Yesterday’s multi-stakeholder roundtable, Understanding LLMs, reinforced a reality that is easy to miss in the hype cycle. LLMs and AI agents can create meaningful public value, but they can also reshape how people interpret political and geopolitical information, often in ways that are hard to detect without strong verification habits and institutional safeguards.
It was also valuable to have AINOW represented in the room, with our president and founder, Suad Seferi, attending the discussion and engaging with the perspectives shared by participants.
In regional work on ethics and bias, a recurring issue is that many systems used in the Balkans are trained primarily on English-language data, which can crowd out local perspectives and context. At the same time, many countries in the region are still building basic strategies and rules for ethical AI, leaving public institutions, media, and education systems without clear operating standards for these tools. This combination increases the risk of unfair outcomes and unaccountable decision-making, especially when AI outputs are treated as authoritative.
What AI agents make possible
AI agents extend LLMs from answering into doing. In practice, this means workflows that can search, summarize, draft, classify, translate, and support case handling or investigative research at speed. In public-interest settings, this can reduce administrative burden, help teams process information faster, and expand access to services or knowledge for people who lack time or expertise.
The opportunity is real, but it depends on a disciplined approach. The systems work best when their role is clear, their outputs are reviewable, and humans remain accountable for final decisions, especially in political, historical, and public communication contexts.
The risk profile that matters most
The main risk is not only that an output is wrong. It is that the output can sound confident while subtly changing framing, tone, and emphasis, which can distort public understanding on sensitive topics. The discussion aligned with a key finding from the referenced research: chatbots should not be treated as stable, independent sources for political and geopolitical issues because their answers can shift in tone and framing.
A second major risk is invisible dependency. When institutions or journalists rely on one model, one vendor, or one interface, small biases and systematic omissions become harder to detect and easier to amplify. This is especially concerning in environments already affected by information disorder and polarization.
A practical governance agenda
The roundtable contribution can be distilled into three core messages.
First, increase transparency from AI developers. Providers should publish meaningful information about data sources and the rules shaping model behavior, and they should document each phase of development to support traceability and accountability. Transparency is essential because LLMs are not fully neutral and can reflect the political and information environments where they were built.
Second, build institutional usage rules fast. Public institutions, schools, and media organizations should implement clear guidelines that state LLM outputs are not facts by default. This is an operational requirement for any organization that wants to prevent AI-generated errors from becoming institutional errors.
Third, require verification and reduce single-model dependence. Users, journalists, and researchers should double-check LLM outputs using trusted sources such as academic literature, official documents, or expert interviews, and they should avoid relying on one model for sensitive topics. Comparing outputs across models can help reveal inconsistencies and bias patterns that are otherwise easy to miss.