Trustworthy AI: Aleph Alpha builds fact-checking into its own model

Hallucinations and the confabulation of plausible-sounding statements that are factually incorrect are considered to be fundamental shortcomings of large AI language models. They make their use in areas where accurate information and safety-critical information is important tricky. According to its own statements, the Heidelberg AI start-up Aleph Alpha has for the first time reached a milestone on the way to artificial intelligence (AI) that is correct in terms of content, explainable and trustworthy.

A now available extension of the in-house language model Luminous is able to understand connections in information and factual correctness on the basis of secured facts, the company explained on Thursday. At the same time, the system is able to show which text passages in a source caused the generated answer or which contradict it.

If the information provided by Aleph Alpha is confirmed in independent tests, the Heidelberg company would have taken a step towards cushioning a systematic weakness in text machines such as ChatGPT. So far, it has often been unclear how and why AI systems like ChatGPT provide certain answers. Validating the answers requires further research by the user – as well as the ability to doubt credible-sounding information beyond one’s own specialist knowledge. As a rule, the models are not able to refer to their sources or indicate degrees of uncertainty in the information provided. Users complain that large language models (LLM) often deliver wrong facts. The topic of explainable AI (XAI for short) has been occupying research teams who want to make AI more transparent for some time. The newly introduced feature is based on Aleph Alpha’s latest research, published academically in early 2023 under the name AtMan in collaboration with TU Darmstadt and DFKI.

AtMan stands for the method of “attention manipulation” presented therein, i.e. the control of attention within transformer models. The name may also be an allusion to the fundamental research work of Google Research, in which transformer models were presented to the public for the first time (“Attention is all you need”, lecture and paper from the NeurIPS 2017 conference). A tweet about the release of AtMan received international attention among researchers who deal with AI security, and the AtMan paper was presented to a broader public by Timnit Gebru’s research institute DAIR, among others.

The transparency and traceability of AI-generated content will “enable the use of generative AI for critical tasks in the legal, healthcare and banking sectors,” said company founder and CEO Jonas Andrulis. With its AI project, Aleph Alpha is considered a beacon of hope in the German software industry. In contrast to the major US competitors OpenAI, Microsoft and Google, the Baden-Württemberg company is not designed for private end customers, but offers its model primarily through partner companies and cooperations. Luminous can be reached indirectly via customers of Aleph Alpha, for example via the public information system Lumi of the city of Heidelberg or via offers from partners such as

Luminous is not a pure language model, but handles multimodal input in words and images in any combination. According to Aleph Alpha, the new feature also applies to explaining image content, as a short demo illustrates:

In view of the forthcoming regulations by the EU AI Act, users and developers of generative AI systems will be subject to stricter requirements. Although the draft has not yet been finalized, it is already becoming apparent that the future rules of the game for AI in the EU will place value on transparency, explainability and data protection – according to legislators, generative AI systems will in future be subject to stricter AI security requirements. Otherwise, the use of AI will have to be limited to non-critical and playful areas.

also read

If you want to delve deeper, you can find the AtMan research paper on (“AtMan: Understanding Transformer Predictions Through Memory-Efficient Attention Manipulation”). Another note is interesting in this context: a plug-in for ChatGPT created by Stephen Wolfram and his team causes the chronically raving AI system of OpenAI to provide factually more correct information, to calculate correctly and to use current data from the Internet to consider. Statistical AI systems that are trained to complete texts according to pattern recognition rules (such as GPT-4 or Luminous) can apparently also benefit from this Supplementing symbolic approaches (like Wolfram’s math and science-centric AI model). More on this can be found in a blog entry by Stephen Wolfram. It remains exciting to see which approaches will prevail and shape future applications.

Anyone who would like to try out the Luminous model family can set up a test account on the Aleph Alpha website – the new “AtMan & Explain” function can be tested in the playground. It can be used either via an interface (API) or via the interactive user interface in the playground. After registration, free credit is available to try out, start-ups and scientific institutions can apply for additional free credits. As the company emphasizes, user data is neither stored nor processed in the playground, and Aleph Alpha also has its own data center in Bavaria.


To home page

Related Posts

Hot News


usefull links

robis robis robis