# Human–AI interaction

Human–AI interaction is the interdisciplinary study and design of how people engage with artificial intelligence systems, covering interfaces, trust, and collaboration. It spans fields like computer science, cognitive science, and ethics, shaping tools from chatbots to autonomous vehicles.

Human–AI interaction (HAI) is the interdisciplinary field concerned with the design, understanding, and evaluation of interactions between humans and artificial intelligence systems. It draws on computer science, cognitive science, human-computer interaction, and ethics to create interfaces and workflows that are effective, safe, and aligned with human values. The field addresses both the technical mechanisms of AI, such as [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) models, and the social and psychological dimensions of how people perceive, trust, and collaborate with these systems.

Unlike traditional software interaction, human-AI interaction involves systems that can learn, adapt, and make decisions autonomously. This introduces unique challenges, including explainability, transparency, and the management of user expectations. The discipline has grown rapidly since the 2010s, driven by advances in [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) and the proliferation of [generative-ai](https://www.wikiprompt.org/wiki/generative-ai) tools that are now used by millions of people daily.

## Historical Foundations

The roots of human-AI interaction trace back to early experiments in [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) during the 1950s and 1960s, when researchers like Alan Turing proposed tests for machine intelligence. The field of human-computer interaction emerged in the 1980s, focusing on usability and user-centered design, but it initially treated AI as a peripheral component. By the 1990s, systems like [chess-computer](https://www.wikiprompt.org/wiki/chess-computer) programs demonstrated that AI could outperform humans in specific tasks, prompting research into how humans could effectively supervise or collaborate with such systems.

A pivotal shift occurred in the 2010s with the rise of [neural-network](https://www.wikiprompt.org/wiki/neural-network) based models and [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)s. These systems, trained on massive datasets, exhibited natural language capabilities that made interaction more conversational. The introduction of the [transformer](https://www.wikiprompt.org/wiki/transformer) architecture in 2017, detailed in the paper "Attention Is All You Need" by researchers including [jakob-uszkoreit](https://www.wikiprompt.org/wiki/jakob-uszkoreit) and [lukasz-kaiser](https://www.wikiprompt.org/wiki/lukasz-kaiser), enabled models to process context more effectively, leading to the development of assistants like those from [openai](https://www.wikiprompt.org/wiki/openai), [anthropic](https://www.wikiprompt.org/wiki/anthropic), and [google-deepmind](https://www.wikiprompt.org/wiki/google-deepmind).

## Core Principles and Design Challenges

Human-AI interaction is guided by several principles that distinguish it from conventional interface design. One key principle is transparency: users should understand what the AI can do, how it makes decisions, and when it is uncertain. This is particularly important for high-stakes applications such as medical diagnosis or autonomous driving, where errors can have serious consequences. Another principle is controllability, meaning users should be able to override or redirect AI actions when necessary.

Designers face challenges such as the "black box" problem, where [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) models are difficult to interpret. Techniques like [model-pruning](https://www.wikiprompt.org/wiki/model-pruning) and [rlaif](https://www.wikiprompt.org/wiki/rlaif) (reinforcement learning from AI feedback) aim to improve efficiency and alignment, but they do not fully solve explainability. Additionally, users often over-trust or under-trust AI systems, leading to misuse or abandonment. Research in this area, including work by scholars like [melanie-mitchell](https://www.wikiprompt.org/wiki/melanie-mitchell) and [brian-christian](https://www.wikiprompt.org/wiki/brian-christian), emphasizes the need for interfaces that calibrate trust through clear communication of confidence levels and limitations.

## Applications and Domains

Human-AI interaction manifests across a wide range of domains. In healthcare, systems like [commure](https://www.wikiprompt.org/wiki/commure) and [intuitive-surgical](https://www.wikiprompt.org/wiki/intuitive-surgical) support clinicians by analyzing medical images or assisting in robotic surgery, requiring careful design to ensure that human oversight remains central. In transportation, [waymo](https://www.wikiprompt.org/wiki/waymo) and [tesla-autopilot](https://www.wikiprompt.org/wiki/tesla-autopilot) develop autonomous vehicles where interaction involves not only the driver but also pedestrians and other road users, raising questions about how AI communicates its intentions.

In the workplace, AI-powered tools from companies like [microsoft](https://www.wikiprompt.org/wiki/microsoft) and [google-cloud](https://www.wikiprompt.org/wiki/google-cloud) assist with tasks such as email drafting, data analysis, and code generation. These tools often rely on [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)s and require interfaces that let users provide feedback and corrections. Creative applications, including image generation and music composition, have also emerged, with systems from [openai](https://www.wikiprompt.org/wiki/openai) and [sora](https://www.wikiprompt.org/wiki/sora) (a video generation model) pushing the boundaries of human-AI co-creation. The field also addresses accessibility, ensuring that AI systems are usable by people with disabilities, and education, where adaptive learning platforms tailor content to individual students.

## Evaluation and Metrics

Evaluating human-AI interaction goes beyond traditional usability metrics like task completion time. Researchers use measures such as appropriate trust, which assesses whether users rely on the AI in the right situations, and situation awareness, which gauges how well users understand the AI's current state and reasoning. Standardized frameworks, such as the NIST guidelines for AI usability, provide checklists for designing and testing interactive AI systems.

Empirical studies often involve user experiments with prototypes, comparing different interface designs or explanation styles. For example, research has shown that providing local explanations (why a specific decision was made) can improve user trust more than global explanations (how the model works overall). However, evaluation remains challenging because AI systems are non-deterministic and can behave differently across contexts, requiring longitudinal studies and real-world deployment data.

## Future Directions and Ethical Considerations

The future of human-AI interaction is likely to involve more seamless and proactive systems, such as ambient intelligence that anticipates user needs. Advances in [multi-head-attention](https://www.wikiprompt.org/wiki/multi-head-attention) and [cross-attention](https://www.wikiprompt.org/wiki/cross-attention) may enable models to better track user intent over long conversations. However, these developments raise ethical concerns about privacy, autonomy, and the potential for manipulation. Researchers and policymakers are exploring frameworks for responsible AI, including the concept of human-in-the-loop, where humans retain final authority over critical decisions.

Another emerging area is the interaction with embodied AI, such as robots from [figure-ai](https://www.wikiprompt.org/wiki/figure-ai) and [sanctuary-ai](https://www.wikiprompt.org/wiki/sanctuary-ai), which require physical and social cues beyond text. As AI becomes more integrated into daily life, the field will need to address issues of digital well-being, such as preventing over-reliance on AI for decision-making. The collaboration between academia, industry, and regulatory bodies will be essential to ensure that human-AI interaction remains beneficial and aligned with societal values.

## See Also

- [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence)
- human-computer interaction
- [generative-ai](https://www.wikiprompt.org/wiki/generative-ai)
- [machine-learning](https://www.wikiprompt.org/wiki/machine-learning)

## References

- Amershi, S., et al. (2019). Guidelines for Human-AI Interaction.
- Shneiderman, B. (2020). Human-Centered Artificial Intelligence.
- NIST (2023). Usability of Artificial Intelligence Systems.

---
Source: https://www.wikiprompt.org/wiki/human-ai-interaction
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-14T06:30:28.197934+00:00