Workforce Guidance Addresses Risks of AI Bias in Decision-Making

A new practical guide from United Ceres Workforce Learning outlines how organisations can approach responsible artificial intelligence use at work. Released as part of the site's broader library on workplace technology, the guide focuses on four core areas: deciding whether a task is suitable for an AI tool, protecting personal and confidential information, checking generated content for bias, and naming who is accountable for the result.

The guide draws on official guidance from two Singaporean authorities. It references the Personal Data Protection Commission's Advisory Guidelines on Use of Personal Data in Generative AI, issued on 20 July 2026, and the Infocomm Media Development Authority's Starter Kit for Testing LLM-Based Applications. The page notes that the PDPC guidelines are advisory only, not legally binding, and do not constitute legal advice. It also states that the Personal Data Protection Act provides the baseline standard of protection for personal data in Singapore.

How the Guidance Works

The framework is structured around three stakeholder groups identified by the PDPC: model providers, who develop the underlying AI models; system providers, who build applications on top of those models; and system deployers, who put the applications into use within an organisation. Each group has distinct responsibilities, but the guide emphasises that all three must take part in ensuring that AI outputs are reliable and fair.

A central recommendation is to treat every AI output as a draft. The guide advises setting the level of human checking before using the tool, meaning that organisations should decide in advance how much human review is needed for each type of task. This approach acknowledges that generative AI systems can produce plausible-sounding but incorrect or biased results, and that human oversight remains essential.

The Five Key Risks

The IMDA's Starter Kit identifies five key risks in testing large language model applications, and the guide highlights one of them specifically: bias in decision making. This risk is especially relevant for workplace tools that assist with hiring, performance evaluation, customer service, or resource allocation. When an AI system reflects biases present in its training data, it can lead to unfair outcomes that affect individuals' careers or access to services.

The guide does not prescribe specific technical fixes for bias. Instead, it focuses on procedural steps: checking generated content for bias, naming a person or team accountable for the result, and documenting the checks performed. This shifts the emphasis from trying to build a perfectly unbiased system to creating a process that catches problems before they cause harm.

One of the key points the guidance makes about AI bias at work is that it is not solely a technical issue. The guide argues that organisational culture, training, and accountability structures matter as much as the algorithms themselves. By requiring that someone be named as accountable for each AI output, the framework encourages a culture of responsibility rather than blind reliance on automation.

Practical Implications for Organisations

For organisations using or considering generative AI tools, the guide offers a structured way to evaluate whether a given task is appropriate for automation. The first step is to decide whether the task suits an AI tool at all. Some tasks involve sensitive personal data, high-stakes decisions, or creative work where human judgment is irreplaceable. The guide suggests that organisations assess these factors before deploying any AI system.

Next, the guide addresses data protection. It reminds organisations that the Personal Data Protection Act applies to any personal data fed into or generated by an AI system. This includes data used to train models, prompts entered by users, and outputs that contain personal information. The advisory nature of the PDPC guidelines means that organisations are not legally required to follow them, but doing so helps them align with the baseline standards of the Act.

The guide also touches on the importance of transparency. While it does not require that organisations publish their AI policies, it recommends that they document their processes for task suitability checks, data protection measures, bias checking, and accountability assignments. This documentation can serve as evidence of due diligence if questions arise later.

Why Bias Checking Matters

The focus on bias in decision making reflects a broader concern in the industry. AI systems trained on historical data can perpetuate existing inequalities, such as gender or racial bias in hiring tools, or socioeconomic bias in loan approval systems. The IMDA's inclusion of bias as one of five key risks signals that regulators see it as a priority issue.

The guide does not attempt to define what constitutes acceptable bias. Instead, it gives organisations the tools to identify and address it on their own terms. By checking generated content for bias and naming who is accountable, organisations can at least ensure that someone is paying attention to the problem. This practical approach may be more effective than a one-size-fits-all rule, because the definition of fairness can vary by context.

Another important aspect of the guidance is its emphasis on human oversight. The recommendation to treat every AI output as a draft means that the final decision should always rest with a human. Even when an AI system performs well on routine tasks, the guide advises that organisations maintain the ability to review and override its outputs. This is particularly relevant for high-stakes decisions where errors or bias could have serious consequences.

The guide's approach to AI bias at work is consistent with emerging international best practices. Many regulatory frameworks around the world are moving toward a model that requires human oversight, transparency, and accountability, without mandating specific technical standards. By drawing on Singaporean guidance, the United Ceres Workforce Learning guide provides a localised perspective that may be especially useful for organisations operating in or with Singapore.

For readers who follow developments in responsible AI, the guide offers a clear, actionable framework. It does not promise to eliminate bias entirely, but it provides a realistic path for managing it. The combination of task suitability checks, data protection measures, bias checking, and named accountability creates a closed loop that can be audited and improved over time.

As generative AI tools become more common in workplaces, the need for practical guidance like this is likely to grow. The United Ceres Workforce Learning guide fills a gap by translating high-level regulatory principles into concrete steps that any organisation can implement. Its emphasis on human oversight and accountability may help prevent some of the more visible failures that have occurred when AI systems were deployed without adequate safeguards.

Organisations that follow the guide's recommendations will still need to invest in training and process design. The guide itself does not provide technical tools for bias detection or data protection, but it points to the regulatory frameworks that do. By aligning with the PDPC and IMDA guidance, organisations can build a foundation for responsible AI use that meets current expectations and prepares them for future regulatory developments.