# Maintaining Critical Thinking When Working with AI Understanding AI's limitations and proper usage strategies - **Source:** [Bruno Rocha](https://brifing.si/source/00ca3dfa-d46d-4465-8860-e8ac51e72676) - **Published:** 2026-08-09 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/44dca16c-1b7b-4b1a-8787-e124305f2a43 ## The Danger of AI Psychosis The term 'AI Psychosis' describes individuals who blindly believe everything AI says, leading to strong convictions in completely unfounded concepts. In software engineering, we frequently encounter long, confident AI-generated explanations that upon closer inspection prove nonsensical. One example is pull request descriptions that appear legitimate at first glance but are actually meaningless word salads. This phenomenon poses serious risks for those using the technology without understanding how it works. ## The Concept of Cognitive Surrender The terms 'cognitive surrender' or 'meat proxy' better describe this phenomenon than 'psychosis.' The issue isn't that people are going crazy, but rather failing to apply critical thinking to AI outputs, often due to not understanding that AI can (and will) constantly make mistakes. When someone merely proxies AI's statements, they add no real value. In an era where anyone can prompt AI, your ability to critically analyze outputs becomes your differentiator. Recognizing AI's limitations is crucial when working with the technology. ## The AI-Assisted Development Paradigm The correct approach is AI-assisted, not AI-independent development. Even the newest models suffer from fundamental issues like jumping to conclusions, forgetting instructions, and hallucinating information. These problems are inherent to LLM technology and will persist until the underlying architecture changes. Claims about solving these issues are often just marketing - remember when OpenAI called GPT-2 'too dangerous' only for it to prove barely coherent beyond a single prompt? ## Supervision Required for Complex Tasks Constant supervision is essential when AI handles complex prompts requiring extensive research. When the AI starts making strange assumptions, immediate intervention is needed to guide it back on track. While unsupervised operation might work for simple requests, this approach rarely succeeds with complex tasks. Regardless of how many details you provide, the AI will almost certainly make wrong assumptions at some point and start building incorrectly. ## Understanding LLM Limitations The best way to deeply understand why LLMs have these problems is to build one yourself. Educational resources like Andrej Karpathy's tutorials on LLM architecture provide invaluable insights into how these models work and their fundamental limitations. These materials offer clear explanations about the technology's inner workings that every serious AI practitioner should understand. ## Developing a Habit of Challenging AI Cultivating the habit of challenging and cross-verifying all AI outputs is crucial. What began as manually asking 'how do you know this?' or 'provide proof' can now be automated with stop hooks that prompt the AI to substantiate its claims. System prompts alone aren't reliable for this purpose because one of the very issues we're combating is the AI's tendency to sometimes ignore instructions entirely. ## The Verification Process Always verify AI-provided evidence rather than accepting it blindly. For instance, when the AI cites a source, actually open the link to confirm the information exists (and isn't itself AI-generated). There are numerous cases where the 'proof' turns out to be completely fabricated by the AI. This verification process should continue until the evidence becomes indisputable. ## Code Quality Preservation Strategies Even as newer models improve, maintaining the practice of line-by-line code review remains essential for preserving coding knowledge. While following previous guidelines may produce working code, that doesn't mean it's optimally structured. Current models typically generate overly complex code filled with duplication, poor separation of concerns, and incorrect edge case handling - essentially all the mistakes a beginner programmer might make. ## Sources covering this story - https://swiftrocks.com/using-ai-without-losing-critical-thinking — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._