Artificial Intelligence Needs Quality Assurance

What are architects' hopes for artificial intelligence? For our 1/2.2025 issue, we asked around 20 experts. Andreas Dieckmann from GMP Architekten warns against using AI blindly – in the end, decisions still have to be made by humans.

For some design teams, text-to-image generators have become an exciting additional tool in the early design phase. They can be used to convert hand-drawn sketches into renderings, to enhance existing simple renderings and, of course, to create image variants of interiors (material sampling) or facades (cladding). Here we use an on-premise solution, i.e. an AI installed and isolated in our own network, which can be trained with our own data (such as project photos and renderings) without connecting to a foreign server. The successful use of such tools requires in-depth training of staff to instruct the text-to-image generators, but the results are useful.

We also use software with AI components in our preliminary design for applications such as optioneering (i.e. the rapid development of numerous design variants), volume studies and environmental analysis. The advantage of such integrated solutions is that by embedding them in the software user interface, these tools are easier to use (optioneering) or even work automatically in the background (analysis).

Another area where we use AI is in our own software development. The results here are rather mixed and must always be treated with scepticism. Overall, our perception in this area is that the LLMs (Large Language Models) do a better job of explaining software/programming concepts than they do of generating directly usable code. We are also testing software with AI components in other areas, such as construction site monitoring or the creation or revision of texts. However, it is too early to talk about our experiences here.

The biggest risk in using AI is the naivety of the users. AI products are currently mostly cloud-based. However, depending on the application, it may not be desirable from the user's perspective for proprietary data to be processed externally or even used to train AI outside the company. This may be sensitive business data (such as contract documents) or intellectual property (architectural photos, renderings, plans). In the interests of privacy, data security and intellectual property protection, it is important to carefully check the software licence terms before using AI products and, if necessary, to refrain from using a product.

The quality of AI output should not be overestimated. In most cases, it does not match the knowledge of experienced experts in terms of reliability and accuracy. In terms of reliability, there are numerous examples of AI systems hallucinating, i.e. producing results that do not stand up to factual scrutiny. This means that all AI work should be subject to additional quality assurance processes: They need to be critically and vigilantly reviewed and evaluated by experienced planners. Thus, AI is (at least in our current experience) more suitable for work that would require internal review and evaluation anyway. For everything else, we still consider (human-programmed) rule-based automation to be a much more reliable tool.

Probably less than the current hype suggests. We are starting to reach the peak of the hype cycle, so hopefully expectations of the future use of AI will be a little more realistic. Also, our industry is not exactly known for its innovative spirit, if you look at how long it took for BIM to become established as a design method.

Larger design and construction firms are likely to use on-premise AI models for their own specific applications, such as generative AI (i.e. AI that generates text, images or even building models) and predictive AI (AI that draws conclusions from historical data based on statistical models). However, the majority of design firms will come into contact with AI through the use of software products. On the one hand, these will be new products that are essentially AI software (e.g. Finch). On the other hand, we will also see the selective integration of AI into existing software – where it makes sense (Autodesk Forma is already a good example of this, using AI to speed up – preliminary – analysis results). At least in the short term, the biggest beneficiaries of AI could be those who either do not need a high degree of precision and planning depth (e.g. in project development) or are highly specialised (e.g. planning offices with a clear focus on housing projects). 

We would like to see intelligent assistance systems integrated into software products. These systems would understand the software architecture of their host system and could be operated using natural language. They would be able to reliably answer tool- and workflow-related questions, but their language models would also need to be trained with company-specific standards and workflows. Ideally, they would also understand the architecture of their host system so well that they could be used to reliably automate simple tasks according to clear instructions. It is particularly important for generative AI that it can be easily trained with its own data in its own network, as is already possible with text-to-image generators. On the one hand, this would enable a wider use of such tools. On the other hand, it could also lead to an increase in quality, since it would be possible to use one's own project pool for training without the risk of producing results that are too generic.


All interviews on the topic of AI can be found in our white paper on Architecture + Artificial Intelligence.

Read more in Detail 1/2.2025 and in our databank Detail Inspiration.