The rapid evolution of AI technology has generated much excitement around its potential to enhance our daily lives. As a result, AI experts are having frequent conversations around the best ways to leverage AI in our society and out of this dialogue has emerged the human-centered AI approach. Human-centered AI isn’t about replacing humans with machines, but rather using machines to enhance the human experience.
With human-centered AI, human input is kept at the center of the design and build process. This approach takes advantage of the strengths of both humans and machines, enabling them to collaborate in a way that mutually reduces blind spots. Human-centered AI is created with people’s wellbeing in mind, focusing on technologies that will integrate seamlessly into our lives for the purpose of bettering our overall experience. It’s a means for bridging the gap between human and machine for the benefit of both.
Human-centered or Autonomous AI
We’re at a pivotal moment in AI discussions, with two approaches to how we ultimately build machine learning models. One is the human-centered approach, while the other visualizes a world of machine autonomy. In the latter, the idea is that machines can handle tasks without human intervention because they’re theoretically free from the bias that’s inherent to humans (which isn’t quite true, as it turns out, but more on that later). This approach posits that algorithms can make autonomous decisions swiftly and accurately, whereas humans don’t always think logically or quickly. Therefore, autonomous machines are more efficient.
The human-centered approach to AI responds to the autonomous approach by claiming that machines can never replace key attributes of humans, for instance the full breadth of our intelligence and our creativity. Humans must reject full control by algorithms and instead take the driver’s seat in directing solutions that will complement and strengthen their lives. A human-centered approach advocates for a more effective experience between machines and humans, one that keeps the safety and health of humans in mind both development of solutions and implementation.
The Benefits of Human-centered AI
Approaching AI as a collaborative effort between machines and humans can lead to many positive outcomes for businesses, customers, and society.
Personalized Customer Experiences
When we interact with technology – be it a chatbot, a personalized email, a social network that is tailored to our needs, a search bar that works perfectly, we walk away feeling more satisfied if the interaction was specifically catered to us and our needs. Personalization like this can only occur, though, if our wants, needs, and behaviors were taken into account during the development of the technology. Basing AI development on human science leads to products that offer a more enriching, fulfilling customer experience.
As mentioned, the human-centered approach leverages the strengths of both human and machine in order to overcome each one’s inherent weaknesses. The result is more precise algorithms built from human values. Businesses benefit, in turn, from being able to make highly informed decisions that have the potential to deliver the best outcomes – through applications of predictive analytics in mission-critical use cases, like cloud operations, for example.
If you think that algorithms are the answer to solving the problem of human bias, that isn’t the case: algorithms can actually perpetuate and amplify biases through feedback loops. Unchecked, a biased algorithm won’t provide objective, neutral decisions, which is especially dangerous if the algorithm is making important societal decisions on things like parole, loans, and job candidates.
The human-centered approach keeps humans in the loop while building AI so they can monitor for bias in algorithmic decisions. The approach enables a checks and balances system wherein neither the human or machine are fully autonomous, therefore making it easier to identify ways to make outcomes more inclusive.
While relying on algorithms alone can seem like a better, more predictable choice, a human-centered approach to AI offers a more dependable solution. If humans become increasingly reliant on fully autonomous algorithms, we’ll lose the ability to handle situations when those algorithms don’t work. People should always be there to be the fall-back for an edge case and the AI is unsure how to respond or responds incorrectly – regardless of the use cases. By keeping humans centered in AI, we’ll avoid the problematic outcome of being forced to rely on algorithms that will sometimes fail.
How to Build Human-centered AI
We’ve talked about human-centered AI in a theoretical way, but how do you actually implement this approach in your organization? There are a few key steps you can take during AI development that may help you achieve the level of balance required for human-centered AI.
Leverage the Human-in-the-Loop
The human-in-the-loop approach goes hand in hand with human-centered AI. It means that humans are involved throughout the training, testing, and tuning process of building an ML model. For instance, humans can label the training data used to help the model learn which features to recognize. Humans can also verify the accuracy of the model’s predictions and provide feedback to the model when it makes an error. In other words, humans are part of a continuous feedback loop with the model.
More advanced tools are now available that advance collaboration between humans and machines even further. For instance, during the annotation process, algorithms can provide an initial best guess, or hypothesis, for the given label. The annotator can then use this to determine their own judgment. Algorithms can also verify annotator judgments before the submission of the job. These types of tools enable the annotator and the algorithm to work together in a way that maximizes the accuracy and efficiency of the work.
Be Aware of Human and Algorithmic Bias
Awareness of bias is incredibly important in AI development to ensure you’re not relying too heavily on human judgment or machine judgment. Think about the biases your team could unintentionally introduce to your algorithm and plan mitigation steps to ensure this doesn’t happen. Algorithms can be useful for compensating for human blind spots, but be sure to consistently monitor the model’s output for bias as well; machines can at times amplify human biases.
Build Diverse Teams
Building off the previous point, AI tends to be less biased when it’s built by diverse teams. More homogenous teams often have similar blind spots, biases, and other gaps that could end up reflected in the model. This doesn’t refer exclusively to developers, either. It’s crucial that the people you select to annotate your data have demographic diversity at the least, and geographic diversity depending on your use case. This opens the process up for a greater variety of opinions, which will result in more inclusive AI.
Think of Your Customers
The human-centered approach relies on keeping the human experience as the focus. When you develop an AI product, you want the end result to enhance and positively augment the lives of your customers. You want to deeply understand who they are (this includes their demographics, backgrounds, needs, and locations, among other identifiers) and how they’re going to use your technology. Involving a subset of your end users in the testing and validation processes of model building can be a great way to capture their feedback. You may think your product will be used in one way, only to find out that end users are using it very differently. The only way to know this is to test with them directly.
Use Autonomy Wisely
It’s important to note that machine autonomy does sometimes have its place in AI. There are a few use cases where it’s ideal for a machine to have full control over a decision, especially in cases where human safety is involved. One example is autonomous vehicles. Human drivers are prone to accidents (over a million people die from car wrecks each year around the world) and generally poor decision-makers behind the wheel. Fully autonomous vehicles hold promise for making more efficient and safe decisions on behalf of human drivers, and could be a use case where autonomy should mostly be embraced. Until that becomes a reality, a mix of human and machine intelligence might be a way to make roads safer for everyone. In other words, machine autonomy can be applied wisely for more sensitive use cases.
Advance Human-centered Approaches to AI
It’s all of our responsibility to further conversations in AI that will have a positive impact on our society. The dialogues we engage in can and do influence the priorities and actions of AI practitioners. We need to advocate for AI that’s equitable and provides a net benefit to all of the people who use it, keeping these end users in mind throughout the development process. It’s also important for companies to engage in knowledge sharing when possible, which can bolster confidence in the human-centered approach.
As AI technology advances quickly, it’s more critical than ever to have these discussions on how we collaborate with and use AI. Ultimately, we want to create a technology landscape where humans are enhanced, not replaced, by machines.
Expert Insight From Phoebe Liu – Senior Data Scientist
At Appen, we have a team of experts to help our customers with their training data needs as they build cutting-edge models utilizing human-centered AI. Phoebe Liu, senior data scientist, leads the team to ensure Appen upholds strict standards in regards to our Crowd Integrity. Here are her thoughts and advice on building human-centered AI:
Test And Understand the Human – AI Interaction
Understanding and testing the interaction of humans and AI is essential for a successful user experience. For automatic speech recognition, test with speakers with different accents and different ways of saying the same thing. For NLU in chatbot and voice AI, test with users who interact naturally as if they were to chat with another human. The more you conduct user testing in real-world situations, the smoother the interaction will be between your users and the AI system.
Human-centered AI is Multidisciplinary
It brings together scholars and practitioners from various domains: engineers, psychologists, designers, anthropologists, sociologists, along with experts from other domains. Creating a successful human-centered AI involves the collaboration from a variety of fields to develop the hardware or software, analyze the behaviors of users when interacting with AI in different social contexts, as well as the required domain knowledge for particular applications. This collaboration can be difficult due to the different disciplinary jargon and practices. The common interest in human-centered AI among this wide variety of participants, however, is a strong motivation for familiarizing oneself with and respecting the diverse ways of acquiring knowledge.
Consider the Specific AI Tasks Candidates
AI is a tool built by people and therefore it’s important to consider what specific tasks it is trying to emulate in human intelligence during the design phase. In designing the system, we need to focus on the specific capabilities that the system is trying to emulate – for example, a voice assistant is capable of factual question-and-answer, but is not capable of navigating autonomously in a snowy road condition. When we design human-centered AI, we should first consider whether the task that the AI is trying to accomplish can be learnt given that objective ground truth exists, and if not, how to best leverage or augment the AI system using human expertise.
What We Can Do For You
At Appen, we leverage the best of human and machine intelligence to provide high-quality annotated training data. Our data annotation platform powers the world’s most innovative machine learning and business solutions. Many of our annotation tools feature Smart Labeling capabilities, which leverage machine learning models to automate labeling and enable contributors to work quickly and more accurately. We value the human-in-the-loop approach to AI and keep humans centered in all of the work that we do.
We understand the complex needs of today’s organizations. For over 25 years, Appen has delivered the highest quality data and services, in over 235 languages and dialects, to government agencies and the world’s largest corporations.
To speak with someone directly about your data needs, contact us today.