Andrej Karpathy revives aerospace language rules as a blueprint for clearer AI output

The controlled language standard ASD-STE100 could help people audit AI-generated instructions by making assumptions, procedures and uncertainty easier to spot.

Andrej Karpathy’s recent comments have helped push an old specialist standard into a new AI context: ASD-STE100, the controlled language originally created for technical documentation in aerospace and defence. The appeal is straightforward. As AI systems make long-form generation cheaper, the harder task for many users is no longer producing text, but checking what it actually says. Concise, tightly controlled language can make that review process less error-prone and less mentally expensive.

According to an overview of ASD-STE100, the standard exists to remove ambiguity by restricting vocabulary and sentence structure in maintenance and technical writing. That design logic fits neatly with one of the main criticisms of large language models: they can be fluent without being precise. In that setting, the value of controlled language is not stylistic. It is operational. Shorter, cleaner phrasing can make assumptions easier to spot, procedures easier to audit, and instructions easier to trust.

Karpathy’s point, as reported by the technology publication BitFan, goes further than simply asking models to write more clearly. He has suggested that users should prompt for output in an ASD-STE100-like style, while recognising that full compliance is difficult and that something like “80% of the way to ASD-STE100” may be more practical. He has also pointed to diagrams, interactive HTML pages and custom explainer videos as alternative formats when the goal is comprehension rather than prose. The underlying argument is that AI output should be treated as an interface, not just a block of text.

That distinction matters because different tasks need different forms of presentation. A procedure may benefit from controlled language. A complex system may be better explained visually. An uncertain claim may need a format that exposes assumptions and dependencies more explicitly. The broader lesson from Karpathy’s comments is that the most effective AI output is not necessarily the most detailed. It is the version that allows a human to inspect it fastest and with the least ambiguity.