Programming has always been the art of telling computers precisely what to do — in their language. But what if we could write software entirely in plain English? This question has captivated researchers, engineers, and visionaries for decades. With the rapid rise of large language models (LLMs) like GPT-4 and Claude, the dream is closer to reality than ever before.
Historical Background: English-Like Languages Have Always Existed
COBOL (1959) was designed to read like English business prose. SQL reads almost like a natural-language query — "SELECT all employees WHERE salary > 50000". Inform 7 lets developers write interactive fiction using full English sentences. Yet all of these still require rigid, structured thinking — they are English-flavored, not truly natural English.
Three Main Technical Approaches
1. LLM as Compiler (Vibe Coding)
You write intent in English, and an LLM translates it into executable code. Tools like GitHub Copilot, Cursor, and Claude Code already demonstrate this. The human writes a "specification" in English; the AI produces working code.
"Create a function that takes a list of integers, filters out negatives, squares each, and returns sorted descending"
AI generates:
def process(nums):
return sorted([x**2 for x in nums if x >= 0], reverse=True)2. Ontology-Based NLP
Natural Language Programming uses a formal ontology so that English sentences compile unambiguously into structured logic. The limitation: the ontology must be pre-built for each domain.
3. Probabilistic Specification
The system uses machine learning to interpret intent probabilistically, then verifies the result via tests. The developer writes acceptance criteria in English; the AI generates code that passes those criteria.
Current Maturity Comparison
| Approach | Maturity | Best Use Case | Main Limitation |
|---|---|---|---|
| LLM-as-Compiler | ★★★★☆ | Prototyping, scripts, CRUD apps | Hallucinations, formal verification |
| Ontology-Based NLP | ★★★☆☆ | Domain-specific DSLs | Requires expert ontology design |
| Probabilistic Spec | ★★☆☆☆ | Test-driven AI development | Incomplete test coverage |
| Hybrid Systems | ★★★★★ | Enterprise applications | High implementation complexity |
Adoption Forecast: 2025–2035
Projected Milestones
Key Challenges That Remain
- Ambiguity: Natural language is inherently ambiguous. "Sort users by name and date" — ascending or descending? Which field takes priority?
- Formal Verification: How do you mathematically prove that AI-generated code matches the English specification?
- Performance: High-performance, low-level code (embedded systems, OS kernels) requires precise control that natural language cannot easily express.
- Security: AI-generated code can introduce subtle vulnerabilities that are difficult to detect without thorough code review.
- Debugging: When English-specified code fails, which part is wrong — the English, the AI interpretation, or the generated code?
Conclusion
English will not fully replace programming languages in the near term — but it is rapidly becoming the primary interface for software development. The shift is already underway: developers today write far less boilerplate and far more intent-driven specifications. As AI models improve, natural language will become the dominant layer through which humans express software requirements. What changes is not the goal (working software), but the abstraction layer we use to define it.