AI coding is gambling

Is AI Coding “Gambling”?

• One camp says yes: outputs are non-deterministic, users repeatedly “pull the lever” with new prompts hoping for a good result, and success can feel like luck.
• Others argue it’s only “gambling” in the trivial sense that all uncertain work is; if success rates are high and properly verified, it’s just engineering under uncertainty.
• Some distinguish “spray-and-pray” one-shot prompting (like slots) from spec-driven, test-driven workflows (more like poker or regular work).

Human Analogies: Interns, Coworkers, Managers

• Comparisons are made to assigning tasks to interns, temps, or mediocre coworkers: you can’t predict quality, and you must review and iterate.
• Critics push back: humans can be trained, build lasting knowledge, are accountable, and have slower, less-addictive feedback loops.
• There’s discomfort with using interns as a metaphor at all; some find it dehumanizing.

Addiction, Variable Rewards, and Psychology

• Multiple commenters report slot-machine–like behavior: rapid re-prompts, running multiple agents in parallel, staying up late “chasing” a working run.
• Variable rewards and near-misses are seen as key hooks, similar to gambling and social media.
• Some explicitly describe disrupted work–life boundaries and dopamine-driven overuse.

Productivity vs Reliability and Maintainability

• Many say agents dramatically speed up boilerplate, translation, and prototypes, sometimes enabling projects they’d never finish alone.
• Others find that generated code often “looks right” but is brittle, hard to maintain, or subtly wrong, especially without strong tests.
• Concerns include long-term maintainability, code bloat, and shallow understanding of complex systems.

Specs, Tests, and Workflows

• A recurring theme: AI coding only works well when paired with clear specs, strong automated tests, and scripted quality checks.
• Some teams reportedly forbid manual coding and enforce spec/TDD + agents, claiming large productivity gains but unhappy developers.
• Others note that in real-world product work, specs are rarely clean or stable, so the “just conform to the spec” story feels unrealistic.

Shifts in Roles and the Meaning of Programming

• There’s sharp disagreement over whether prompting + reviewing counts as “programming” versus “managing a coder.”
• Some relish focusing on ideas and product design while offloading typing to AI; others value the craft of writing code itself and avoid AI to preserve skill and joy.
• Worries appear about training pipelines for juniors and a future where many devs can orchestrate agents but few deeply understand systems.

Control, Ownership, and Industry Dynamics

• Some fear dependence on proprietary LLM providers, subscription lock-in, and future price hikes.
• Others see current subsidies and promotions as analogous to casinos’ free chips: the “house” ultimately wins.
• Broader ethical concerns include damage to online trust and “looting the commons” of public code and text to train models.