The AI changed my mind. Should it?
Training AI to fairly reflect diverse views might help us navigate conflict better than seeking AI 'neutrality'
Would you want an AI to change your mind? Or to help you understand how your political adversaries see the world? Next week’s Paris AI Action Summit is set to address these very issues, including ensuring AI aligns with humanist values for the collective good as AI global adoption accelerates.
Promoting Meta AI’s 2025 ambitions this month, Mark Zuckerberg stated, "I expect that this is going to be the year when a highly intelligent and personalized AI assistant reaches more than one billion people.” With personalized AI at scale, how will it shape our beliefs, political opinions, and sense of reality? Will the AI give us facts and be politically neutral? This is the substance of “A Practical Definition of Political Neutrality for AI,” by Senior Scientist at the Center for Human-Compatible Artificial Intelligence, Jonathan Stray.
The neutrality myth
In theory, AI could be trained to be ‘neutral’ and have no impact on human beliefs. But in practice, Stray argues, “such a machine would lie to try to prevent you from changing your mind." An AI trained to have zero persuasive effect would have to actively resist influencing a user, possibly engaging in omissions, distortions, or even outright falsehoods to maintain its non-influence.
So if LLMs do influence users’ beliefs, who decides what narratives are amplified and which perspectives are given to the users? Stray points to studies showing that large language models (LLMs) often exhibit political leanings. A recent study found that engaging with LLMs subtly shifted users’ voting intentions toward Biden over Trump, even though the LLM was not explicitly designed to influence their choice.
Determining maximum equal approval
The quest for political neutrality led Stray to propose an empirical solution: maximum equal approval, where an AI-generated response is deemed neutral if it fairly represents each side of a debate and ensures an equal percentage of approval from both sides. This concept mirrors Wikipedia’s "neutral point of view" policy, which prioritizes comprehensive and balanced representation over the pursuit of an absolute "truth."
Stray argues that AI responses should be evaluated by how fairly different communities perceive them, rather than through traditional objectivity metrics. However, he highlights that this approach carries its own risks:
Who decides which perspectives are legitimate and which are 'fringe'?"
Bad actors could game AI fairness metrics, amplifying extremist views or distorting factual accuracy.
Those with greater access to AI developers and training data could disproportionately influence the system’s 'neutral' defaults.
These challenges highlight the need for continuous human oversight, transparency in AI decision-making, and mechanisms for users to challenge AI-generated outputs.
Asking AI about conflict
Stray’s inquiry becomes even more critical when AI engages with users on violent conflicts and historical injustices. When the question is whether genocide is underway in Gaza, should AI simply present all viewpoints as equally valid? When asked which country Nagorno-Karabakh belongs to, should AI offer different answers depending on the language the question is asked in?
Stray experimented with LLaMA-generated responses to the question ‘Is the war in Gaza a genocide?’' to illustrate three possible approaches:
Definitive Answer: AI takes a firm stance based on legal definitions and existing evidence.
Ambiguous Answer: AI highlights the complexity of the issue without taking a position.
Multi-Perspective Answer: AI presents arguments from both sides in a structured manner.
Under the maximum equal approval model, Stray argues that the third approach is preferred—not because it satisfies everyone, but because it maximizes fairness across competing narratives.
This doesn’t solve the matter of persuasion. There’s a tendency to equate persuasion with manipulation even though influence exists on a spectrum. If AI presents a bouquet of facts and perspectives, and a user changes their mind as a result, is that manipulation—or is it intellectual growth? Stray notes: "If there really are better facts and arguments on one side, they should persuade; a news bot that could never change your views would be dysfunctional."
AI as multi-partial
Stray’s arguments suggest that AI could be trained to function more like a conflict mediator than an information authority, ensuring that users are exposed to multiple perspectives, not just dominant or algorithmically favored ones.
These dilemmas reveal a deeper truth: neutrality is a myth. Peacebuilders have long recognized that in high-conflict settings, neutrality or impartiality is rarely achievable. Instead, a concept called multi-partiality—where all sides are engaged fairly but not necessarily treated as morally equivalent—may be a better guiding principle for AI design.
For those at the intersection of technology and peacebuilding, the goal isn’t a perfectly neutral AI, but one that reduces harm, polarization, and fosters constructive engagement with complex realities. AI systems will inevitably influence human perception—but we must decide whether that influence will fuel conflict or facilitate understanding.
If we get this right, AI could serve as a tool for social cohesion, helping societies navigate contentious issues with greater nuance and empathy. If we get it wrong, we risk deepening divides, entrenching fear and dangerous misinformation, and allowing invisible algorithms to shape our world in ways we neither control nor fully understand.
Lena Slachmuijlder is a Senior Advisor for Digital Peacebuilding at Search for Common Ground, and Co-Chairs the Council on Tech and Social Cohesion