Three Reasons AI Is Now More Reliable Than Ever

TL;DR

WSJ column explains three reasons AI models, despite hallucinations, now handle real work effectively.
Clues came from a recent leak of Anthropic's Claude Code source code.
Engineers built systems around prone-to-error models to boost practical reliability.[[1]](https://www.wsj.com/tech/ai/ai-model-reliability-hallucinations-a3bc0497)

The story at a glance

Christopher Mims argues in the Wall Street Journal that AI models still hallucinate and give inconsistent answers by nature, yet engineers have made them useful for actual tasks. The piece draws on a recent unintentional leak of source code for Anthropic's Claude Code, a popular AI coding agent. This is timely after the late-March 2026 leak exposed how companies like Anthropic wrap safeguards around base models.[[1]](https://www.wsj.com/tech/ai/ai-model-reliability-hallucinations-a3bc0497)[[2]](https://www.wsj.com/tech/ai/anthropic-races-to-contain-leak-of-code-behind-claude-ai-agent-4bc5acc7)

Key points

AI models inherently produce false information (hallucinations) and vary in responses, but recent advances make them suitable for productivity.[[1]](https://www.wsj.com/tech/ai/ai-model-reliability-hallucinations-a3bc0497)
A software leak revealed inner workings of frontier models from firms like Anthropic, Google, and OpenAI—trade secrets worth billions.[[1]](https://www.wsj.com/tech/ai/ai-model-reliability-hallucinations-a3bc0497)
Anthropic's Claude Code, leaked in late March 2026 via an npm packaging error, is a "crown jewel" that helped the firm edge toward profitability ahead of OpenAI.[[3]](https://johnlothiannews.com/how-a-233-year-old-wall-street-institution-went-all-in-on-crypto)[[2]](https://www.wsj.com/tech/ai/anthropic-races-to-contain-leak-of-code-behind-claude-ai-agent-4bc5acc7)
The three reasons (detailed in paywalled article) likely involve engineering techniques shown in the leaked code, such as error-checking loops and orchestration to mitigate model flaws.[[4]](https://www.reddit.com/r/BetterOffline/comments/1sarsx4/yall_gotta_read_this_engineer_eviscerating_the)
Leak exposed about 500,000 lines of code across 1,900 files, but no model weights, customer data, or core AI math.[[5]](https://www.axios.com/2026/03/31/anthropic-leaked-source-code-ai)

Details and context

The article focuses on practical improvements in AI reliability amid hype over superintelligence. Base large language models remain probabilistic and error-prone, but layered systems—like those in Claude Code—use retries, linters, sub-agents, and tool routing to produce dependable outputs for coding and other work.[[4]](https://www.reddit.com/r/BetterOffline/comments/1sarsx4/yall_gotta_read_this_engineer_eviscerating_the)[[6]](https://www.linkedin.com/pulse/claude-code-leak-your-healthcare-ai-agent-strategy-bill-russell-jaele)

Anthropic's leak stemmed from human error in a release process, bundling a source map file publicly; the firm issued takedowns for over 8,000 GitHub copies and added safeguards.[[2]](https://www.wsj.com/tech/ai/anthropic-races-to-contain-leak-of-code-behind-claude-ai-agent-4bc5acc7)

This fits broader trends: companies prioritize "agentic" wrappers over raw model power for enterprise use, explaining why specialized AI tools cut head count in tech firms.

Key quotes

"A funny thing happened on the way to AI superintelligence: Engineers made AI suitable enough to help humans get real work done. Clues to how they achieved it emerged in a recent software leak."[[1]](https://www.wsj.com/tech/ai/ai-model-reliability-hallucinations-a3bc0497)

Why it matters

AI reliability gains expand tools from toys to workplace staples, affecting jobs and efficiency across industries. For businesses and developers, it means safer adoption of agents like Claude Code for coding, with less risk of bad outputs if wrapped properly. Watch Anthropic's profitability push and similar leaks from rivals, which could reveal more on scaling reliable AI amid rapid competition.[[3]](https://johnlothiannews.com/how-a-233-year-old-wall-street-institution-went-all-in-on-crypto)