Feedforward in Generative AI: Opportunities for a Design Space

Generative AI (GenAI) models have become more capable than ever at augmenting productivity and cognition across diverse contexts. However, a fundamental challenge remains as users struggle to anticipate what AI will generate. As a result, they must engage in excessive turn-taking with the AI’s feedback to clarify their intent, leading to significant cognitive load and time investment. Our goal is to advance the perspective that in order for users to seamlessly leverage the full potential of GenAI systems across various contexts, we must design GenAI systems that not only provide informative feedback but also informative feedforward—designs that tell users what AI will generate before the user submits their prompt. To spark discussion on feedforward in GenAI, we designed diverse instantiations of feedforward across four GenAI applications: conversational UIs, document editors, malleable interfaces, and agent automations, and discussed how these designs can contribute to a more rigorous investigation of a design space and a set of guidelines for feedforward in all GenAI systems.

Introduction. Generative AI (GenAI) models are capable of generating a diverse range of content including stories, videos, code, and more. HCI communities have integrated these models into a diverse range of applications to augment learning and teaching [15, 32], support academic literature review [1, 12, 16, 19], generate and customize software on the fly [6, 17, 24, 36], and automate web tasks [11, 29], showcasing their potential to enhance productivity and cognition on a universal scale.

Despite its potential, users inevitably face challenges in trying to anticipate GenAI’s responses and aligning their prompts to match their needs. Subramonyam et al. have labeled this set of challenges as the gulf of envisioning, which describes the user’s gap in identifying and articulating their desired task, providing effective instructions, and anticipating the model’s response [30]. This gap makes it difficult for novices to formulate effective prompts to AI and understand the cause of misalignment [38]. While researchers have explored various approaches to help users better learn and anticipate AI’s responses with dedicated spaces for experimenting with prompts [3, 4], generating multiple outputs in a single space [13, 31], asking clarifying questions [28], and providing widgets to modify generated outputs posthoc [14, 23, 36], these systems primarily focus on offering feedback—helping users understand what happened after they submit a prompt. While feedback helps users develop an understanding of the system through exploration, users cannot always anticipate the outcome of their prompts. Instead, they must rely on iterative trial-and-error, which can incur significant cognitive load and time investment. If we want users to easily leverage the potential of GenAI systems across diverse contexts, they must be able to anticipate AI outputs without solely relying on feedback. To achieve this, we aim to advance the perspective that GenAI must not only provide informative feedback but also informative feedforward—designs that help users anticipate outcomes before performing an action [37]. Providing such feedforward in GenAI systems can help users understand what to expect from the AI model before submitting their prompt. This can reduce excessive conversational exchanges, bridge the abstraction gap between users and AI [21, 30, 35], and encourage meta-cognition when constructing prompts [33]. The HCI community has extensively explored feedforward design in various domains, such as gestures [5], VR/AR [25], widgets [8, 34], and proactive systems [2]. Furthermore, the community has begun exploring feedforward designs for GenAI, examining trade-offs between the level of detail provided and the cognitive load involved in assessing generated code and images [10, 39]. However, GenAI’s expanding capabilities are shaping an increasingly broad design space of GenAI interfaces. This rapid growth calls for a more systematic approach to designing effective feedforward across diverse domains—uncovering opportunities for novel interaction techniques, assessing user benefits and challenges across design dimensions, and establishing comprehensive design guidelines. This paper aims to spark discussion on how we may achieve these goals, enabling users to effectively engage and interact with all forms of GenAI systems. To lay the groundwork for this discussion, we designed diverse instantiations of feedforward across four GenAI applications: conversational UIs, document editors, malleable interfaces, and agent automations. We then discussed the

Method. This section aims to explore possible ways to provide feedforward in generative AI systems. We view that effective feedforward in generative AI interfaces should enable users to:

• Anticipate what the AI’s response or action may be. • Disambiguate their prompt without excessive exchanges in conversation or interaction. • Directly engage with the feedforward content. • Reflect on their prompting practices with less friction.

To explore various ways to design feedforward in GenAI systems, we first explore feedforward in an LLM-powered conversational interface. We then describe scenarios for three other designs of feedforward in GenAI applications.

2.1 Feedforward in Conversational LLM UIs 2.1.1 Feedforward Components. Research has shown that users benefit from multiple representations when interacting with LLMs [14, 32]. Therefore, we developed feedforward components that each aim to provide a specific representation that would help users anticipate the LLM’s response. As the user types a prompt, the conversational UI generates feedforward content inside two feedforward components: 1) a key topic phrase alongside a high-level outline of the anticipated response (Fig. 1.1a) and 2) a minimap view with paragraph blocks (Fig. 1.1b). These representations help users anticipate the topics and subtopics the LLM will generate, as well as whether its response might be too short or too long.

2.1.2 Manipulating Feedforward Content. We enable users to manipulate feedforward content in two ways. First, when users edit their prompt, the feedforward content updates to reflect the changes. This allows users to identify misalignments between their intent and the feedforward content, encouraging them to add details to the prompt before submitting it. For instance, they can clarify an ambiguous term (Fig. 1.2) or request a shorter response (Fig. 1.3). These instant updates to the feedforward components may also spark reflection, such as realizing that it may be useful to understand the advantages and limitations of Wizard of Oz Prototyping rather than just its definition.

Discussion. We invite researchers in this community to critically examine past and present GenAI interfaces through the lens of feedforward, draw from feedforward literature in other domains, explore new interaction techniques and frameworks for delivering informative feedforward, and work toward concrete design and implementation strategies applicable to all GenAI systems.