AI and the Abandonment of Design: The Empty State in Question

An alarming observation: the empty state and user abandonment

In the context of a Chrome extension co-developed by the author, a concerning phenomenon is emerging: approximately 70% of new installations are never reused. This alarming figure is primarily attributed to the empty state of the interfaces, which fails to engage and guide users from their very first interaction. Experience shows that the version of the extension that performed the worst was the one that adhered most closely to current standards for AI products, often characterized by a minimalist design.

Historically, the empty state—this crucial first impression of software—was meticulously attended to by designers. Elements such as pre-filled templates, sample data, guided tours, or explanatory videos were integrated to orient the user. However, with the advent of AI products, this state has been reduced to its simplest form: an empty text field, often accompanied by a generic message like "Ask me anything." This approach, presented as minimalism, is in reality a lack of design.

The historical importance of the empty state in software design

The empty state has long been considered one of the most crucial elements in the design of a software product. Companies like Linear, Notion, Figma, Airtable, Slack, Trello, Intercom, and Asana have invested years in developing their first-use experiences. According to Don Norman, in his book The Design of Everyday Things, good design communicates its function through signifiers: visual cues that help the user understand how to use a product. An interface devoid of these signifiers is, by definition, defective.

The Nielsen Norman Group has clearly established that for an empty state to be effective, it must explain what is supposed to appear there, show how to make those elements appear, and offer an initial action to the user. These elements are not mere decorations but essential tools to ensure that the user returns after their first interaction.

The evolution of AI products towards minimalist design

For the past three years, every major AI product has adopted an empty state reduced to a centered text field, accompanied by a few prompt suggestions. Tools like ChatGPT, Claude, Gemini, Copilot, Cursor, Perplexity, Grok, and Le Chat de Mistral all follow this model. This uniformity is not due to a technical constraint but to a convention that solidified in 2023, when the first language model-based products were launched without a genuine design phase.

Bret Victor had already warned against this type of failure in Magic Ink, emphasizing that treating interactivity as the main surface of an information product forces the user to relearn with each new interaction. An empty prompt box requires the user to start each session by inventing a question, which can be discouraging.

The consequences of a high abandonment rate

The 70% abandonment rate during the first session is telling. Users who installed the extension had shown clear interest by searching for the product, reading its description, and granting the necessary permissions. However, the first session confronted them with a cognitive challenge rather than a simple and intuitive action.

A/B testing has shown that versions of the extension that presented a concrete example or a sample result drawn from user data significantly outperformed those that merely featured an empty field with a generic instruction. Kathy Sierra described this gap as the Suck Threshold: the distance between installation and competence. Reducing this threshold is crucial for retaining users.

The loss of vocabulary in AI interface design

By adopting the prompt box, AI products have abandoned decades of research in human-computer interaction (HCI) on how to transform a novice into a competent user. Concepts such as signifiers, affordances, recognition rather than recall, and progressive disclosure have been set aside.

Jakob Nielsen's usability heuristics, particularly "recognition rather than recall," emphasize that users should not have to remember information to navigate an interface. An empty prompt box violates this principle by forcing the user to recall what the product can do before they have even experienced it.

Imagining a truly designed empty state for AI

A well-designed empty state for an AI product should include four key elements: a worked example, a starting verb, clearly stated limits, and the ability to act before speaking. These elements are technically feasible but have not been widely adopted because the prompt box remains the least costly option to develop.

A worked example replaces the blinking cursor by showing a response to a representative question on the first screen. According to Bret Victor's framework in Magic Ink, information software should show rather than interact. A worked example is a visual demonstration, not an interaction.

A starting verb, such as "Create your first issue" in Linear or "Type slash for commands" in Notion, guides the user towards a concrete action. "Ask me anything" is the antithesis of a verb; it represents a lack of clear direction.

Stating the model's limits from the outset replaces silent failure. According to Norman's feedback principle, the system must clearly indicate its capabilities and limitations. Current AI products often conceal this information, forcing the user to discover the limits through trial and error.

Finally, allowing action before speech fosters intuitive discovery. The moment a user accidentally realizes something useful is crucial, and a well-designed empty state should facilitate this experience without waiting for the user to ask the right question.