The Day the Spin Doctors Meet the Machine

Why politicians are about to have a frightfully unpleasant decade — and why, on balance, this might be rather good for the rest of us.

It’s 2030. You’re sitting at home watching a political debate. Two chaps in suits, lecterns, the usual ghastly business. One of them is doing what politicians do — sliding past awkward facts, smuggling in loaded words, telling you that his opponent voted seventeen times against puppies and Christmas. You half-notice. You always half-notice. But it slides past because it always slides past, because the human brain, marvellous though it is, was not designed by natural selection to score the rhetorical hygiene of strangers in real time. We were designed to nod along with our tribe and feel vaguely uneasy about the other one.

Now imagine a small bar at the bottom of your screen. As the chap speaks, it lights up. Loaded language: 3.2 times the rate of his opponent. Cherry-picked statistics: four. Strawmen: two. Pattern over the last six months: consistent. You glance at it. You don’t have to read a thinkpiece. You don’t have to consult a fact-checker the next morning. The number is just there, the way the score is there at a football match.

This is, I am reliably informed, more or less imminent.

Why this works when fact-checking doesn’t

Here is the thing that the political-science chaps have established beyond serious doubt over the last two decades: telling people that the politician they like has said something untrue does almost nothing. You can publish all the Snopes articles you wish. The voter shrugs, mutters something about the lying media, and proceeds. Humans are not Bayesian reasoners updating gracefully on evidence. Humans are tribal primates running on coalitional software, and the software treats inconvenient facts the way the immune system treats a transplant.

So you might think: well, if fact-checking doesn’t work, this rhetoric-scoring business won’t work either. People will simply ignore it.

But it’s not the same mechanism. And this is the bit worth dwelling on.

Fact-checking attacks individual claims. Did he say the thing? Was the thing true? The voter’s tribal software has excellent defences against this — denial, whataboutery, source-impeachment. But what the new tools attack is pattern. Not “this statistic is wrong” but “this fellow uses three times more manipulative language than the other fellow, every time, over six months, across every venue.” That’s a different sort of signal. It bypasses the truth/falsity argument entirely and goes straight to character. And character — the perceived sort of person somebody is — is precisely the variable that the tribal primate brain does update on. It is, in fact, the variable it was built to update on. Our ancestors weren’t checking each other’s sources. They were watching, over years, who was reliable and who was a slippery sod.

So this is, if I may, the cable-news mechanism in reverse. Cable news weaponised pattern recognition for tribal loyalty: here are all the times the other side did something bad, repeatedly, across years. It worked beautifully. It also corroded everything it touched. The new tools take the same mechanism and point it at rhetorical conduct itself, irrespective of which tribe is performing it. It’s the same psychological lever but a different target.

What I’d expect to follow

A few predictions, dear reader, for which you may hold me to account in the fullness of time.

The score becomes the story. At the moment, the morning after a debate is decided by pundits — that grim parade of former staffers and PR people telling you who “won.” Within a few election cycles, post-debate coverage will lead with the asymmetry numbers. “Senator Plonker used 3.2 times the loaded language of his opponent — here is the breakdown by technique.” Pundits will hate this, naturally, because it commoditises pundits, who are extraordinarily expensive replacements for what is essentially a confidence trick. But networks will license the rhetoric-scoring rails anyway, because the numbers will rate. Spectacle finds spectacle.

Speechwriters adapt — and here we have a fork in the road. Two possible futures, both plausible, neither yet decided.

The cheerful one: speeches actually clean up. The speechwriter, knowing that obvious manipulation will be flagged in real time on screens around the country, edits it out. Baseline rhetoric improves. We get something a touch closer to grown-up political discourse.

The grim one: speechwriters learn precisely where the classifier is blind, and write around it. We get AI-clean rhetoric — still misleading, still manipulative, but on dimensions the model isn’t yet measuring. This is what the engineers call adversarial co-evolution, and historically the optimisers win, because there are thousands of them and only a few classifiers. The honest answer is that we’ll get some of both, with the second probably outpacing the first unless the rubrics keep evolving — which means the people building these tools have to treat them as a permanently moving target, not a finished product.

The town hall dies. The AI grilling replaces it. A clip went around recently of Bernie Sanders being interviewed by Claude. It was striking because it bypassed two layers that politicians ordinarily rely on. Friendly interviewer, friendly audience: gone. Replaced with an interlocutor that doesn’t get tired, doesn’t get charmed, doesn’t get intimidated, and doesn’t have a producer in its ear telling it to wrap up because the next segment is on weather. If even two or three serious candidates do well in that format, and one or two visibly duck it, then refusing the AI grilling becomes a political signal — the way refusing presidential debates was a damaging signal between roughly 1960 and 1976. By 2028 I’d expect a “live AI Q&A” to be a major campaign format. The candidate who runs as the AI-tested candidate will probably be a fellow with a strong substantive record and feeble charisma. The format will favour him for the same reason a written exam favours the swot over the chap who can charm the examiner in viva.

Charisma is decomposed for parts. And here, dear reader, we approach a topic dear to my evolutionary heart. Charisma at present is a black box — “she’s just somehow likeable,” “he has presence,” “there’s something about him.” Once the machines start telling you that the something is, in fact, 73% vocal cadence, 8% argument structure, and 19% well-deployed tribal cues, the mystique evaporates. Charisma has historically been an enormous political asset — it is, after all, how primates have selected leaders since long before language. Strip it down into measurable components on a screen, and it loses much of its black-box premium. Reagan, in this environment, is a slightly diminished Reagan. Some painfully wonkish fellow nobody can presently bear to listen to becomes, relatively speaking, more powerful. Whether that’s good for civilisation is genuinely unclear. Charisma is, among other things, how democracies have built coalitions across difference. Take it away and you may find yourself with more honest politicians and less functional politics.

The fifteen-second gotcha clip dies. At present, a great deal of journalism consists of catching a politician contradicting himself in a single clip and circulating it for a day. Once we have proper longitudinal profiles of how every politician speaks and writes over years, the single contradiction matters less — because it’s no longer a slip, it’s a data point in a pattern. Patterns matter more. Journalism shifts from clip-hunting to pattern-narrating. The politicians who survive are the ones whose patterns hold up under that scrutiny. Most won’t.

The earpiece scandal is coming. Once the tools exist to grade rhetoric in real time, they also exist to coach it in real time. Some candidate, somewhere — probably first in a city council race or a congressional primary, where standards are slacker — will be caught wearing a concealed device fed real-time rebuttals from a staff-run AI. Once it’s caught, it will be a scandal. Once it’s normalised, it will be the norm. I should rather like to know when the first big detection lands. I’d guess within five years.

And the deepest worry

Now the awkward bit.

Once politicians start optimising for the classifier, the classifier stops measuring what it was supposed to measure. Goodhart’s Law — the moment a measure becomes a target, it ceases to be a good measure. The honest version of any such tool keeps reminding the user: this is one frame on framing; there are others; reasonable people disagree. The successful version, the one that gets a million users, probably can’t afford to keep saying that, because users want a verdict. They want to be told who’s lying and who isn’t. They want the green light or the red light, not the nuanced essay.

That tension — between honesty and influence — gets harder to hold the more influential the tool becomes. The serious response is something like: publish your rubric, publish your disagreements, surface dissents prominently, refuse to give a single number where multiple are honest. Whether the market rewards that level of restraint is an open question. I’d not bet the family farm on it.

The moment to watch most carefully is the first major debate where some superPAC takes the asymmetry numbers and runs them as the entire content of an attack ad. “Here is the score. Vote accordingly.” That is the moment this technology stops being a literacy tool and becomes campaign infrastructure. And the politics of building it — funding it, governing it, contesting it — change overnight.

We are, dear readers, perhaps two election cycles from finding out which version we get.

That’s why I’ve partnered with Parallax — a company that wants to create the most ethical and useful version of this technology. You can download the Parallax extension for your Google Chrome browser and see the techniques YouTubers and writers use to persuade you. And in the spirit of intellectual honesty, I scanned the article you just read — you can view Parallax’s analysis of it here.