Strategic campaign slop

Working across forty countries ruins you for a certain kind of business model. You learn quickly how little you can assume. Then you learn it again. What carried an election in one place falls flat two borders away, and the reasons are almost always specific, local, and completely invisible from outside. It's a beautiful reflection of humanity in many ways.

But it is also inconvenient, because the most profitable thing you can possibly do in this industry is build a strategy once and sell it forty times.

Plenty of people do exactly that. It is a good part of what the consultant class I was complaining about yesterday actually sells — a whole approach, lifted from campaign to campaign and organization to organization with the names swapped out. It is an excellent business. It is also structurally incapable of noticing the differences that decide things.

Because strategy is very largely a question of position. What you should do depends entirely on where you are standing: what the terrain is, who is already organized and who is not, which relationships exist and which were broken last year, what a particular word means here that it does not mean two hundred miles away. Change one piece and the right move changes with it.

So a strategy built somewhere else was aimed from somewhere else. It does not miss by a little. It misses in the same direction every time it is deployed, and because the deck is coherent and the framework is complete, nobody in the room can see what it failed to account for.

Which brings me to AI slop. Let's start by establishing a definition.

Slop is not simply bad work. Bad work has a person behind it, and you can usually tell what they were reaching for. Slop is generic by construction. It is what comes out when the thing producing it is a prediction engine. Prediction, by its nature, returns to you the middle of the distribution. What does this mean? The most likely sentence. The most expected image. The average of everything already said on the subject, rendered confidently. (And usually in a group of three.)

You already know what this looks like, because it is on every noticeboard in your neighbourhood. Not the warped hands and the garbled lettering and the early tells that have largely been fixed. It is the sameness. The same composition every time, the same three-part layout, the same glow, whatever is being advertised and wherever it is going to hang. Nothing in the image production process is built around the room it will end up in.

I look at this and my eyes glaze over before I have read a single detail. Not because they are ugly. Because they announce, in the first half-second, that nobody was genuinely handed the creative part: the unique vision, the lived experience, the craft. And once you know that, why would you bother reading on?

We have a crisis of authenticity in our politics, and we built it ourselves, long before any of this. Decades of persuasion work have gone into making candidates resemble whatever a population is thought to want. Positions triangulated. Biographies sanded. Language tested until it means as little as possible to as many people as possible. Not bringing anyone to a case. Not inspiring anybody. Adjusting the product to fit the research and calling the result a campaign.

People can smell it. All of them, more or less instantly, and for years now.

They can smell the structural version too. They know they are being handled as a resource to be harvested for a vote or a donation at a frequency set by somebody's quarterly target.

Which is how we end up here.

People voted for the man who posts misspelled, unhinged, semi-coherent messages at two in the morning, apparently from the bathroom.

I do not think that is because people are stupid. I think it is because those messages are obviously, verifiably, unmistakably written by an unpolished human being, and almost nothing else in our politics is.

We have made bad-but-real a rational preference over good-but-fake. That is not the public's failure. It is ours. What looks like bad judgement from where we sit is the ordinary psychological response to being gaslit and manipulated for years.

It is also why the defence of democratic institutions keeps landing so badly when it comes unpaired with any recognition of the rot in the democratic culture those institutions stand on. When people who have watched those institutions drift away from them are told that the institutions must be protected, and the telling contains no acknowledgement whatsoever of the drift, the argument confirms exactly what they already suspected: that the people defending the system are the people the system still works for.

So consider the moment we have chosen to introduce the most generic patterning machine ever constructed.

The flyers are tier one. Embarrassing, obvious, and everybody can already spot them.

Tier two is the messaging, and it is already everywhere. Harder to catch than a warped hand, because prose does not have fingers. But the same flatness runs through it — predictable sentences, now produced faster than any editor can review them.

Tier three is the one that worries me, and it is the least obvious because it is not public facing. Strategic slop. Not the machine writing the email, but the machine deciding what the campaign is.

I want to be careful here, because there is a loose version of this complaint that I do not hold and that I hear constantly. The loose version is that AI has no business anywhere near strategic work at all. That is wrong, and it is wrong in a way that will cost us — it hands the field to the people making the opposite mistake, and it is not what I have spent the last two years building.

So let me be exact, which means being exact about a word our whole industry uses sloppily.

Tactics are what you do. Strategy is the choosing — continuously, while the ground moves. Marshall Ganz's line, which I used yesterday, is that strategy is a verb rather than a noun: a creative, continuous stream of tactical adaptation. Not a plan you own. A thing you keep doing, in contact with a situation that will not hold still. Tactics can be written down. Strategy cannot, because by the time the deck is printed the position has changed.

And AI gets more useful the further you move away from that.

At the implementation end it is genuinely good, and I use it every day. Turn this into forty local versions. Draft the follow-up nobody has time for. Find the thing buried in the spreadsheet.

In tactical planning it is useful. Sequence this launch. Sort these two hundred notes into themes. What are we forgetting. What would somebody who disagreed with us say first.

And in strategy it has a role as well — a role we named long before computers existed. It is the scribe. The best strategy sessions I have ever sat in had somebody in the corner writing everything down: catching what was said, reading it back, noticing that nobody had answered the third question, pulling forty people's contributions into something the room could look at. That is real work, and a session without it is worse. Nobody has ever confused the scribe with the strategist.

Which is the whole of my worry, stated properly. Not that the machine will do strategy well. That it will produce something shaped like strategy — coherent, confident, complete, nicely formatted — and nobody in the room will be able to see what it could not see. The same failure as the imported deck I started this chapter complaining about, except now it is free, instant, and arriving at a volume no one can review.

There is good evidence for exactly this shape, and it is the reason I am not making the loose argument. A large field experiment run with Boston Consulting Group — Ethan Mollick was among the authors — gave hundreds of consultants access to GPT-4 across two different tasks. On the creative task, inside what the researchers called the jagged frontier of the model's competence, quality rose by more than forty percent. On a business problem where the available data was subtly misleading, outside that frontier, the consultants using AI were nineteen percentage points less likely to reach the right answer than the ones working without it. Same tool, same people, opposite result. And the ones who did worst were the ones who interrogated the output least.

Which brings me, unavoidably, to the strawberry.

A language model does not read words. It reads tokens — fragments. Strawberry arrives as pieces, which is why for a couple of years models could not reliably tell you how many r's it contains. They have been patched since, and the patch is the instructive part: they were taught to spell it out and count, one letter at a time. Which is not the same as ever having known.

It has never seen the word. It has seen the tokens for ‘straw’ and ‘berry’. It hasn’t seen the word ever. It’s never seen a real strawberry either. And it has certainly never tasted one.

‘Strawberry strategy’ of the AI is what we are about to get an enormous quantity of — plans that have never seen the place they are for, never smelt the air, and never know the face of a constituent. And here is what makes it worse than the flyers. A strawberry tastes more or less the way it did when I was a child. A community does not. The situation you are strategizing about is changing while you read the deck.

None of which, finally, is really a claim about silicon, and this is the part I most want to land. A consultant alone in a hotel room with a framework fails in exactly the same way. So does a brilliant executive director who stopped going to meetings four years ago. Strategy done well is collective — not because collaboration is pleasant, but because no single mind holds enough of the context. You need the person who knows what happened on that street last year, and the one who knows what that word means to her congregation. A person in a box cannot do strategy either.

The failure is not artificial intelligence. It is disconnection from community by the planning class of our politics. AI is simply the fastest and cheapest method we have ever invented for industrializing it.

I will come back to all of this properly, because it deserves its own chapter and it is going to get one.

And all of this sits on top of an awkward fact, which is that the output does not work anyway.

I said earlier in the week that I would come back to how thoroughly this is settled, so. Kalla and Broockman's meta-analysis across forty-nine field experiments found the persuasive effect of campaign contact in general elections to be approximately zero. Coppock and colleagues found the same, and found that it held regardless of context, message, sender or receiver. In 2024, Democratic campaigns and allied groups spent around four and a half billion dollars on advertising, a billion more than the other side, and lost every swing state.

That is about as settled as social science ever gets. And I should say that I have been on the selling side of it. For a long stretch of my career, a large part of what my industry sold was efficient participation. Lower the barrier. Fewer clicks. One-tap action. Sign here. Every incremental reduction in effort was a number we could improve, so we improved it.

Here is what I think we got wrong.

Participation that costs nothing gives nothing back. And it does something worse than nothing, because it discharges something. A person who signs a petition has, in their own mind, done the thing. The itch is scratched. Whatever responsibility they were carrying toward the issue — the responsibility that might have taken them to a meeting, or a conversation with a neighbour, or a room with other people in it — has been quietly settled for the price of an email address.

I am now fairly sure petitions are doing more harm than good at this point. Not because collecting names is wrong. Because we allowed the name-collecting to become the participation, and then told everybody that this was democracy.

It is the trash bin of civic life. You did a thing. You feel better. The system that produced the problem is entirely undisturbed.

So what is actually scarce here is not reach, and it is not content. We have more content than at any point in human history and less trust than at any point in mine.

What is scarce is the sense that a real person, who knows something, meant this.

That cannot be generated. It genuinely cannot. No amount of model capability gets you there, because the thing being detected is not quality, it is authorship.

But it can be supported, and the difference matters more than anything else I will write this month. There is an enormous distance between a machine that writes the message for you and a machine that helps you write the message you would have written yourself, if you had the time, the craft, and somebody to think it through with. The first one produces slop. The second one produces you, at your best, more often than you manage alone.

That is what my team has spent the last stretch building, and on September 15th we unveil it.