WPT details: Web platform tests will cover the WebIDL interface, availability() states, parameter and schema validation, AbortSignal cancellation, and return type structure. While specific semantic decisions and probability outputs depend on the underlying model, interface compliance and error handling are fully testable.
Provides high-speed, deterministic semantic decision-making directly on the user's device as a built-in AI capability. Unlike open-ended generative language models (such as the Prompt API) that decode text token-by-token over hundreds or thousands of milliseconds, the Decisions API (window.DecisionModel) evaluates natural language input text against a structured set of questions and candidate options in a single forward pass using logit scoring. This delivers micro-latency evaluations (tens or low hundreds of milliseconds on typical consumer laptop CPUs and GPUs) while outputting calibrated confidence values. This brings non-autoregressive "System 1" decision models (e.g. Jev, Laya, Kev, Open-Jev) natively into the web platform, providing developers with a pragmatic, zero-marginal-cost "Semantic If" primitive for real-time client-side routing, interactive input guardrails, and adaptive user interfaces.
Web applications and agentic workflows increasingly need to make fast, structured decisions over unstructured input and page state: Which search filters match this natural-language query? Which UI action or tool fulfills the user's goal? Would this draft trigger a moderation review or miss key details before submission? What does this page section or form field represent? [1] Without a built-in decision primitive, developers are forced to choose between several suboptimal alternatives [2]: - Brittle Heuristics: Keyword rules and regular expressions that are fast and local, but break on phrasing variations, typos, and multilingual input. - Bespoke Model Selection & Training Tax: For most web products, curating training data, evaluating models, integrating runtimes, and bundling multi-megabyte model weights is a prohibitive operational hurdle. Users also face downsides of duplicating network and storage costs, and unmanaged resource contention across origins. - Cloud AI Endpoints: Capable, but introduces hundreds of milliseconds network round trips, triggers privacy tradeoffs by transmitting user drafts off-device, and incur recurring or high-frequency cloud / server costs. - Generative LLMs (LanguageModel): Autoregressively generating structured JSON via client-side LLMs is 10–50x heavier than necessary, consumes significant battery/RAM, and lacks calibrated option probabilities. Recent advancements demonstrate that compact, non-autoregressive decision models can evaluate candidate choices in tens of milliseconds with calibrated confidence distributions. We believe Built-in AI has a natural role to play: by offering a platform-level decision engine, we can make fast semantic branching an ubiquitous primitive. Web applications gain instant access to binary verification, categorical routing, and ordinal scoring tasks with amortized overheads, on-device privacy, and efficient hardware utilization. This prototype phase aims to confirm whether the community's enthusiasm translates into concrete web production use cases and architectural viability. [1]: https://github.com/explainers-by-googlers/decisions-api#use-cases [2]: https://github.com/explainers-by-googlers/decisions-api#considered-alternatives
Samples: https://web-ai.studio/playgrounds/decisions
Explainers: https://github.com/explainers-by-googlers/decisions-api