- Generative AI can produce inaccurate or biased content that appears credible, which can affect government decisions, services, and communications, and hurt public trust.
- In the public sector, verification, transparency, the identification of biases, and human judgment are essential practices for the effective adoption of AI.
- These issues are explored in a recent IDB publication to help governments develop sound AI policies.
A response may sound convincing and be well-written, yet still contain errors.
That is one of the most significant challenges that generative artificial intelligence (AI) poses for governments and public organizations.
A government official prepares a report. A communications specialist drafts a proposal for an institutional campaign. A citizen looks up information about a government procedure. Increasingly, behind these everyday tasks are tools capable of producing text, images, and summaries in a matter of seconds.
This capability offers significant opportunities to improve productivity, streamline processes, and enhance the services provided to citizens, an agenda that the Inter-American Development Bank (IDB) actively supports. To capitalize on these opportunities, several governments across Latin America and the Caribbean are increasingly incorporating generative AI tools to strengthen public management and support decision-making.
However, these technologies also pose new challenges. Unlike other digital tools, generative AI can invent data, omit relevant information, or reproduce biases while presenting results that appear plausible and reliable.
The challenge is no longer simply to harness the benefits of generative AI, but to ensure that its results are reliable and subject to the necessary oversight.
When a Convincing Response Contains an Error
One of the most unique challenges of generative AI is that its errors often appear credible. Responses may be well-written, sound convincing, and even mimic the format of reports, technical documents, or institutional communications. As a result, detecting an erroneous statement can be more difficult than identifying a traditional technological flaw.
A widely publicized case occurred in 2023 in the United States, when lawyers used an AI tool to assist in drafting a legal brief. The document included references to court decisions that had never existed. The citations appeared authentic and followed the expected format, but they had been generated by the system without any basis in reality.
The limitations of these tools are not confined to isolated cases. Recent research has found that generative models can produce nonexistent bibliographic references, omit relevant information, or present false claims with high levels of confidence. These situations are particularly relevant for governments and entities that produce knowledge, draft technical documents, or use evidence to inform public decisions.
The challenges are not limited to written content either. Generative tools can produce seemingly realistic images that contain errors, omissions, or stereotypical representations of certain social groups. UNESCO has documented how some systems can reproduce biases present in the data used to train them, including gender stereotypes and other forms of discrimination.
These limitations do not negate the value of the technology, but they underscore the importance of critically reviewing its results before incorporating them into governance processes.
Citizens interact with institutions, not with algorithms.
Why These Errors Matter to Public Institutions
When a response contains errors, a communication conveys inaccurate information, or a decision appears unfair, institutional credibility can be undermined regardless of the technology used to generate it.
Decisions remain the responsibility of organizations and the individuals who make them. Therefore, the use of these tools is not merely a technological issue; it also influences the quality of public services and the trust that citizens place in them.
Four Best Practices for the Reliable Use of Generative AI in the Public Sector
As these tools are integrated into the day-to-day work of governments, there are certain best practices that can help improve the quality and reliability of the results.
1. Verify the information. Generating content is becoming easier, but validating data, cross-checking sources, and reviewing evidence remain essential.
When a response seems convincing, verification is no longer an optional step.
2. Identify potential biases. AI systems learn from large volumes of existing information. Critically reviewing results involves asking who might be missing from the response, what perspectives were not considered, and whether the content reproduces stereotypes or generalizations that could affect a decision or public communication.
3. Act with transparency. People need to know when an organization uses AI tools to support processes or services. Transparency also involves explaining how results are monitored and who bears ultimate responsibility for the decisions made.
The OECD’s AI Principles highlight the importance of people understanding when they are interacting with AI systems and having adequate oversight and accountability mechanisms in place.
4. Apply human and contextual judgment. A response may be technically correct and yet still prove insufficient for public policy, institutional communication, or interaction with the public.
Human judgment remains indispensable for interpreting contexts, assessing risks, and making decisions.
AI Also Requires Public Sector Capabilities
Generative artificial intelligence offers significant opportunities to improve public administration. However, its impact will depend less on the ability to generate content and more on the ability of governments to use these tools judiciously, transparently, and with adequate oversight.
Harnessing the potential of AI in the public sector requires more than just access to technology. It also demands the capacity to verify information, identify risks, monitor results, and ensure that decisions continue to serve the public interest.
In this vein, the Reference Model for AI Policies in the Public Sector in Latin America and the Caribbean (available in Spanish), developed by the IDB, highlights the importance of complementing technology adoption with mechanisms for governance, risk management, and capacity building.
These capabilities are not built automatically. They require continuous learning, the exchange of experiences, and access to practical tools that enable public servants to understand both the potential and the limitations of these technologies.
Trust does not arise from technology alone; it is built based on how institutions use it.
From Knowledge to Action
At the Inter-American Development Bank, we support governments in Latin America and the Caribbean in the responsible adoption of artificial intelligence through knowledge development, practical tools, and capacity building.
That is why the IDB is promoting ImplementaLAC, the pioneering regional hub where public servants in Latin America and the Caribbean turn their ideas into policies that deliver results. This IDB platform offers free access to courses, tools, and resources on policy implementation, digital government, artificial intelligence, and regulatory improvement for the public sector.
If you want to build your capacity to use generative AI in a critical, responsible, and public-interest-oriented manner, join ImplementaLAC.