Trust in AI Models

The question constantly arises of how we select LLM providers, which criteria guide our decisions, and whether—beyond costs, runtimes, and other technical factors—there are additional criteria. We group these considerations under the term trust.

What this is about:

This discussion is less about the end user or private customer, and more about how an LLM behaves within an integrated system—for example, as an agentic AI solution such as a file-handling assistant.

Admittedly, it depends heavily on what is needed, since the use case determines whether an LLM should be more talkative or the opposite—rather “laconic.” This must be factored into the evaluation as a criterion: talkativeness can be beneficial in a generative AI context, but less so for a document assistant in the insurance sector.

In advance:

Trust in AI is not a technical problem. It is an issue of intentions, governance, and the cultural infrastructure from which a model emerges.

The outcome:

The result is a classification of currently available market LLMs, categorized by trust level: high, medium, or low.

Here, apart from cost and IT-related metrics, we evaluate whether and how we trust the outputs—i.e., qualitatively. Regarding the trust criterion, we tend to exclude LLMs with lower trust levels if they show strong ideological bias. For our use cases, we prefer LLMs that act pragmatically and whose underlying intentions we can understand.

The Trust Report on current LLMs can be requested directly from sol4data at info@sol4data.com. Please book an appointment with an AI architect for this purpose.

Summary:

Models differ less by parameters and more by the value system of the organization that builds them. Their generated responses reflect this.

We provide guidance on selecting LLMs—even when the considerations go far beyond token pricing.