“What kind of data do we produce, and how do we use it?”

According to Liang et al. (2024), improving trust and acceptance of AI within organizations in the AECO sector requires a focus on transparency, ethics, and the involvement of internal stakeholders. In the AECO sector, ensuring the accuracy and traceability of AI-generated data is challenging, because there are still few technical mechanisms available for this. Transparency is central in AI development, as users must understand how AI systems function, make decisions, and what data they rely on. For trust to grow, users must have the opportunity to be involved in AI-supported decision-making (Human-in-the-loop) and to evaluate the AI’s suggestions.

This section of the AI playbook for the AECO sector examines the accuracy, traceability, reliability, and transparency of AI-generated data. We review different challenges related to AI-generated outputs, such as bias, and consider how reliability can be improved in AI applications within the sector.

Bias and Fairness

Two of the most significant factors challenging the reliability of current AI systems are bias and fairness. According to Ferrara (2023), biases in generative AI emerge when models produce content that reflects prejudiced or unjust perspectives, often stemming from biases in the data used to train the model. Fairness, in turn, means that AI systems treat all users and situations equitably, without discrimination or favoritism. Bias in AI can lead to unfair treatment of individuals or groups and reinforce existing inequalities.

Explainability and Interpretability of AI-Generated Information

Explainability and interpretability are critical for ensuring trust, transparency, and ethical use when applying AI systems. Explainability refers to an AI system’s ability to provide reasons for its decisions or actions. This means that system should indicate to user how and why the system arrived at a particular outcome. Interpretability refers to how understandable the internal process of an AI model is. The difference between these two can be summarized as follows: interpretability relates to the model’s internal structure and how easily it can be directly understood, while explainability focuses on how well the model can offer meaningful and useful explanations of its outputs, even when the model itself is not inherently easy to understand.

Monitoring and Traceability

Establishing clear monitoring mechanisms for AI-based processes is essential to ensure that all decisions and actions taken by AI systems can be traced. Maintaining logs of AI operations and decisions enables oversight and helps ensure compliance with ethical standards. This transparency supports the identification and correction of potential errors or biases in AI-generated outputs.

Data Transparency

Data transparency means that users of an AI system must be able to easily understand how the AI arrived at its conclusion and what data it used in its decision-making process. Many AI algorithms, especially deep learning models, are opaque, and this “black box” characteristic highlighted by Liang et al. (2024) can weaken user trust and the accountability of the AI system’s owner. Filipsson (2024) suggests emerging methods in explainable AI (XAI) as a solution. These methods aim to enable AI systems to clarify how they reached a given outcome and to verify the data sources used.

Weber-Lewerenz (2021) emphasizes that the reasoning behind AI-driven design decisions in the AECO sector should be transparent and understandable to both designers and clients in order to maintain trust and enable informed adjustments. AI should complement, not replace, human judgment. Recommendations produced by AI should be evaluated critically by experts, especially in AECO sector, before they are adopted, since human safety as well as economic and environmental sustainability are at stake.

An organization in the AECO sector can demonstrate transparency by clearly communicating how the developed AI systems operate, including information about the data they use and their decision-making processes. Documenting decision-making processes through logs and audit trails is also essential to maintain and remember when developing AI systems. The adoption of explainable AI methods can help end users understand how the AI arrives at its results. Highlighting the limitations of AI systems and involving end users in their design process further enhances transparency.

The EU AI Act (2024) emphasizes increasing openness, developer accountability, and transparency. Under the Act, the data sources or datasets used to train general-purpose AI models, such as large private or public databases or archives, must be described in a summary. In addition, the Act stresses the responsibility of developers of high-risk AI systems to ensure transparency. High-risk AI systems must undergo conformity assessments and provide clear, transparent descriptions of how the systems operate.

High-risk AI systems are those that are either integrated into products subject to product-safety regulations or that may pose significant risks to human health, safety, or fundamental rights. Product-safety–regulated items include, for example, machinery that moves people, health technology, and toys. If a product already requires CE marking, any AI integrated into it is considered high-risk. Other high-risk use cases include student admissions, employment decisions, access to essential public services, elections, insurance, the operation of critical infrastructure, and other situations in which AI may make decisions that restrict human rights or endanger safety. In the AECO sector, there is still limited information available on high-risk AI systems, but likely examples include AI-controlled construction machinery and AI-based control of critical infrastructure.

According to Emaminejad et al. (2022), explainability and interpretability in AI have not been widely studied in the AECO sector. However, integrating blockchain technology with AI algorithms can make it easier to track decision-making chains, thereby improving AI explainability. This development, however, requires significant research investment before practical deployment.

Various ethical perspectives

Liang et al. (2024) highlight, in addition to transparency in AI within the sector, several ethical concerns, particularly the reduction of jobs, the importance of data privacy, information security, and conflicts in decision-making. For example, concerns about decision-making conflicts arise from the fact that AI used in design processes may not adequately consider factors that influence sustainability and user comfort, such as natural light availability in floor plans or the responsible use of materials. With current data availability, these aspects are still difficult to measure, and experiential knowledge is challenging to translate into AI algorithm behaviour. In such situations, human and AI-driven decisions may conflict, making it essential to ensure that the reasoning behind AI-assisted decisions is transparent and understandable to both designers and clients. This transparency increases trust and clarifies the conscious design choices made by professionals.

Ensuring the reliability of AI is an ongoing process

Deodhar et al. (2024) emphasize that the ethical principles of AI (just like its technological components) are in constant evolution. Organisations should therefore never consider themselves “finished” when it comes to ensuring the reliability and ethical governance of AI-generated information. In the AI era, ensuring reliability is a continuous process that requires regularly updated practices within companies in the AECO sector.

Strong attention must be paid to AI security and especially to individual data privacy, as the number of cybercriminals continues to grow. From the perspective of cybercrime, the AECO sector is attractive due to the large capital flows involved in projects.

New methods will be needed in the future to ensure reliability, such as managing data biases in algorithms and introducing automated monitoring of AI-generated material. Emaminejad et al. (2022) stress that improving AI reliability requires clearer definitions of human–machine collaboration and designing AI solutions so that they integrate directly into traditional operational processes of the AECO sector.

Read more