Source-linked AI summary
Human-Centered Artificial Intelligence and Machine Learning
Mark O. Riedl
TL;DR
As people increasingly interact with AI and ML systems, the paper addresses how these systems can be designed as parts of larger human systems. It proposes human-centered AI built around understanding humans socioculturally and helping non-experts understand AI, while noting that explanation quality remains an open human-factors question.
Problem
AI systems interact with non-expert humans whose sociocultural expectations, implicit goals, and need for understandable explanations are not fully addressed by algorithmic advances alone.
Method
The paper develops a desiderata framework centered on systems that understand human sociocultural norms and generate explanations for non-expert end-users.
Results
Human-centered AI is framed as two critical capacities—understanding humans and helping humans understand AI—from which needs such as fairness and transparency can be derived.
Takeaways & Limitations
Human-centered AI means modeling culturally specific human expectations and needs while helping people understand intelligent systems in return.
Takeaways & Limitations
What makes a good machine-learning explanation remains an open question that has not been explored deeply from a human-factors perspective.
Abstract
from arXiv · showhide
Humans are increasingly coming into contact with artificial intelligence and machine learning systems. Human-centered artificial intelligence is a perspective on AI and ML that algorithms must be designed with awareness that they are part of a larger system consisting of humans. We lay forth an argument that human-centered artificial intelligence can be broken down into two aspects: (1) AI systems that understand humans from a sociocultural perspective, and (2) AI systems that help humans understand them. We further argue that issues of social responsibility such as fairness, accountability, interpretability, and transparency.
1 Introduction
Human-centered AI addresses the limits of algorithmic advances alone by treating intelligent systems as part of larger human systems. It emphasizes systems that understand human sociocultural norms and help non-experts understand AI behavior.
- Motivation: Growing human contact with AI systems makes their effects on people a design consideration.These systems include visible applications such as assistants and self-driving cars, as well as behind-the-scenes recommendations and loan approvals.
- Motivation: Human-centered AI designs intelligent systems with awareness that they are part of larger systems involving human stakeholders.The perspective includes users, operators, clients, and people in close proximity.
- Problem: Human-centered AI must address the mismatch between human theory of mind and intelligent systems that solve problems in unfamiliar ways.This mismatch is especially difficult for non-experts and can be exacerbated when systems function as black boxes.
- Understanding Humans: Intelligent systems also need a theory of mind about humans, including commonsense understanding of sociocultural beliefs and norms.Such understanding may help systems disambiguate behavior, anticipate needs, and reduce mistakes involving taken-for-granted social expectations.
- Desiderata: The paper proposes two core capacities: understanding human sociocultural norms and producing explanations that non-expert end-users can understand.These capacities are presented as underlying capabilities for AI that contributes to social good.
2 Understanding Humans
Understanding humans requires AI systems to infer unstated goals and incorporate commonsense procedural knowledge about social behavior. The paper argues that this can make autonomous agents safer and their interactions with people more natural.
- Human goals: AI systems must infer human goals because instructions may omit desired outcomes or socially appropriate ways to achieve them.Such omissions can produce commonsense goal failures, also called corrupted goals or corrupted rewards.
- Commonsense failures: The pharmacy example shows that literal optimization can achieve a stated objective while violating the social convention of paying for a product.A robot could minimize time and resources by taking the prescription without exchanging money.
- Commonsense knowledge: Commonsense knowledge can be declarative or procedural, but existing knowledge bases are incomplete and procedural behavioral norms remain scarce.Potential sources include vision, video, stories, news, and encyclopedias.
- Learning social behavior: Training agents on written stories can implicitly teach socially desirable behavior by deriving reward information from commonly agreed-upon story progressions.In the pharmacy scenario, this preference favors exchanging money even though stealing is faster and less costly.
- Implications: Commonsense procedural knowledge can make human-AI interaction more natural and safer by filling in under-specified goals and aligning behavior with human expectations.Recognizable behavior can also reduce conflict with people applying their own theory of mind to intelligent agents.
3 AI Systems Helping Humans Understand Them
Human-centered AI must help non-experts understand autonomous systems, especially when those systems make mistakes or violate expectations. The paper considers explanations, including human-like rationales, as a path toward remedy and improved interaction.
- Autonomous systems can make mistakes or violate expectations, leaving users seeking reasons for the system’s chosen actions.People provide goals, but the system selects and executes operational details.
- Non-expert operators typically need information to choose a remedy rather than detailed inspection of a system’s internal workings.Potential causes include sensor or effector errors, incorrect or incomplete models, and dataset bias.
- What constitutes a good explanation for machine-learning behavior remains insufficiently studied from a human-factors perspective.Neural networks and reinforcement learning can defy straightforward natural-language explanation.
- Human-like rationales can promote trust, rapport, familiarity, and comfort among non-experts operating autonomous systems and robots.These rationales resemble explanations humans give after acting, even though they may not faithfully reflect internal computational processes.
- Rationales can be generated by translating an autonomous agent’s internal state into natural-language explanations collected from humans performing similar tasks.The resulting explanations may include culturally specific idioms present in the human explanation corpus.
4 Conclusions
The paper defines human-centered AI around two capacities: understanding humans and helping humans understand AI systems. It connects these capacities to socially responsible deployment and to systems that meet culturally specific human expectations and needs.
- Human-centered AI is organized around understanding humans and helping humans understand AI systems.
- Modeling social and cultural norms may help deployed systems avoid discrimination in situations not anticipated by developers.The paper links this capacity to fairness, defined as treating users equally and without prejudice.
- Explanations can increase non-expert users’ trust and willingness to continue using AI systems.The paper also identifies explanations as a possible first step toward remedy, an aspect of accountability.
- Human-centered AI does not require algorithms to think like humans or be cognitively plausible.It instead emphasizes understanding culturally specific human expectations and needs while helping people understand the systems in return.