Conversational AI: measuring influence while protecting public judgment

The finding

Salvi and colleagues conducted a preregistered experiment with 900 participants in short debates with humans or GPT-4. Their results provide evidence that AI conversations can influence expressed agreement in a controlled setting. The outcome was measured attitude change, not observed voting behaviour or an election result. [1]

Evidence, correction and context

A September 2026 correction changed an important comparison: personalized GPT-4 did not significantly outperform unpersonalized GPT-4 in that direct test. The corrected analysis therefore does not establish an additional personalization benefit from this comparison alone. Any account of the original paper should include that qualification. [2]

Lin and colleagues separately tested AI dialogues in three national election contexts and reported changes in candidate preferences. These experiments expand the settings studied, while still leaving a distinction between survey responses and actual electoral behaviour. They do not measure a Latin American population. [3]

Oxford's official account of related Science research reports a tension between persuasive performance and factual accuracy. That finding supports scrutiny of both outcomes rather than using influence alone as a quality measure. The university also notes the gap between controlled conversations and real-world engagement. [4]

Editorial perspective: assess the public purpose

The following agenda is independent editorial analysis. A public institution considering a conversational system should first state what service it wants to provide and how success will be assessed. For a civic-information service, useful criteria might include accurate explanations, visible sources, correction of errors, and the user's ability to question an answer. Changing a user's opinion should not silently substitute for those goals. An evaluation should examine whether people understand the system's role and can reach a human when needed. The proposed principle is straightforward: institutions should judge a tool against its declared public purpose and publish enough evidence for outsiders to challenge that judgment.

Editorial scrutiny: evaluate understanding and accuracy

A proposed audit should separate factual accuracy, comprehension, user satisfaction, and willingness to act. Those measures should be defined before testing and reviewed independently. Researchers should document the exact system version and record whether answers cite sources that actually support their claims. A correction procedure should explain how an identified error reaches the people responsible for the service. These are proposed governance tests, not evidence that any particular institution has failed them. An audit should also state what its sample cannot establish. A small successful pilot would justify further evaluation, not an unrestricted claim that the system is reliable for every language, topic, or public decision.

Editorial questions for Latin America

A regional project could ask whether a nonpartisan information tool answers equivalent Spanish and Portuguese questions with comparable accuracy and understandable sourcing. How should evaluations include Indigenous-language expertise without assuming that translation alone establishes equivalence? Can users identify uncertainty and distinguish an official rule from an explanation or opinion? Social scientists, linguists, accessibility specialists, computer scientists, journalists, and public-service users should help design the assessment. Local validation would require language-specific review and realistic service questions. The objective should be public understanding and accountability, not discovering which personal characteristics make a group easier to persuade. Results should describe the tested setting and preserve the privacy of participants.

Editorial research pathway

A first output could be a public evaluation protocol, a small reviewed question set, and a transparent correction log. Independent teams could assess the same answers and publish disagreements with their reasons. Subsequent testing should examine whether users learn accurate information and recognize its limits. The work would remain a proposed research programme until institutions, participants, and appropriate reviewers agree to undertake it. The studies cited here are external research; this briefing does not claim that LAANSSS tested a political system or operates a persuasion programme.