Wie erstellt man KI-Systeme, die für alle funktionieren?

Creating an AI system that works for everyone requires a multi-faceted approach. Firstly, a diverse and representative dataset should be used to train the AI, encompassing various cultures, languages, and perspectives. Secondly, an inclusive design process is crucial, involving input from diverse teams with interdisciplinary expertise. Ethical considerations, like fairness, transparency, and accountability, should be at the forefront of development.

In simpler terms, imagine if all your recipes came from one chef; you'd miss out on a world of flavours. Similarly, AI, when built by a narrow group, lacks the richness and fairness of diverse inputs. Think about a classroom where only one kid gets to answer every question – that's not a complete picture. So, for AI that understands and serves everyone better, we need voices from all walks of life.

First, I disagree with the premise of appling AI to facial recognition. The use of facial recognition has very limited legitimate uses but many illegitimate uses that themselves are discrimiation in of itself. Facial recognition is a bad example to use for the argument of making AI fair , ethical, and non-discriminating. However, assuming a legitimate use that was not for the purpose of discrimination, data should be collected from a very wide variety as wide a net as possible and weighted equally across each sub-type and domain, in addition it should be reviewed by a round table of reviewers and stakeholders that are also themselves inclusive of the wide diversity that is being attempted to be achieved. Humans are very unique.

The angle to approach AI working for everyone is not to train the AI models with data scientists in place, to deal with those collected data based on standard training pipeline. Instead, but to see thru to the core, what do people want to do with the AI. It's quite obvious if we can turn it upside down, look at this puzzle from the end.

The field of healthcare AI holds great promise, but we must be mindful of biases that could be life-threatening. Recently, I challenged three different AI chatbots with a scenario: a patient with Adrenal Crisis in the ER. Prompt treatment is vital, and noticing patients who may present atypically is essential. While the chatbots performed well in some areas, they exhibited failures in others. I questioned ChatGPT-4 about its faulty assumptions, and it acknowledged its own bias along with the need for intentional inclusion of rare diseases in the training data. By incorporating patient feedback and actual cases we can work toward more effective integration of AI into the care delivery system.

The path to ethical AI has many steps. But walking it with diverse perspectives lights the way to innovation that empowers all. First, assemble a cross-functional team with diversity of backgrounds and experiences. This foundation sets the tone for equity. Next, engage a wide range of stakeholders early for input on use cases and requirements. Don't make assumptions - learn people's actual needs. Now, thoughtfully source representative data. Test models with excluded groups to reveal blindspots. Iterate based on feedback. Throughout, conduct ethical risk assessments on potential biases and harms. Drive decisions with impact, not just accuracy. Implement accessible and transparent user experiences. Enable recourse for issues.

Unlocking AI's full potential requires embracing diversity and inclusion. By involving diverse voices in every step of AI design, we can create technology that is both innovative and equitable. Making sure everyone can have access to AI no matter their gender, skin, location, religion, ETC, will be the most efficient way to create a diverse system that works for all.

AI systems that work for everyone, also known as building AI systems that are inclusive and unbiased, is a complex and multifaceted process. It involves addressing various challenges related to data, algorithms, design, and ethics. Here's a detailed explanation of how to approach this goal: Diverse and Rep Data: Collect Diverse Data: Ensure that the training data for AI systems is diverse and representative of the entire population. This includes different demographics, backgrounds, cultures, and perspectives. Mitigate Bias in Data: Identify and mitigate biases present in the training data. Biases can lead to unfair and discriminatory outcomes. Use techniques like data augmentation, oversampling, and reweighting to balance the data.

When we talk about a micro-level perspective, we run the risk of falling into biases due to the behavior of population groups. However, this challenge should be addressed on a macro level. Considering the significant cultural migrations we are experiencing and the shifts in awareness towards previously unexplored topics in recent years, I believe that inclusion will become an inherent part of the artificial learning process. It will naturally be considered through techniques of equalization to address it. But first, we must fully address this issue in society in order to effectively convey and teach it.

AI by itself is not biased or unbiased, but how it is trained. As it is well said "charity and kindness begins at home", in the same way, training models made with any AI technology are by humans, and we humans needs to be unbiased.

I recently completed a course at Stanford’s HAI - Center for Human-Centered Intelligence called the Foundation Models Scholars Program. It combined industry with academia in a way that allowed for a much more inclusive approach to innovation. Take a look at some of the incredible research such as HELM for holistic metrics to help organizations measure not only accuracy but also fairness and transparency. Another project I enjoyed learning about was FrugalGPT, which invited deeper thought around using the right model, for the right task, at the right time. I celebrate academia’s involvement in industry solutions to build better AI that serves everyone.

Agreed. But I think a first principle should be that AI should not be used to assess, judge, rank, evaluate humans. Humans are each unique in their own way. No matter how much diligence is put into making it as fair as possible, there will always be a group of people that are being treated unfairly. And that unfair treatment could be causing harm to humans. For that reason alone, AI should be avoided in assessing humans. In some cases, there may be beneficial healthcare applications where AI can provide benefit in achieving the results. From that perspective, the AI should be kept as minimalistic as possible to human exposure and exclusive to the science and math facts as much as possible and avoid other human related aspects.

Ever wondered about fair AI? It’s all about embracing the 𝐅𝐄𝐀𝐓 principles - a recipe for inclusive AI. Unveil AI's choices with Explainable AI (XAI) to understand how AI reaches its decisions. (1) 𝐅airness is a level field as we must avoid data biases that might lurk in AI’s training set. We should think of AI as a game where everyone plays by the same rules. (2) Then comes 𝐄thics, guides AI's conscience – helping, but not harming. (3) 𝐀ccountability answers to 'what if' and who is responsible if something goes wrong. (4) Lastly, 𝐓ransparency is AI explaining like a friend telling why they made a certain decision. Dive deeper in my article on Avoiding Biased AI on McKinsey Medium. Let's shape unbiased AI for all, a shared AI future!

It's vital that AI solutions are not implemented as a "black box". All stakeholders and users need to have some level of transparency into how an AI works and be able to challenge its assumptions. An example I point to is credit scores. For a long time, consumers had little to no idea what their credit scores were or what formulas where being used to calculate them. Now, consumers can get a copy of their credit and all its components and challenge incorrect items. They can even see a generalized formula for how their score was determined. I hope that AI adopters follow a similar track of transparency and allow those affected to challenge incorrect assumptions. Especially those that directly impact our health, career, or finances.

Frameworks are compasses, communities the navigators. Together they steer AI towards empowerment, not exclusion. This journey belongs to all of us.

2. Fair Algorithms: Algorithmic Fairness: Use fairness-aware algorithms that prevent discrimination and favoritism. Ensure that the AI system's decisions are not disproportionately affecting certain groups. Regular Auditing: Regularly audit the AI system for bias and fairness issues. Evaluate the system's outcomes across different groups to detect and address potential disparities. 3. Transparent and Explainable Models: Interpretable Models: Choose models that are inherently interpretable. It should be possible to understand how the model arrives at its decisions. Explainable AI: Develop techniques to explain the rationale behind AI decisions. This is important for building trust and ensuring accountability.

Sooner or later, artificial intelligence will affect our lives much more than it does now, which is already quite substantial, though we might not be fully aware. Eventually, through natural evolution, humanity will have to establish supervisory groups delegated by governments that, under a unified protocol, can exert control over whether a boundary should be crossed or not. However, speculating about the future, the rigidity of this boundary could vary depending on the case and the larger interests at stake. I believe that as creators, we have an obligation to contribute transparency, legality, and trust under unified standards.

Having inclusive user research practices is crucial. Some of the best teams I've worked with are intentional about recruiting a diverse and representative user study group. 

By reaching out to such diverse groups, we ensure that we are collecting comprehensive data about our users, their needs and pain points. 
 Whenever possible, we also like to go back to these users and show them our solution or get them to use it firsthand to further uncover feedback on the real-world use of the AI solution. These practices are key to helping build a strong foundation for designing experiences that are inclusive, user-friendly, and safe for all.

Different AI systems may require different user research and testing methods. LLMs for e.g. already have 'internet scale biases' so should be accounted for. User research and testing (UXR/T) data should be complemented by other data sources and checked for biases, errors or limitations. LLMs always have a fixed-in-time nature and might not fit the use-case data needs! (UXR/T) should also be continuous and iterative throughout. A product needs signals from users as fast as possible and often needs feedback at every moment of value provided to the user e.g. 👍 👎 use Data feedback loops should in place to improve unique data over time and pilot rollout/testing is often simplest method to learn if the solution is close to ideal PMF

User-Centric Design: Inclusive Design: Design AI systems that are accessible to a wide range of users, including those with disabilities. Consider factors like screen readers, voice interfaces, and other assistive technologies. User Feedback: Gather feedback from diverse users during the design and testing phases. Incorporate their feedback to make improvements that cater to various needs. Ethical Considn: Ethics Guide: Establish ethical guidelines and principles for the development and deployment of AI systems. Ensure that the technology respects human rights, dignity privacy. Diversity in Develop: Build diverse teams with varied backgrounds and expertise to avoid blind spots and ensure a well-rounded perspective in decision-making.

With the right foundations, vision becomes reality. Tools remove friction so teams can focus on people-first innovation. ↳ Open source libraries expand access for more to build AI. ↳ Cloud platforms enable scaling to serve larger communities. Align with the Enterprise Cloud Strategy. ↳ Automation and MLOps multiply team productivity. Frees up time to improve UX. ↳ Visual tools democratize development beyond coders. ↳ Built-in fairness toolkits reduce harmful bias risks. ↳ Interpretability features bring transparency to users.

Continuous Monitoring: Monitoring and Maintenance: Continuously monitor AI systems in real-world scenarios to identify any emergent biases or issues. Regularly update and improve the system to address new challenges. Feedback Loop: Create mechanisms for users to provide feedback on the AI system's performance and impact. Use this feedback to make necessary adjustments. Regulatory Complian: Adherence to Regulation: Ensure that the AI system complies with relevant regulations and standards related to fairness, privacy, and data protection. Transparency: Be transparent about the data sources, model architecture, and decision-making processes used in the AI system. This helps regulators and stakeholders understand the system's functioning.

I believe it's not about handling a vast array of tools, but about knowing how to solve each problem by seeking the best solution. In some cases, it will be creating a customized tool to oversee the entire process; in others, the optimal solution might come from an API performing model inference in a cloud environment. I think we're on the brink of experiencing even greater growth in fast machine learning systems, creating inclusive environments that were previously accessible to only a few.

Every person and community has gifts. AI should amplify them. By expanding its reach, we expand possibilities for all. Great examples btw.

Alongside my colleague Antonio at Maiacorp, we are preparing a virtual tutor that can assist students' doubts 24/7 in a school and guide them through processes like enrollment, prework, exercise resolution, and questions. Often, we don't put ourselves in others' shoes, and an anxious student requires prompt answers that can't wait for days. Sooner or later, all educational institutions will need to adapt and create their own solutions

Education and Awareness: Public Awareness: Educate the public about how AI systems work, their limitations, and the steps taken to ensure fairness and inclusivity. Training: Train AI developers, data scientists, and decision-makers on ethical considerations, bias mitigation techniques, and responsible AI practices. Continuous Improvement: Learning from Mistakes: Recognize that achieving perfect fairness and inclusivity is challenging. Learn from mistakes, be open to feedback, and continuously improve the system. Collaboration: Collaborate with other experts, organizations, and researchers in the field of AI ethics to collectively address challenges and share best practices.

Creating AI systems that work for everyone involves several key steps. First, ensure diverse representation in the development team to consider different perspectives. Second, collect diverse and inclusive training data to prevent biases. Third, implement continuous testing and monitoring to detect and mitigate biases and errors. Lastly, involve user feedback and iterate on the system to improve its performance for all users.

As we harness the immense potential of AI across multiple domains, we must be aware of its limitations and biases, particularly in complex, sensitive contexts like healthcare and specifically rare diseases (1 in 10 Americans have a rare disease). By addressing these biases head-on, we can ensure that AI becomes a reliable, useful tool in the medical field, augmenting human decision-making and ultimately improving patient outcomes. If we do not focus on rigorous training in a transparent environment, we risk causing great harm in violation of the hippocratic oath. The stakes are high!

Designing AI that works for all is like playing chess with a diverse team. The pawns - representing communities - move together to scope needs and set direction. Nothing moves without them. Knights and bishops - technical leads - employ specialized skills to execute the vision. But they only go where pawns guide. Rooks and queens - stakeholders - defend boundaries and clear obstacles. They protect the shared mission. The king - ethical AI - relies on the collective. No one piece works alone. Each contributes unique value. When every piece participates, the board transforms. Limitations become strengths. New strategies emerge. Together, we checkmate exclusion. Our diversity is AI's greatest superpower.

In summary, creating AI systems that work for Overall Summary: Everyone requires a holistic approach that considers diverse data, fair algorithms, transparent models, ethical considerations, and continuous monitoring. It's an ongoing effort that requires collaboration, research, and a commitment to building technology that benefits society as a whole.

Prioritizing the end user should be our fundamental starting point. In the current context, open and accessible solutions sometimes are often developed later in projects. However, this perspective must change, placing the user at the center from the beginning. The premise is to create a solid foundation for optimal functioning and stability. This involves comprehensively addressing identified needs in relation to the problem. Just as we responsibly use natural resources, technological resources must be employed efficiently. Envision not just a vast ocean, but a small lake capable of effectively supplying a community.

Before creating an AI system, it is important to consider whether or not you really need it. There may be simpler, more traditional solutions that can achieve the same results.