Ad Machinam

On a logical fallacy we thought we had outgrown — and what it is costing us. — from A Renaissance of Thought by Jake W. Casselman.

On a logical fallacy we thought we had outgrown — and what it is costing us.

If you have been following this series, you now have a few things in hand. A name for the tool — the conversational posture, the Socratic workbench — and a name for what it makes possible — the minimum viable scaffolding that allows curious people to cross into fields that were never theirs. And, from the last chapter, an honest accounting of what these tools can cost you when you use them badly.

The argument began optimistic. Deliberately so — and then it turned, just as deliberately, to give the counterargument its due. These tools are genuinely new, and the possibility space they open is genuinely large. The dangers are real too.

But there is something happening in the culture around AI that deserves honest examination. Not the doomer concern — that has been covered exhaustively elsewhere. Something quieter, more pervasive, and in some ways more corrosive. A logical error that we spent centuries learning to identify, and are now committing again without noticing.

A Rule We Already Established

There is a principle in clear thinking, old enough to have a Latin name, that most people know intuitively even if they have never heard the term.

It works like this. Someone makes an argument. Instead of engaging with the argument — examining whether it is true, whether the reasoning holds, whether the evidence supports the conclusion — you dismiss it by pointing to who made it. He has a financial interest in saying that. She has never run a business. He went to the wrong school. The argument never gets evaluated. The person does.

This move has a name: ad hominem. To the person. It is considered a logical fallacy not because the speaker’s identity is never relevant, but because identity tells you nothing about the truth value of an idea. A corrupt politician can state a true fact. A child can identify a flaw in expert reasoning. The idea stands or falls on its own merits, independent of the vessel it arrived in.

We know this. It is taught in schools, invoked in debates, cited whenever someone wants to signal they are thinking clearly rather than reactively. Attacking the messenger rather than the message is understood, broadly, as the mark of a weak position.

What We Are Doing Now

Run a simple experiment. Post something well-reasoned, carefully structured, genuinely useful. Then mention, or simply imply, that AI was involved in producing it. Watch what happens.

The engagement changes. The credibility shifts. The work is evaluated differently — not because anything in the work changed, but because the perceived source changed. The idea is the same. The logic is the same. The insight is the same. But the source is now suspect, and so the work becomes suspect.

This is ad machinam. To the machine. The same logical move, aimed at a new target.

It is worth stating plainly: this is not a principled epistemological position. It is the old fallacy in new clothes. If we accept that source identity does not determine idea quality — and we do accept this, we built a whole framework around it — then the source being a machine, or a human working with a machine, does not change the analysis. The idea is still either true or false, useful or useless, well-reasoned or not. The vessel it arrived in remains irrelevant.

The Survey

A while ago I ran a small survey on LinkedIn. The question was simple: have you changed the way you write to avoid appearing like you used AI?

The results were surprising. Not because a few people said yes. Because many people said yes — and said it without embarrassment, as if it were the obvious and reasonable response to the situation.

People were deliberately avoiding certain sentence structures. Removing particular punctuation. Roughing up their prose. Introducing small imperfections. Performing humanness.

The LinkedIn poll, as it closed.

Sit with that for a moment.

We have arrived at a situation where people are degrading their own authentic voice in order to satisfy a heuristic that has no logical foundation. The em dash — a punctuation mark with a long and legitimate history in English prose — became suspect because AI used it frequently. So people stopped using it. Not because it made their writing worse. Because it made their writing look like something they were now afraid of being associated with.

The heuristic swallowed the thing it was supposed to measure.

System 1 Wearing a Serious Face

There is a useful distinction in psychology between two modes of thinking. Daniel Kahneman, in Thinking, Fast and Slow, called them System 1 and System 2. The first is fast, automatic, pattern-matching — it runs in the background, makes quick assessments, operates on shortcuts and heuristics. The second is slow, deliberate, effortful — it examines assumptions, evaluates evidence, actually does the work of thinking.

Scanning a piece of writing for em dashes and clean sentence structure as a proxy for AI involvement is System 1. It feels like critical discernment. It is not. It is pattern recognition dressed up as judgment, moving so fast that it never has to make contact with the actual question: is this any good?

Actually evaluating whether an idea is well-reasoned, whether the argument holds, whether the conclusion follows — that is System 2. It is harder. It takes longer. It cannot be done by checking for suspicious punctuation.

The intellectually honest version of the question is always: is this true? Is this useful? Does this hold up? Not: what was the process by which it was produced?

The process question is easier. That is the only reason it has taken over.

What The Fallacy Is Protecting

Logical fallacies rarely persist without a reason. The reason is usually that they protect something — a position, a status, a way of organizing value.

Ad machinam is protecting the conflation we examined in the previous chapter. If writing and thinking are the same thing — if the quality of the transcription is the quality of the mind — then a tool that assists with transcription threatens the entire framework. The people whose credibility rests on their prose style, whose authority is signalled by their command of language, whose position in a hierarchy of ideas is maintained by the difficulty of producing polished writing — those people have a structural interest in maintaining the idea that AI-assisted output is categorically lesser.

This is not a personal failing. It is a rational defence of an existing value system. But it is still the fallacy. And it will not survive contact with the question it is trying to avoid: is the idea good?

The Self-Defeating Loop

Here is the part that should concern everyone, regardless of where they stand on AI.

The em dash heuristic — and all the variations of it — is a proxy. It was never measuring quality. It was measuring surface features that happened to correlate with AI output at a particular moment in time, in a particular stylistic register. As a quality signal it was always weak.

But now it is not even that. It is a heuristic that human writers are actively optimizing against. People are roughing up their prose, avoiding their natural cadences, introducing deliberate friction into their writing — not to improve it, but to survive a filter.

The signal has been corrupted by the thing it was trying to detect.

And the writers are the ones paying the price. Their authentic voice — the thing that actually carries genuine value, the thing no tool can replicate — is being sanded down in service of a test that was never measuring what it claimed to measure.

This is the cost of ad machinam stated concretely. Not a philosophical problem. A practical one. Real people, writing real things, making their work worse to avoid a fallacy.

Not Every Word Needs To Be Yours — But Some Do

None of this is an argument that writing doesn’t matter. It is an argument that writing’s value is not uniform across all contexts, and that we should be honest about when the transcription layer needs the human and when it does not.

Consider the range. At one end: an HR form, a meeting summary, a standard email confirming a time. The thinking required is minimal. The stakes are low. The idea, if there is one, is simple and does not need to be wrestled into existence through the act of writing it. Assistance here costs nothing because there is nothing to lose. The resistance is just friction.

At the other end: a book. Something meant to last. Something where the ideas are load-bearing and the writing is not merely carrying them but forming them. This is where the tool of writing stops being neutral. When you sit with a sentence that will not cooperate, when the paragraph refuses to close, when you write and rewrite the same passage because something about it is not yet true — that resistance is not an obstacle. It is the process. The friction is the thinking. The difficulty of getting the words right is inseparable from the difficulty of getting the idea right. Removing it removes something real.

Between those ends sits the blog. And this is where the distinction becomes practically useful.

The purpose of writing a blog — at least in this series — is not to produce a finished artifact. It is to get an idea into the world fast enough to test it. Minimum viable scaffolding for the ideas themselves. If the concept is real, reality will confirm it. If it has holes, readers will find them. The iteration cycle matters more than the polish. Here, AI assistance lowers the cost of getting the core idea out without distorting the idea itself — and that is a genuine advantage, not a shortcut.

But the blog is not the destination. The ideas that survive, that prove load-bearing under pressure, that deserve to last — those ideas eventually need to be written by hand. Not because AI assistance is illegitimate, but because the resistance of writing is itself a test. If you can sit with every word and know why it is there, the idea is ready. If you cannot, it is not yet finished.

The question is not whether to use the tool. The question is whether the stakes are high enough to require the resistance.

A Note on the Name

The term already exists. I am not the first to coin it.

I arrived at ad machinam independently, working from the parallel with ad hominem — but the parallel is clean enough that other people got there first, and it would be dishonest not to say so. The earliest documented use I can find is from August 2024: a Stratnav blog post titled Embracing AI in Strategy: Beyond the Argumentum ad Machinam, which defines the same concept directly — "The term argumentum ad machinam parallels the concept of argumentum ad hominem, where the focus is wrongly placed on the creator rather than the content. In this AI-driven variant, the emphasis is improperly put on the use of AI tools, ignoring the quality and substance of the work itself."

There are others. Will Falk has a LinkedIn post, Introducing ad machīnam: A new fallacy for AI-generated…, from earlier this year. And an October 2025 Medium piece by Anecoica Studio, Beyond Ad Machinam, frames it as "the negative authority fallacy — an ipse dixit with a minus sign. It deems a claim false because it comes from a disfavored source (a machine). A related move is the modern variant of the argumentum ad hominem: the argumentum ad machinam."

I take the convergence as confirmation rather than competition. When several people, working separately, reach for the same Latin construction to name the same error, that is evidence the error is real and worth naming. The point was never to own the term. The point is that the fallacy is common enough to need one.

The Standard We Already Hold

There is nothing new required here. No new principle, no new framework, no new rule.

The standard already exists. We established it. The idea stands or falls on its own merits. Source identity is not evidence of quality. Attacking the vessel rather than the contents is the weaker move.

Apply that standard consistently — which is all any logical principle asks — and ad machinam dissolves. The em dash is either helping the sentence or it is not. The argument is either sound or it is not. The insight is either useful or it is not.

The question was never where it came from. The question was always whether it is true.

We knew this. We just forgot, briefly, in the excitement of having a new target.

What This Means For The Series

In the first chapter we established a new way of using these tools — not transactional, but conversational. In the second we showed what that posture makes possible at the boundary between fields. In the third we gave the counterargument its due — the real costs, the genuine dangers, the case against. In this one we have cleared some of the debris that accumulates around any genuinely new thing: the reflexive rejection that masquerades as principle.

The examined life, as we said at the start, was never only an internal project.

It also requires examining the frameworks we use to dismiss what we have not yet decided to understand.

Take-home reference card Ad Machinam — the logical structure A one-page companion: the two fallacies side by side, the legitimate exception, and the substitution test.

PDF →

Written in Honolulu, June 2026