Persuasion just became a commodity. Influence did not

01 Editorial Content

By Jacopo Garone

PERSUASION JUST BECAME A COMMODITY. INFLUENCE DID NOT

In June, researchers from Oxford, the UK AI Security Institute, the LSE and Stanford published a result that our profession has been half-expecting for two years. Across 18,978 conversations with 6,923 people, AI systems were reliably more persuasive than the best human persuaders the researchers could find, from tournament winners and professional canvassers to world championship debaters. In a real-world test, AI was nearly three times as effective as professional fundraisers at raising money for Save the Children. If a machine can win the argument, what exactly are firms like Voice selling?

The easy answer is relationships. Relationships matter, and always will, but on their own they are an incomplete answer. I think the research says something more precise, and more useful, about where the value in influence is moving.

What the machines are actually good at

The detail that matters is not that AI won; it is how it won. The authors found converging evidence that its advantage came from "rapidly deploying larger quantities of information". When the AI was forced to write human-length messages at human speed, its lead over the best coached debaters fell to zero.

The humans tried to compete on the same ground. After practising against the AI that had beaten them, elite debaters wrote longer messages and used 54% more fact-checkable claims per conversation. Their persuasiveness did not significantly improve. Volume is a contest humans cannot win, and the study shows it.

This fits an earlier and larger result. In December 2025, a team led by the same first author published in Science the findings of nearly 77,000 conversations with 19 language models. The most persuasive systems were those that produced information-dense arguments, packed with claims that can be checked. Read this way, persuasion is less a talent than a throughput problem, and throughput is what machines are for.

The price of volume

Two further findings complicate the picture, and both matter to anyone whose reputation is exposed.

First, the same Science study found a trade-off: the more persuasive a model is, the less accurate its information tends to be. Optimising for conviction erodes the very thing that makes a claim safe to repeat.

Second, in September, Reed Orchinik of Carnegie Mellon and David Rand of Cornell tested a single warning on 3,208 Americans: that language models can be prompted to persuade or manipulate, and may present information selectively. It cut the persuasive effect of AI by roughly half (48%) without denting general trust in the technology. One line of disclosure, and half the effect is gone.

Read together, these results suggest that the argument was never the whole of persuasion. Who is speaking and whether they can be trusted to mean it carry a weight that no volume of information can replace.

The industry has already noticed

The commercial response is already visible. In August, Burson, part of WPP, acquired Limbik, whose platform predicts "the virality and believability of messaging to determine impact before it reaches the market" across more than 60 markets. It will run across the WPP network. Large groups are turning persuasion into infrastructure: tested, scored and deployed at scale.

That is a rational move, and it will work for mass audiences. But it confirms the point. If persuasion can be pre-tested, generated and optimised by software, it will be priced like software. Anyone whose offer is "we build the strongest arguments" now competes with a machine that builds more of them, faster, for less.

The wrong reflex

AI has already transformed advocacy and influence, above all in digital spaces. The real question is how institutions respond, and the early signs worry me.

When the EU tightened its rules on political advertising last autumn, the largest platforms did not adapt. Meta and Google stopped carrying political advertising in the EU altogether, citing legal uncertainty. Bad actors lost a channel. So did every legitimate organisation that wanted to make its case openly, with its name on it. Researchers tracking the aftermath have already seen political influence shift into less transparent channels. A rule written for transparency ended up shrinking the most transparent form of paid advocacy there is.

The Orchinik and Rand study points to a better answer. One line telling people that a machine may be trying to persuade them halved its effect, and left their trust in the technology intact. That is not a case for silencing the messenger. It is a case for educating the audience. Citizens who understand how persuasion works, and who is behind a message, are harder to manipulate by anyone, human or machine. Restricting legitimate professionals does the opposite: it leaves the field to those who never intended to sign their work.

What remains scarce

So what does not become a commodity? I see three things, and none of them is sentimental.

01. Vision: choosing the argument, not just winning it

In these studies the goal was fixed in advance: move this person, on this issue, now. The AI was very good at that. Nobody asked it which argument was worth having, how to frame it, when to raise it, or whether to raise it at all. That is strategic vision, and the creativity that goes with it: seeing the fight before it starts and choosing which one to pick. It also includes knowing what not to say. A system optimised to convince will trade truth for effect unless someone decides it should not.

02. Access: the relationship and the room

The studies measured members of the public, in text, in paid online sessions, and the largest of them is still a preprint awaiting peer review. The authors say so themselves. That is not where the decisions that matter in regulated sectors are made. Those happen in small rooms, with small, informed and sceptical audiences: officials, regulators, investors, journalists who have read the file and know the people making the case. What opens those rooms is a relationship built over years; what keeps them open is competence. In those rooms, information is the least scarce thing on the table.

03. Attribution: a name that answers

A message is worth what the name behind it is worth. When people know a machine may be arguing a brief, they discount it by half. A named person or organisation that will answer for a claim, next month and next year, is the one thing an AI cannot supply on anyone's behalf.

Persuasion is winning an argument. Influence is being the party whose word still counts after the argument is over. The first has just become a commodity. The second has not.

What changes if you operate in a regulated sector

Three practical consequences follow.

First, separate the two. Much of what organisations call communication is argument production: briefings, Q&As, position papers, social copy. Much of it can and will be automated. The part that cannot is the mandate: who decides the strategy, who signs, who speaks, who answers. That is where the investment should go.

Second, treat accuracy as capital, not compliance. If the most persuasive systems are the least accurate, the organisations that keep their claims checkable will be the ones still believed when a claim is challenged, whether by a journalist, a regulator or an activist.

Third, disclose before you are asked. Audiences discount what they suspect is machine-made. Saying clearly where AI is used, and where it is not, turns a vulnerability into a signal of confidence.

At Voice, we use AI every day to read more, and faster, than any team could. We do not let it decide what a client should say, or to whom. That is where our work begins.

On whose behalf

The authors close their paper by noting that the question is no longer whether AI can out-persuade humans, but how, where and on whose behalf this capability will be exercised. For those of us who advise organisations on how to be heard, that last clause is the job.

So here is the question worth asking this month: if a machine can make your case better than you can, will anyone still believe it is yours?

Sources and further reading

Hackenburg, K. et al., AI systems out-persuade expert humans, arXiv preprint 2606.16475, June 2026

Hackenburg, K. et al., The levers of political persuasion with conversational artificial intelligence, Science, December 2025

Orchinik, R. and Rand, D., A light-touch AI literacy intervention helps protect against AI political persuasion, arXiv preprint 2609.16432, September 2026

Votta, F., What data reveals about Meta and Google's political ad ban in the EU, Tech Policy Press, 3 November 2025

Burson, Burson acquires cognitive AI company Limbik, 12 August 2026

SHAPING THE CONVERSATION

Voice has a new identity and a new website. Both start from one distinction: adding to the noise is not the same as shaping the conversation. The organisations that handle complexity best know exactly what to say, when, and how to reach their audience. Our positioning and four core capabilities are now built around that idea. Be the voice, not an echo.

BIOREFINERY: A STORY TO BE TRUSTED

Our team spent a day at Xanthos, a biorefinery in the Adamant Group, developing content and shaping its communication strategy for the months ahead. The energy transition is as much a communication challenge as an industrial one: technical detail has to become something credible and clear to stakeholders with very different levels of knowledge. That translation is the work we do best.

700 ALUMNI IN LONDON. NOW THEY HAVE A CHAPTER

Our co-founder and Co-CEO Jacopo Garone has been appointed Head of the London Chapter of IULM Alumni, after two years on the association's Steering Committee. More than 700 Università IULM alumni work in London across finance, fashion, media, consulting and tech: a community that already exists but does not yet know itself. The first gathering is on 4 November in London.

If AI is changing how influence works, the real question is who is shaping the rules of visibility - and who is reacting too late. We will continue exploring this throughout the year.

Enjoyed this edition? Forward it to a colleague or subscribe to receive future insights from Voice.