When people talk about AI and political communication, the conversation usually centers on speed: more content, faster turnaround, greater reach. A more interesting question—and one for which reliable data is now available—is this: How effectively can language models actually shift political attitudes?
The answer is more nuanced than the headlines suggest. And for Public Affairs, it leads to a different place than expected: away from the question of how to use AI to persuade more effectively, and toward the question of how to use AI to evaluate one’s own position before it reaches the ministry’s desk.
What the studies measured
A study by Bai et al., published in *Nature Communications* in 2025, which included three pre-registered experiments and 4,829 participants, found measurable shifts in attitudes resulting from texts generated by large language models—across a range of controversial policy issues. On average, the AI-generated texts were just as persuasive as those written by humans (LLM-generated messages can persuade humans on policy issues).
A PNAS study from the same year tempers the hype surrounding hyper-personalization. Neither tailored nor interactive messages performed significantly better than a well-crafted generic message (Testing theories of political persuasion using AI).
The most insightful study is the large-scale Science study by Hackenburg et al., which involved 76,977 participants, 19 models, and 707 political topics. It shows that prompting and targeted post-training boost persuasiveness significantly more than personalization or model size. One factor in particular proved effective: dense, fact-based reasoning. However, the study also checked 466,769 statements for accuracy—and found that the models that argued most persuasively were systematically less accurate (The levers of political persuasion with conversational AI).
Taken together, this paints a picture of public affairs that is more sobering than threatening. What works is essentially what should work in good advocacy anyway: a well-documented argument, clearly structured, without exaggeration. And precisely when the pressure to persuade increases, staying true to the facts becomes a problem.
The boundary you have to read between the lines
All three studies primarily examine citizens in controlled or digital experimental settings. It cannot be concluded from this that professional decision-makers would react in the same way.
A staff member at the ministry does not read a position paper as a private individual. She reads it against the backdrop of the ministry’s priorities, a government agenda, ongoing review processes, and the question of who else contributed to it. She also assesses the sender, the underlying interests, and the feasibility of implementing the proposal within the available timeframe. Expertise, institutional constraints, and organized opposing interests act as filters here—factors that simply do not exist in the experiment.
Its application to public affairs is therefore a hypothesis, not a guarantee. That is precisely what makes it useful in practice: AI can improve arguments, but it does not replace the political context.
The actual use case: putting pressure on your own position
Anyone who concludes from these studies that we should now be rolling out AI messages on a large scale has missed the point. The benefit lies before the deadline, not after it.
One person can take on the role that no one on the team usually has time for: that of the harsh reviewer.
Which assumption falls apart if the other side uses a different data set? Which figure in our paper is more than three years old? How does our proposal sound from the perspective of a social partner who will reject it, and from the perspective of a civil servant who would have to implement it? What happens to our line of reasoning if the responsible department changes?
This forces you not only to articulate your own position, but also to test its resilience. And it’s different from having someone else write a text for you. It makes the preparation harder, not easier.
Comparing alternatives falls into the same category: Which argument holds up in reality? Where is there a risk of reactance? Which phrasing is politically acceptable without watering down one’s own position? Here, too, we’re still dealing with hypotheses. Knowledge of stakeholders doesn’t replace this; it simply allows us to make better use of it.
Where the border runs
In this field, integrity is no mere footnote. Public affairs can only work as long as it is clear to the other side of the table who is writing and on whose behalf. Fabricated sources, artificially generated grassroots voices, or a concealed sender undermine the very resource on which political relevance depends in the long term. The gain in efficiency is always the smaller factor in this trade-off.
Just how seriously this is being taken is already evident in Austria. As of August 2, 2026, ORF has expanded its labeling of AI-generated and AI-edited content, explicitly citing credibility, public trust, and the transparency requirements of the EU AI Act (ORF press release dated July 31, 2026). For Public Affairs, it is not so much the rule itself that is interesting as the logic behind it: As soon as it becomes unclear where a statement comes from and how it came about, a speed advantage turns into a credibility problem.
Parliament is also addressing this issue. The report, prepared by the Institute for Technology Assessment at the Austrian Academy of Sciences, analyzes, among other things, disinformation, fake accounts, political microtargeting, and the role of algorithmic selection (“Generative AI and Democracy,” final report in PDF format). Anyone who engages in unethical practices in this context must expect to be caught.
The real measure: better representation of interests, better laws
But that still doesn't answer the crucial question: What's the point of all this?
Advocacy plays a role in the legislative process. Ministries and lawmakers make decisions on matters whose practical implications they cannot know firsthand. They rely on others to point out the real-world consequences of a regulation: which deadlines are technically unfeasible, which definitions disproportionately affect certain industries, and which reporting requirements lack a corresponding data set. This is precisely why there are discussions with political stakeholders, review procedures, committee hearings, social partnership, and so on.
The goal of AI should therefore not be to make one’s own position more assertive. It should be to make one’s own contribution more useful. And there is a real lever for achieving this: impact assessments can be calculated more accurately, interactions with existing law can be identified earlier, data sets can be processed more quickly, and unintended side effects can be identified before they become apparent during implementation. A position paper that not only sets forth one’s own demands but also provides a robust analysis of the consequences of implementation—including alternatives—is simply more valuable to a government department.
In the best-case scenario, this improves not only the quality of advocacy but also the quality of the legislation itself. Fewer regulations that need to be tightened shortly after they take effect. Fewer transition periods that are out of touch with reality. Fewer provisions that don’t work in practice. Those who understand public affairs in this way have a standard for AI deployment that extends beyond their own success—and which, incidentally, is also the best protection against suspicions of manipulation.
Five Steps to Implementation
The value doesn't come from purchasing a tool, but from following a defined process. In practice, five steps are sufficient:
- Define the goal—including the question of what intermediate political outcome counts as progress.
- Provide context—primary sources, data, and your own position, all in a verified form.
- Have someone proofread it—to identify counterarguments, different perspectives, and alternative phrasing.
- Verify—check every fact against the primary source and ensure political compatibility within the team.
- Taking responsibility—the final message rests with the person who must also defend it in conversation.
Step four is not merely a formality. The findings in Science regarding declining factual accuracy alongside increasing persuasiveness are precisely why no AI-generated set of facts should be included in a position paper without being verified. This also includes documented prompts, clear approvals, and internal rules for confidential data. AI governance here is not a compliance add-on, but rather an integral part of communication quality.
How to Recognize Progress
The number of texts produced using AI is not an indicator of success. Questions like these are more useful:
- Are objections identified earlier?
- Does the number of unresolved issues decrease before a meeting?
- Are Q&As becoming more consistent across departments?
- Can management decide more quickly between different courses of action?
- And now, the toughest test: Are our opinions actually being used in the department?
If only output increases and these factors remain unchanged, the productivity gain is smaller than it appears.
Conclusion
The research shows that AI makes argumentation scalable. It does not show that it makes it better. The benefit arises where technology enhances the quality of human work: better preparation, more perspectives, earlier identification of weaknesses, more consistent messages—and, overall, a contribution that truly helps lawmakers. Those who, instead, merely produce more content are primarily automating mediocrity.
Do you have any comments on the text, or do you disagree completely? Feel free to email me at: bernhard.zink-spiel@mastermind.cc
Politische Entwicklungen früh erkennen und strategisch einordnen: Mit dem Mastermind Monthly erhalten Sie einmal im Monat kompakte Analysen zu relevanten politischen und regulatorischen Themen – mit Hintergründen, Einschätzungen und konkreten Implikationen für Unternehmen und Organisationen. Jetzt kostenlos abonnieren und keine Ausgabe verpassen.
