AI bias against niche stacks

Copilot-style tools feel omniscient until you point them at a legacy module written in a stack the internet barely indexed. The model confidently invents APIs, misreads vendor-specific semantics, and produces diffs that compile in the demo repo and detonate in yours. That is not a temporary bug; it is a bias baked into training data toward popular languages and greenfield patterns.

AI rewrite pilots — the 2023–2025 enterprise fad — sell autonomy on slides and deliver autocomplete in practice. Teams feed samples, get impressive translations, and discover production needs tests nobody has, edge cases nobody documented, and compliance paths the model never saw. The failure mode matches big-bang rewrites because the organizational work was skipped, not because the tokenizer was too small.

Low-code and LLM-augmented citizen dev share a darker symmetry: they accelerate the easy 80% and hide the dangerous 20% — integrations, idempotency, PII, concurrency — until governance panics and senior engineers get hired to untangle graphs and diffs they did not author. AI did not remove engineering; it moved it to cleanup at a premium.

Bias toward niche stacks hurts maintainers of the boring systems that run banks and logistics. The median GitHub repo is a JavaScript todo app, not a J2EE integration from 2006. Models optimize for plausible continuation, not verified behavior — dangerous when "plausible" means "looks like Spring" but your container is WebLogic with proprietary hooks.

Augmentation works when humans remain accountable: small slices, parity tests, code review with domain experts, strangler migrations. Anything marketed as "autonomous modernization" without those guardrails is procurement theater — useful for demos, lethal for prod if leadership believes their own Q3 timeline.

Vector databases and RAG do not fully fix this; they retrieve snippets, not ownership. If your institutional knowledge lives in tickets and tribal memory, retrieval just gives the model better fiction prompts. The fix is boring documentation and incremental migration — the same fix as before AI, except now you also pay token bills.

The fad will mutate into sensible tooling: assistants inside IDEs, test generators, migration helpers with human gates. It will fail as engineer replacement for the same reason low-code did — exceptions are code, and prod pages do not care about keynote adjectives. Bias against niche stacks is a reminder that fashion-forward AI chases median repos; your business might still run on the long tail.