Use Cases for Trustworthy AI

How can AI technology be used safely and securely? Discover our examples of powerful and trustworthy AI systems from various application areas.

There is a high demand for AI solutions in industrial applications. These solutions need to be efficient, trustworthy and secure in order to be used in series production or quality control, for example. However, the new possibilities opened up by generative AI also raise questions. Can users rely on what a chatbot says? How can vulnerabilities in an AI model be detected early on during development? 

We at Fraunhofer have therefore developed solutions and tools that are powerful, trustworthy and reliable, as well as compliant with European data protection standards.

Are you interested in working with us? Please contact us!

There is a high demand for AI solutions in industrial applications. These solutions need to be powerful, trustworthy and safe in order to be used in series production or quality control, for example.

How can companies ensure this? Our AI Assessment Catalog and AI assessment tools are two key instruments in achieving this.

Trustworthy AI thanks to the AI Assessment Catalog: Guidelines for companies

Fraunhofer IAIS has developed the AI Assessment Catalog, a detailed and extensive catalog that helps companies operationalize requirements relating to the trustworthiness of their AI applications. These guidelines, which have already been tested with successful results in numerous pilot projects, are available free of charge. Our goal is to establish standards for AI assessments and pave the way for independent AI certification.

Assessment tools for trustworthy AI from the company, developer, auditor and assessor perspective

Fraunhofer IAIS supports companies in the mechanical and plant engineering sector in designing the use of AI to be safe and reliable. Our AI assessment tools provide support with systematic quality assurance for AI systems. They enable companies and developers to identify and improve vulnerabilities. For example, systematic tests show whether AI models have learned errors or weak points. If there are systematic weaknesses, this indicates distortions the models have learned from the training data — in that the model only recognizes a certain kind of quality defects in certain scenarios, for example. It is impossible for humans to identify these kinds of systematic errors without technological tools. Our assessment tools offer suitable instruments for this. They can support developers, auditors, and assessors alike. This makes the tools useful both for development of AI systems within companies and during quality assessments by testing and certification bodies.

Your partner for safe and trustworthy AI technologies

As part of the "ZERTIFIZIERTE KI" KI.NRW flagship project, the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS is working with high-profile partners such as the German Federal Office for Information Security (BSI) and the German Institute for Standardization (DIN) to make the future of artificial intelligence (AI) safe and trustworthy.

Additional links:

• Website "ZERTIFIZIERTE KI" (in German)
• AI Assessment Catalog

Dealing with uncertainty plays a crucial role in AI predictions, especially in security- and safety-critical domains such as medicine and autonomous driving. The uncertainty wrapper from Fraunhofer IESE is a method of managing uncertainty in AI. It is used to conduct reliable uncertainty assessments of predictions made by AI models. This makes it possible to quantify and communicate uncertainties in the prediction.

Effective uncertainty management allows for more accurate and more reliable predictions, which in turn brings greater acceptance of and trust in AI systems. Improved understanding of the uncertainties of AI predictions also enables sounder decisions in various fields of application.

Reliable predictions in medicine

In medicine, effective uncertainty management can help to more accurately adapt AI models to various patient groups and medical conditions. For example, they can use MRI scans to make diagnoses or predict how certain treatments might affect the patient. However, these kinds of predictions always carry a certain amount of uncertainty stemming from factors such as image quality or the complexity of the model. The uncertainty wrapper can be used to quantify the uncertainties surrounding diagnoses, thereby helping with decision making.

Handling uncertainty in autonomous driving

In the field of autonomous driving, the uncertainty wrapper can also help to identify and assess uncertainties in detecting and monitoring for obstacles. This improves safety on the road. In inclement weather, a street sign may be hard to see for either a driver or AI. This is where the uncertainty wrapper comes in, quantifying the reliability of the predictions made by the AI. That gives the driver or autonomous system a better sense of whether they can rely on the AI-generated predictions.

Additonal links:

• "KI-Vorhersagen auf der Spur – oder: Wie ein gutes Uncertainty Management den Umgang mit KI-Modellen erleichtert"
  Investigating AI predictions — or: How good uncertainty management facilitates work with AI models — Fraunhofer IESE blog (only available in German)

Artificial intelligence (AI) is coming to play an increasingly important role across almost all industries and processes. Generative AI (GenAI) models in particular will take center stage in the future. Safety and trustworthiness are crucial in this development, but they are not the only issues. Training data and algorithms are also becoming key competitive factors. For example, European languages are significantly underrepresented in the data used to train models from American providers.

Open-source AI model

The OpenGPT-X consortium project, which is receiving funding from the German Federal Ministry for Economic Affairs and Climate Action (BMWK), is developing large language models under the leadership of the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS and the Fraunhofer Institute for Integrated Circuits IIS. In contrast to the American market leaders, OpenGPT-X takes into account a large number of European languages, including German, Spanish, and English. The objective is to make OpenGPT-X available to European companies with open source code, thereby helping to ensure the digital sovereignty of Germany and Europe.

Data sovereignty and data protection

Companies looking to implement generative AI are forced to choose between cloud-based and locally operated infrastructures. Especially in sensitive areas such as healthcare and public administration, safeguarding data sovereignty and ensuring transparency in relation to the use of data are crucial. The open-source approach taken by OpenGPT-X allows companies to adjust algorithms to reflect their specific needs while also maintaining control over their data at all times.

Use cases from the OpenGPT-X project

The AI model can be used for automatic claims adjustment in motor vehicle insurance, for example. AI-assisted document analysis makes it easier to settle claims. A digital assistant can help customers have their claims processed quickly and fairly. In the automotive industry, too, a conversational AI system allows vehicle users to ask questions and get answers via an interface. Large language models (LLMs) are used for this to manage and control entries and provide domain-specific information.

Your partner for a trustworthy large language model

Fraunhofer IAIS is contributing its expertise to the OpenGPT-X project as a leading European research institution in the field of artificial intelligence and big data, and Fraunhofer IIS is doing the same for the domain of audio and voice signal processing. 

Additional links:

• Website OpenGPT-X

In times when specialists are in short supply and businesses are increasingly forced to rely on inexperienced personnel, the likelihood of operator errors is rising rapidly. This not only brings increased waste in the form of defectively manufactured products but also an unnecessary increase in the materials and energy consumed. But there is a solution: Artificial intelligence (AI) can identify these operator errors and defects due to wearing processes early on, thus reducing them — all with the help of energy data.

Federated learning

DigiWeld, an AI solution developed by the Fraunhofer Institute for Manufacturing Engineering and Automation IPA and the University of Stuttgart, uses an innovative approach known as federated learning for this. This method makes it possible to utilize data from various users in industry without any need to store the data centrally. This is achieved through decentralized training of individual AI models and joint use of model parameters instead of the actual data. In this way, sensitive production data stays safely in the hands of users while the AI receives effective training.

The AI solution detects anomalies in industrial workflows purely by analyzing energy data, which is typically collected without additional sensors. The experts from Fraunhofer IPA use a specific process used in production to show how this works: arc welding. Operator errors can be detected early on and prevented using just a few pieces of data on electricity consumption, voltage, and wire feed. The research team also continuously evaluates how to strike a good balance between privacy and the detection rate by adjusting the federated learning parameters.

Your partner for AI in production

In DigiWeld, Fraunhofer IPA has devised a solution that allows makers of machinery and systems to tap into the benefits of artificial intelligence without having to collect sensitive production data from their customers.

Additional links:

• Website Fraunhofer IPA »Energiedatenanalyse« (in German)
• Website Fraunhofer IPA »Fehler erkennen durch Energiedaten« (in German)
  
• Press release Fraunhofer IPA »Fehler beim Schweißen automatisch erkennen« (in German)

The spread of disinformation and fake news in relation to events such as the coronavirus pandemic and the wars in the Middle East and Ukraine is a serious challenge. Forwarded and shared without checking, fake news can spread swiftly on social media platforms. This causes uncertainty and loss of trust across society.

Researchers at the National Research Center for Applied Cybersecurity | Fraunhofer SIT have tackled this challenge from the perspective of forensic text analysis and forensic multimedia analysis as part of the DisCo (Disinformation and Corona) project. Their first step was an in-depth analysis of the current fake news landscape in text and images. As part of this analysis, they concentrated on studying the typical methods and techniques used to produce fake news in the context of current crises.

Your partner in the fight against disinformation

One of the team’s key areas of focus was developing a demonstrator that shows how AI technologies can help stakeholders such as journalists by automatically identifying and highlighting passages of text that should be verified. To this end, they studied machine learning methods that predict which sentences should be prioritized for fact checking. Tools such as inverse image searches provided by various search engines were also tested with an eye to effectiveness and performance. These tools make it possible to identify images that have been taken out of their original context, a common feature of fake news.

The human element in the fight against fake news

Even with cutting-edge technological solutions, human judgment is always needed to tell truth from fake news. With this in mind, the researchers working on the DisCo project worked closely with fact-checking sites to ensure that the results of their technological analysis were evaluated by qualified specialists. Ultimately, deciding what is true and false is still the responsibility of journalists and fact checkers.

Additional links:

• Podcast "Deepfakes"(in German)
  
• "Fake News im Netz automatisch erkennen"(in German)
• Project DisCO at Website ATHENE

The more important the role of artificial intelligence (AI) in our world becomes, the more urgent the need to take human-centered and ethical issues into account when developing and using AI systems. These days, companies need to do more than just recognize the economic benefit of AI. They also have to ensure that solutions that put humans and their requirements and needs at the forefront are both developed and used. Human-centered consideration of AI is the focus of the research on human AI done at the ADA Lovelace Center by Fraunhofer IIS.

Acceptance of AI in healthcare

While others are concentrating primarily on the economic or technological possibilities opened up by AI, this area of the project (ADA Lovelace Center) focuses instead on research on human-centered AI. Pursuing what they call “human AI,” the researchers explore the individual and societal roles of AI and the ethical principles that should apply during development and use of AI. Drawing on the social and behavioral sciences, researchers here specifically aim to explain why (or why not) people accept AI and what the consequences of using AI are for individuals.

AI has vast untapped potential for use in the healthcare sector in particular, from analyzing people’s vital signs in preventative care to imaging and big data analyses for diagnosis and treatment purposes through to robots that deliver care. But at the same time, this industry has a greater need than almost any other for human connection, to get people involved, address their concerns, and deal sensitively with their data.

User-centered recommendations

Research in the field of human AI takes place across various projects, including “KI-BA: Künstliche Intelligenz in der Versorgung — Bedingungen der Akzeptanz von Versicherten, Ärzten und Ärztinnen” (Artificial Intelligence in Healthcare — Conditions of Acceptance by Patients and Doctors), which is receiving funding from the Innovation Committee at the Federal Joint Committee. The project is working to identify individual and contextual factors affecting acceptance and use of AI applications across various areas of healthcare. They include aspects like education, income, and gender, but also technological affinity, personality, healthcare and living situations, and personal networks.

The goal of the acceptance study, which is being conducted with 500 doctors and 1,500 patients, is to understand the factors influencing acceptance among these groups and use this information as a basis for establishing practical, user-centered recommendations for the use of AI in healthcare. The ultimate aim is to make these recommendations available to patients, family members, physicians, medical centers, health insurance funds, insurers, and government agencies as well. By working with various stakeholders, the researchers plan to craft concrete recommendations for actions to take to counter the acceptance risks that have been identified and ensure ethical and human-centered use of AI in healthcare.

Your partner for human-centered AI

With its research in the field of human AI, the ADA Lovelace Center for Analytics, Data and Applications at the Fraunhofer Institute for Integrated Circuits IIS will drive the future of AI toward human-centered design. The goal is to use AI not just effectively but also responsibly and in ways that inspire trust.

Additional links:

• KI-BA: Artificial Intelligence in Healthcare
• www.ada-lovelace-center.com