From simulation to strategy: new AI developments that will transform political communication

Archive note: this guest essay was published on 8 January 2025. Its account of the studies and its interpretation of their implications are preserved as written. Quotations from research are translated from the Spanish article, rather than presented as verified wording from the original papers.

In this issue of estrategIA, we welcome a contribution from Martín Sosa Dirié, a journalist, communications consultant and graduate of the Master’s in Image Advice and Political Consultancy directed by Jorge Santiago. As our guest author, Martín examines how advances in AI simulation are transforming political communication.

In November, researchers published a study from the universities of Washington, Northwestern and Stanford, and Google DeepMind, in which they conducted qualitative interviews with 1,052 people and created the same number of AI agents to try to simulate their attitudes and behaviour.

The results were surprising and lend support to using AI to conduct social experiments and, ultimately, try to predict human behaviour.

1. The study¶

The researchers explained that, to assemble the study, they recruited “more than 1,000 participants using stratified sampling to create a representative sample of the United States in terms of age, gender, race, region, educational attainment and political ideology”.

“We present a generative agent architecture that simulates more than 1,000 real individuals using two-hour qualitative interviews,” they stated. They also noted that “large language models (LLMs), which encapsulate extensive knowledge of human behaviour”, make it possible “to build an architecture that can accurately simulate behaviour across multiple contexts”.

What is striking is that the generative agents replicated “participants’ responses to the General Social Survey with 85% accuracy, comparable to the accuracy with which the participants themselves reproduce their responses two weeks later”. They also achieved “similar performance in predicting personality traits and outcomes in experimental replications”.

In addition, across a series of five social science experiments, the researchers compared human responses with those of AI agents. In four of these studies, the AI-generated responses closely resembled the human ones. Statistical analyses revealed a correlation coefficient of 0.98.

2. Some context¶

The study represents another advance in simulation and social experimentation through AI agents, a field in which Project Sid and AI Town had already stood out — and which began almost four decades ago with the legendary Little Computer People on the Commodore 64.

LITTLE COMPUTER PEOPLE | Commodore 64 (1985)

Project Sid succeeded in creating a complex virtual society within Minecraft, populated by 1,000 autonomous AI agents. These agents developed sophisticated social behaviours: they formed alliances, established trading networks and created a currency system using gems to exchange goods and services. Power dynamics even emerged, including instances of “corrupt priests”, reflecting complex social structures.

AI Town grew out of a study published in August 2023 that analysed the interactions of 25 AI agents over two days in Smallville. What made this research interesting was observing how the agents formed — and broke off — relationships, collaborated, learnt through dialogue with one another and resolved conflicts.

3. Predicting human behaviour: the decision-maker’s dream¶

Political communication will not be the only field seeking to make the most of these innovative uses of AI, which bring together programming, mathematics, sociology and political science.

In commerce, marketing and advertising, for example, this technology could mark a turning point: a paradigm shift.

Being able to simulate the emotional and behavioural responses of real consumers would make it possible to test different messages, formats and approaches in a virtual environment before a major market launch.

This would not only reduce the costs associated with trial and error but also increase effectiveness by creating personalised campaigns aligned with consumers’ motivations, maximising impact and loyalty.

In political communication, tools of this kind are undoubtedly a fundamental development that will affect election campaigns in the short term, as well as government management.

Faced with particular situations, politicians and consultants used to commission opinion research — generally online or by telephone, and sometimes face to face. Now, once we have conducted a substantial study using this approach, we will be able to put an endless range of questions and issues to it over a given period.

This could become an indispensable tool for politicians and consultants seeking advance insight into the reception of particular government measures or potential electoral alliances.

Yet, like every innovative technological tool, it also has a downside.

It could encourage a cautious political leadership, excessively preoccupied with the public mood, shying away from the slightest conflict and lacking the resolve or boldness that political action often demands.

4. A call for professionalism¶

AI agents can provide a way of approaching the thoughts — and, above all, the feelings — of particular social groups. But we must not embrace their results blindly.

This is not simply because, like any study, they are not infallible and still have a considerable margin of error. Their findings must never be analysed in isolation.

Growing up in a family steeped in sociology, I have always carried with me the idea that the fullest and most accurate interpretations emerge from dialogue between different studies and findings.

Likewise, the human role in critically analysing information still cannot be delegated.

A scene in the film Irresistible, starring Steve Carell, illustrates this very well.

The film is a satire of contemporary political campaigns. An election is approaching in a small Wisconsin district and, a few days before polling day, the polls predict a tie.

Then — spoiler alert — a member of the Democratic campaign team, immersed in the war room, notices that a very specific part of the district has a high concentration of single women who live alone.

Rather than examine these profiles — based on age, gender and marital status — more closely or compare them with other sources or colleagues’ assessments, he accepts them uncritically and decides to mount a last-minute campaign in that area in favour of contraception.

The result? Complete rejection: it turns out to be a neighbourhood of nuns.

This example simply highlights the importance of building interdisciplinary teams that work thoroughly.

Data analysed without due rigour lead to the wrong strategy and, consequently, poor results.

5. AI and the social sciences¶

AI is already bringing about a paradigm shift in molecular biology, where tools such as AlphaFold — recognised by the Nobel Prize — have predicted protein structures with unprecedented accuracy.

Similarly, in physics, machine-learning algorithms are helping to analyse the immense quantities of data generated by particle accelerator experiments, such as those at CERN, making it possible to identify patterns and discover phenomena that might otherwise have gone unnoticed.

In computational chemistry — the branch of chemistry that uses mathematical models to simulate interactions between the atoms of substances and thereby solve chemical problems — AI models can predict the properties of new materials, accelerating the search for sustainable solutions to global problems such as energy and climate change.

Research such as the work on AI agents also opens up opportunities to intensify collaboration between AI and the social sciences.

In short, AI has already demonstrated its capacity to revolutionise the basic sciences.

Now it is time for it to strengthen the social sciences, where AI can help analyse and understand complex social situations and generate valuable input for decision-making.

Martín Sosa Dirié

Journalist and communications consultant

• @conversacioncontinua — his newsletter on Substack.