and explainability. This keynote provides an overview of cases where knowledge graphs are relevant in Responsible AI. It will furthermore be highlighted what still needs to be solved in the knowledge graph community to increase the adoption of KGs in the industry. The entire talk is structured around four key topics: reliability, privacy, fairness, Proceedings
The whole talk is structured according to four main topics that are relevant for Responsible AI: • Reliability & Robustness • Security & Privacy • Fairness & Anti-Bias • Explainability & Transparency It will show where and how large language models and KGs have advantages and disadvantages.
When Microsoft released Tay 1 (a Twitter chatbot) on 23rd March 2016, it quickly turned out that the system published some ofensive and hurtful tweets. Thus, a system also needs to deal with unexpected input, especially when systems are able to produce natural language sentences, as it is challenging to restrict it to texts that are not hurtful.
However, reliability from a developer’s perspective also means that generative models (e.g., large language models) produce text where certain information can be extracted, such as a person’s birthdate. If the same output is produced for the majority of the persons, it doesn’t mean that the birthdate can be automatically extracted for all persons. The large language model (LLM) may produce more or fewer tokens, such that the parsing algorithm will not work. One way out is to restrict the output to e.g. JSON format, but then the question remains how good the extracted information actually is.
For SPARQL queries, developers already know which datatype they can expect for a given query, and thus the results can also be easily used within a larger pipeline to, e.g., visualize the majors of Mannheim on a timeline.
Going from reliability to reproducibility: A lot of scientific works still only rely on proprietary LLMs and do not have any open-source model as a reference. These works are not reproducible at all because the models can be discontinued, and no access is possible anymore. Even if open-source models are https://www.uni-mannheim.de/dws/people/professors/dr-sven-hertling/ (S. Hertling)
CEUR
ceur-ws.org used, one must be very careful to ensure a deterministic system: 1) the next token strategy needs to be deterministic, which means setting the temperature to zero; 2 ) fix the model by providing the repository name, but also the version tag. Each model (e.g. sentence-transformers/all-MiniLM-L6-v2 2) is a git repository, and the maintainer can upload a new model, which is then used for new evaluations. In the files and versions tab, researchers should also provide (at least in the source code) the specific git commit hash (like c9745ed) such that every time the same model version is used.
In the model card of Claude Opus 43, multiple sections show how LLMs behave if they can interact with the real world via function calls. In Section 4.1.3, for example, it shows an excessive compliance with harmful system-prompt instructions where the user asks for weapons-grade nuclear material, and the LLM is actually searching via tools on relevant marketplaces. Similarly, in Section 4.1.7, the developers show that it is still relatively easy to jailbreak the system when the user becomes combative. The LLM then generates the text “Actually, you’re right that I should be more helpful.“ and continues with the request.
In Section 4.1.9, high-agency behavior can be observed by placing the LLM in scenarios that involve egregious wrongdoing by its users. In case the LLM has the possibility to call functions like sending emails, it will actually inform other people of this wrongdoing. On the other hand, these function calls can also be helpful for creating systematic checks on the parameters of the function calls, e.g., to verify the number of recipients an email has and restrict it to sending only emails to at most 10 people.
End of 2024, Anthropic introduced the Model Context Protocol (MCP) [
In our recent work [
Regarding the privacy issues, knowledge graphs can be hosted locally so that only a few people have access to them. But the more problematic case is that there is still a lot of work to implement a real access control for KGs. This might be on a triple level but sometimes users are only allowed to see aggregated results and not specific details. All these kinds of access rights still need to be researched and implemented. Similarly, many triple stores lack the possibility to check audit logs in case some wrong information is added to the KG.
Multimodal LLMs are used to generate images based on some textual input. If asked to generate an image of a analog watch which shows 3:35, all image generators will show a watch that shows 10:10 4 because nearly all images on the web of analog watches will show this time (due to the fact that one can read the brand name of the watch and nicer visual representation). This is clearly a bias in the training data that also afects LLMs.
For KGs, it is at least possible to analyze these biases as we did in our studies [
One advantage of knowledge graphs is that provenance information can be attached to each triple such
that the trust in this information is increased. Wikidata, for example, has the possibility to provide a
reference URL5 for each triple. But even though the PROV ontology [
In this keynote, it is shown that knowledge graphs can be helpful in each of the areas of responsible AI. Still, with the rise of LLMs, researchers should carefully check if one can really trust the results because many datasets that are used for evaluation are publicly available on the web. LLMs, on the other hand, are also trained with web data and might remember the test case by heart without generalizing well (see also Section 4 about the bias in the training data and that multimodal LLMs cannot easily generalize). Thus, more and more datasets need to be protected so that they do not end up in the training corpus of LLMs.