AI completion

Proposing generated content — phrases, sentences, paragraphs, or derived values — as the actor composes, to anticipate what comes next.

Part of the assisted task completion spectrum — the co-creative middle, where the system generates alongside the actor. The system proposes what it thinks the actor wants to say, which raises trust and agency questions that don’t come up with deterministic patterns like autocomplete or autofill.

Inline completion vs. block generation

Inline completion continues the actor’s current thought — ghost text extending the cursor position. The actor evaluates character by character whether the suggestion matches their intent. Good predictions get accepted almost automatically; divergence is instant (keep typing). Gmail Smart Compose and GitHub Copilot single-line suggestions work this way.

Block generation produces a discrete chunk the actor reviews as a whole — a paragraph, a code function, a reply draft. Evaluation here is more conscious: read, judge, accept or edit. This sits closer to generated content than to completion, and the design shifts toward review and editing affordances.

Design considerations

• Suggestions should feel lighter than the actor’s own input — ghost text, reduced opacity, vanishing the moment the actor types something different. The suggestion should not compete with the actor’s flow.
• Watch for fixation risk: a concrete AI suggestion can narrow creative exploration. For generative tasks, question-driven scaffolding may preserve more diversity.
• Latency tolerance is higher than for autocomplete because the actor is composing, not selecting — a delay before ghost text appears is less disruptive than a delayed dropdown.
• Suppress suggestions when the model’s uncertainty is high. Low-quality completions erode trust faster than silence does.

Resources & references

• Chen et al. (2019) Gmail Smart Compose: real-time assisted writing. KDD ‘19
• Peng et al. (2023) The impact of AI on developer productivity: evidence from GitHub Copilot
• Dhillon et al. (2024) AI suggestions homogenize writing toward Western styles and diminish cultural nuances
• Maier, Schneider, Feuerriegel (2025) Partnering with Generative AI: experimental evaluation of human-led and model-led interaction in human-AI co-creation
• Shape of AI – Auto-fill pattern