Aleph Alpha reaches first milestone on the way to content-correct, explainable and trustworthy AI

13 April 2023, Heidelberg

• New functionality makes AI-generated results traceable and verifiable
• Successful prevention of hallucinations and results without a trustworthy source
• Generative AI is now applicable for information-critical areas in law, healthcare, and banking

The Heidelberg-based AI start-up Aleph Alpha has achieved a breakthrough in the field of transparency and explainability of large AI language models, which even transfers to multimodality. An extension of the in-house language model Luminous, available as of today, is able to trace correlations in information and factual correctness of the AI language model Luminous on the basis of verified facts. The extension also shows which text passages in a source have caused or contradict the answer generated by the system. Since 2021, Aleph Alpha has also offered multimodality in the Luminous family, and so this function can also be fully used for facts from image content. This is an important step towards more explainability and traceability – a regulatory requirement for the use of generative AI under the forthcoming EU AI Act that has not yet been finalised but is suspected. “Luminous thus offers a unique opportunity for transparency and traceability of AI-generated content,” says Jonas Andrulis, CEO and founder of Aleph Alpha. “This transparency will enable the use of generative AI for critical tasks in law, healthcare, and banking – areas that rely heavily on trustworthy and accurate information.”

Building new trust in AI

Until now, it has not been possible for the new generation of large language models to understand the decision-making processes of AI language models in such a way that an expert could quickly identify errors. The resulting associations of AI as a “black box” and concern about invisible “bias” often led to a loss of confidence in new AI technologies. For example, according to a study conducted by the digital association Bitkom e.V. in 2022, around half of the companies surveyed worried about a lack of comprehensibility of the results. This is where Andrulis sees the particular strength of the newly developed method: “We now have the opportunity to strengthen trust in generative AI and create a better understanding of how our technology works, so that the use of AI is no longer limited to non-critical and game-changing applications.”

A tool against false content and hallucinations

In recent weeks and months, generative AI has not only often shown impressive capabilities, but also repeatedly produced results that were wrong in content but sounded plausible. With the new Luminous function, content-related facts in the model output can not only be inferred in a comprehensible way, but outputs without matching trustworthy sources can also be withheld directly, says Andrulis: “Our customers often have verified internal knowledge with great value. We can now build on this and ensure that an AI assistance function only uses this knowledge successfully and always provides context.”

Progress through research

The feature is based on Aleph Alpha’s latest research, which was published academically earlier this year under the name “AtMan”. As an AI company with a strong internal research team, Aleph Alpha is uniquely placed to implement findings from its own research directly into its products. In cooperation with globally active partners, the transition period between research and product development is thus quickly closed. For AtMan, Aleph Alpha’s internal research team collaborated with Prof. Kristian Kersting from TU Darmstadt. “We are very proud that together with our partners we can create new crucial opportunities, from research to development to product launch.” (Jonas Andrulis)