Yeah, bot no, but yeah, bot…

I was cc’d on an email exchange last week that I keep thinking about.

Two people, both senior, were going back and forth about the importance of using AI to create content. The emails were becoming longer. More detailed. Thoroughly argued. Full of structured reasoning, carefully weighted counterpoints, and that unmistakable cadence that AI gives to written English when left to its own devices.

They were arguing about whether organisations should embrace AI-generated content - using AI-generated content.

Neither of them appeared to notice, or appreciate the irony.

I read both sides. Or rather, I started to. By the third volley, each response had grown into something closer to a white paper than a reply. I suspect neither of them was reading the other’s emails in full either. They were pasting them into their respective AI tools, generating a rebuttal, and hitting send.

Two machines performing a debate while two humans supervised from a distance.

Is this another form of shadow AI? Not a rogue employee uploading client data to ChatGPT. Not a security breach or a compliance failure. It’s the quiet replacement of thinking with “output”...

The Pendulum

We have been through this cycle before.

In the olden days we had the handwritten letter. Every word costs effort. You think before you write because crossing out looks terrible - the same with the typewriter. The composition was the thinking.

Then word processing and email arrived and this was amazing. We got to edit cleanly, send to multiple recipients and attach pictures of cats playing the piano. Soon we were sending to everyone in the building to make sure we were “seen” to be doing our job and to cover our backs. I have worked in organisations where email overload nearly paralysed the operation. Everyone cc’ing everyone on everything, not because the recipients needed the information, but because the sender needed proof they had shared it. Cover-your-back culture dressed up as transparency.

My response to mitigate against this overload was often to set up wikis. Internal knowledge bases where information had a home, so it did not need to be broadcast to every inbox on the floor. Give information structure, and you reduce the need to spray it everywhere.

But even without that fix, we adapted. We stripped our responses back to one-line replies. “Noted.” “Thanks.” Eventually, the thumbs-up emoji as a complete response to a three-paragraph question. Communication became fast and thin.

Now AI has swung the pendulum violently in the other direction. That same person who would have sent a thumbs-up last year can now produce a 600-word response in seconds. It looks thorough. It feels responsive. It feeds the ego, because suddenly you sound articulate and well-prepared and decisive, even if you spent less time on it than the emoji would have taken.

The length is no longer a sign of thought. It’s a sign of delegation...

Brandolini’s Law

The Italian programmer Alberto Brandolini once observed that the amount of energy needed to refute bullshit is an order of magnitude greater than the energy needed to produce it.

He was talking about misinformation specifically, but the principle applies perfectly here. When AI can generate a detailed, plausible-sounding argument in seconds, the effort required to actually verify, interrogate, and respond to that argument properly is enormous by comparison. And nobody has that kind of time. The loop closes. The humans step further away.

What Gets Lost

The risk that most commentary focuses on is data security. Sensitive information going into public AI tools. That is real and worth addressing, but it is not the interesting problem to me.

The interesting problem is that meaning is draining out of professional communication. When I receive an AI-enhanced email, I do not know which parts the sender actually thought about and which parts the machine filled in. I cannot tell what they believe from what was auto-generated to sound convincing.

The signal-to-noise ratio has collapsed.

And I am guilty of it too. I have fed someone’s email into AI to draft a response. It felt efficient, not dishonest. That is precisely what makes it difficult to address. It does not feel wrong. It feels like good time management.

The Question I Keep Coming Back To

If both sides of a conversation are being generated by AI, and neither side is being fully read by a human, is it still a conversation? Or is it performance?

And if it is performance, what decisions are being made on the back of it?

I do not have a tidy answer to this. But I notice it more every week, and I suspect I am not alone.

Just me?

If you’re wondering what to actually do about it, we wrote up the three things that work (and the things to avoid!)