat down and wrestled it into a memo. That wrestling was cognitive work. She decided what to include, what to cut, how to frame it. She chose which caveats to keep and which to drop. The memo was lossy, inevitably, but she remained the compressor. She knew what she'd left out. If her VP read it and asked a follow-up question, she could recover the context that didn't make it onto the page. The uncompressed version still lived in her head.

Now watch what happens when AI enters both ends of that exchange. The product leader dumps her rough thinking into AI and asks it to draft a memo. AI decides what matters, what to cut, how to frame it. It removes the hedges, tightens the caveats into confident assertions, drops a weird aside about a customer conversation that might not be relevant. The memo reads beautifully. But the product leader didn't do the compression. She may not even notice what's missing, because the output looks complete.

Her VP receives it, skims it, and asks his AI to pull out the key takeaways. Three bullet points arrive. Crisp. Actionable. Completely disconnected from the anxiety she felt about the market shift, the uncertainty she was trying to signal, the customer conversation that was bugging her for reasons she couldn't fully articulate yet. The VP didn't do the decompression either. His AI did.

Both people feel productive. Yet neither one did the cognitive work of engaging with the material. And nobody in the chain holds the uncompressed original.

If you grew up playing Operator (or Telephone), you already understand the mechanics. A message passes through a chain of people. Each person compresses and decompresses it through their own understanding. By the end, the message has drifted. Everyone laughs.

But there's something important about the original game: every person in the chain at least processed the message through their own mind. They engaged with it. They might have garbled it, but they thought about what they heard before passing it on.

What we've built with AI-mediated communication is an Operator game where the players hold the phone to a machine instead of their own ear. The message might actually be more accurate at each individual handoff. AI is probably a more faithful compressor than a distracted human. But the humans have removed themselves from the chain. They're no longer doing the cognitive work of understanding what they received or deciding what to pass along.

That's the trade we've made. More fidelity per link, less human engagement per link. And it turns out the engagement was doing something important that we didn't appreciate until we optimized it away.

I wrote about compression architecture last October in "The Recombination Illusion." That piece was about what happens inside the machine: AI excels at compression with constraint but can't expand beyond its training distribution.

This is about what happens between people when they stop doing their own compressing and decompressing.

Eric called it the "sanded edges." Each AI layer sands off the rough parts of the original thought. What emerges might capture the essential shape, but it misses what he called the messiness of the original conception. Hesitation. Tone. The specific word choice that betrayed someone's uncertainty. The half-formed thought that would have sparked a follow-up question if you'd heard it directly.

The messiness was part of the signal.

The danger isn't that AI gets things wrong. It's that AI gets things close enough that nobody checks what was lost.A polished artifact doesn't invite scrutiny. It looks finished. A rough draft with hedges and visible uncertainty signals to the reader that the thinking isn't settled, that a conversation is needed. AI strips those signals out because it's optimizing for clarity. And clarity, in this context, is the enemy of understanding.

On the call with Eric, I raised the idea that there needs to be some kind of verification mechanism. A handshake. You read back what you heard from me, and I tell you whether you've understood me correctly. A recursive loop of truth-seeking.

Current AI-mediated workflows don't have this. There's no read-back. The polished artifact looks complete, so nobody asks what was lost in translation. We've built a communication infrastructure that optimizes for speed and surface quality while systematically eliminating the feedback loops that catch misunderstanding.

If you lead a team, this should worry you. Not because AI is producing bad content. It's usually producing very good content. But "good content" and "accurate representation of what someone actually thinks" are increasingly different things.

The conversation that produced this piece was messy and iterative. Eric and I riffed on an idea for thirty minutes, built on each other's language, interrupted each other, and arrived at something neither of us had when we started the call. That's what unmediated human communication does when it works. It creates meaning that didn't exist before the exchange.

The question I will leave you with: how many of your team's most important communications now pass through layers of AI mediation before they reach the person who needs to understand them? And when they do, is anyone on either end still doing the cognitive work of compressing and decompressing the ideas themselves?

Or have we all just started holding the phone to the machine?

Break a Pencil,

Michael

P.S. The compression framework I referenced from "The Recombination Illusion" is foundational to how I work with product teams on AI adoption. If your organization is stacking AI tools without thinking about what's getting lost between people, let's talk about that.

P.P.S. Know someone who's been polishing every communication through AI without thinking about what's being stripped away? Forward this to them.