As AI researchers explore different domains such as sales, finance, and human resources, IBM Research is taking a human-centered approach in developing AI systems.
Despite increasing levels of automation enabled by AI — whether it’s AI driving our vehicles, designing our drugs, determining what news and information we see, and even deciding how our money is invested — the common thread among these systems is the human element. AI’s long-term success is contingent upon our acknowledgement that people are critical in its design, operation, and use.
Human-Centered AI (HCAI) is an emerging discipline intent on creating AI systems that amplify and augment rather than displace human abilities. HCAI seeks to preserve human control in a way that ensures artificial intelligence meets our needs while also operating transparently, delivering equitable outcomes, and respecting privacy.
IBM Research’s HCAI strategy is to rigorously investigate and design new forms of human-AI interactions and experiences that enhance and extend human capabilities for the good of our products, our clients, and society at large. To do that, we take an interdisciplinary approach that involves researchers who specialize in human-computer interaction (HCI), computer-supported cooperative work, data visualization, and design in the context of AI. We explore three important themes that we consider critical for the success of AI systems.
Adhering to the core value that “human + AI” is better than either one individually, we can develop novel user experiences and visualizations that foster human-AI collaboration. We can also create frameworks for designing or evaluating human-AI interaction models and conduct theoretical work that develops and extends theories of human-AI collaboration or co-creation.
Data science offers an excellent example of how people and AI can work together, to augment our ability to glean meaningful insights from data. One of the biggest challenges for data scientists is identifying and analyzing large, disparate data sets in ways that derive new insights that help solve complex problems. To do that, data scientists must build models and measure their performance. Then they optimize those models by tuning their hyper-parameters and evaluate those models for fairness and robustness.
We learned that the future of data science work will be a collaboration between humans and AI systems, in which both automation and human expertise are indispensable.
Using a human-centered approach, we first sought to understand the attitudes that data scientists had in automating their work. In doing so, we learned that the future of data science work will be a collaboration between humans and AI systems, in which both automation and human expertise are indispensable.1 Those insights led IBM to develop our AutoAI technology, which enables data scientists to produce more, higher-quality machine learning models faster, with fewer errors.2 This also includes novel visualization techniques that make it easier for data scientists to compare pipelines and understand how they rate against different metrics of accuracy and bias.3
This area covers all aspects of how At NeurIPS 2021, we organized the first workshop at NeurIPS on Human-Centered AI, exploring research questions that stem from the increasingly wide-spread usage of ML algorithms across all areas of society. It focused on understanding technical and design requirements for HCAI systems, as well as how to evaluate the efficacy and effects of HCAI systems. The workshop was major step toward building a hybrid field of AI and human-computer interaction.human-centered AI systems can provide positive and beneficial outcomes to their direct users, to those affected by their operation, and to society in general. To achieve these outcomes, HCAI must be fair and unbiased, secure, applied ethically, and used in service of users’ needs.
Efforts to develop responsible and human-compatible AI take several factors into account, including the need to understand how people engage with and trust AI systems. There’s also a need to explain the operation of AI models and improve peoples’ understanding of how AI systems operate. An HCAI strategy must also gauge the negative consequences or potential misuse of AI systems, include ways to mitigate human and AI biases, and be able to measure peoples’ perceptions (or misperceptions) of AI systems.
Our collaborations with the IBM’s internal business units offer real-world examples of responsible and human-compatible AI. In one instance, IBM Research, the Chief Analytics Office, and the Chief Information Office partnered to create a set of assets that infuse AI into the tools of IBM Sales and our business partner ecosystem, including the SCORE engine.
SCORE uses AI techniques to learn from sales data and operational feedback from IBM’s sellers, business partners, and clients to recommend the optimal BP to receive each sales lead.4, 5 Explainable recommendations that provided clear evidence as to why a business partner is recommended were critical to the success of this AI system engendering trust in the system. The deployment and uptake of SCORE has led to $214M in incremental revenue since 2018 and won several prestigious awards.
Explainable recommendations that provided clear evidence as to why a business partner is recommended were critical to the success of this AI system engendering trust in the system.
As we explore different domains where AI can be applied, such as sales, finance, and human resources, our researcher often leads to a generalizable, deeper understanding that can result in AI design guidelines for ML practitioners. For example, our research team has been developing explainable AI (XAI) guidelines that can be leveraged by practitioners in multiple domains and for multiple use cases. We are also leveraging data visualization approaches to support interpretability, understanding, and interactive explorations of AI models, such as GLTR, a tool to help humans detect automatically generated text, and GenNI, an interactive visual system for high-level human-AI collaboration in producing descriptive text.
Conversational user interfaces (CUIs) powered by advanced dialogue systems have grown in popularity, and many intelligent assistants have been developed for business, social, and emotional purposes. Under this theme, we are aiming to understand the kinds of tasks appropriate for this medium, the task effectiveness, and how to design engaging and delightful interactive experiences.
In one project, our team is investigating the use of formal and informal styles of language by customer service AI-powered agents, both in speech- and text-based chats. Since the use of formal language is connected to gender in many languages and cultures, we are exploring cases where the machine impersonates either female or male agents and considering a comparison with expectations for human-based customer service.
Our initial findings seem to indicate a direction that challenges traditional gender stereotypes of customer service in Brazil. In a Brazilian Portuguese context, we found a preference for human agents to use more formal language but no preferred use of formality for machine speech-based systems, such as Siri or Alexa. However, in the case of text-based chatbots, we saw a strong preference for female-portraying systems using informal language and male systems using formal language.6, 7
As the technologies we develop become more intelligent and autonomous, our interactions with those systems will transform. To effectively support true The ACM IUI Conference on Intelligent User Interfaces is the premiere venue for work that spans across cutting- edge artificial intelligence and human-computer interaction. From March 21-25, 2022, IBM Research participants presented recent advances related to our Human-Centered AI research agenda. This year, many of our papers focused on generative AI across various areas including code translation, co-creative systems, UX modernization, and explainability.Human-AI partnerships, one important question is how to facilitate the AI system and the user to negotiate a common objective, where traditionally AI solutions focus on optimizing performance or accuracy, these metrics do not encapsulate other objectives, like personalization, fairness and the many other objectives humans trade-off on a day to day basis when making decisions.
A second frontier for Human-Centered AI is to investigate and understand the design of AI systems that become creative partners. In the business world, human-AI co-creativity involves subject matter experts working with an AI system generating code, co-designing user experiences, and accelerating scientific discovery.
We envision the user experience of creating both physical and digital artifacts will become a partnership in which people will take the role of specification, goal setting, steering, high-level creativity, curation, and governance, whereas AI will augment human abilities through inspiration, creativity, low-level detail work, and the ability to design at scale.
Human-Centered AI: We design and investigate new forms of human-AI interactions and experiences that enhance and extend human capabilities for the good of our products, clients, and society at large.
Future of Finance: We're working on core technology innovations, from secure and elastic microservices-based hybrid cloud solutions and advanced analytics for finance, to AI blockchain, and quantum computing for the banking industry.
Date31 Mar 2022
- Note 1: At NeurIPS 2021, we organized the first workshop at NeurIPS on Human-Centered AI, exploring research questions that stem from the increasingly wide-spread usage of ML algorithms across all areas of society. It focused on understanding technical and design requirements for HCAI systems, as well as how to evaluate the efficacy and effects of HCAI systems. The workshop was major step toward building a hybrid field of AI and human-computer interaction. ↩︎
- Note 2: The ACM IUI Conference on Intelligent User Interfaces is the premiere venue for work that spans across cutting- edge artificial intelligence and human-computer interaction. From March 21-25, 2022, IBM Research participants presented recent advances related to our Human-Centered AI research agenda. This year, many of our papers focused on generative AI across various areas including code translation, co-creative systems, UX modernization, and explainability. ↩︎
Wang, D., Weisz, J. D., Muller, M., et al. Human-AI collaboration in data science: Exploring data scientists' perceptions of automated AI. Proceedings of the ACM on Human-Computer Interaction. 3(CSCW), 1-24. (2019). ↩
Wang, D., Andres, J., Weisz, J., et al. AutoDS: Towards Human-Centered Automation of Data Science. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems CHI '21: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. May 2021 Article No.: 79 Pages 1–12. ↩
Weidele, D. K. I., Weisz, J. D., Oduor, E., et al. (2020, March). AutoAIViz: opening the blackbox of automated artificial intelligence with conditional parallel coordinates. IUI '20: Proceedings of the 25th International Conference on Intelligent User Interfaces March 2020 Pages 308–312. ↩
Alkan, Ö., Mattetti, M., Barros, S., Daly, E. Exploiting Heterogeneous Data Sources through User Feedback for a Business Partner Recommender System. Proceedings of the Joint Proceedings of the ACM IUI 2021 Workshops co-located with the 26th ACM Conference on Intelligent User Interfaces (ACM IUI 2021) ↩
P. Cavalin, V. H. A. Ribeiro, M. Vasconcelos, C. Pinhanez, J. Nogima and H. Ferreira, Towards a Method to Classify Language Style for Enhancing Conversational Systems. 2021 International Joint Conference on Neural Networks (IJCNN), 2021, pp. 1-8. ↩
From Disjoint Sets to Parallel Data to Train Seq2Seq Models for Sentiment Transfer (Cavalin et al., Findings 2020) ↩