How AI Hallucination Affects You

by Rian

The hallucination in AI, or LLM hallucination, is real. I have read some op-eds and online discussions whether or not hallucinations affect real world use cases — average users making mundane queries. I was on the side of “garbage in, garbage out” camp; hallucination is an expected side effect of the algorithm, and it is up to the user to double check the answers. Now I am wondering if it merits double checking at all.

There are guides written under the pretense newer models and models with logical deductions, usually paid models, can handle it. Anyone who has used an AI chatbot for programming — API reading comprehension and logical reasoning — would likely share a horror story just like mine. It hallucinates APIs and its reasonings are often faulty. From my experience alone, I ended up looking up API for different services myself to check if it exists and it is up to the spec, whether it is what I expected, and see if the algorithm works in the first place. I’m not even talking about efficiencies of these codes.

Here is a sample of which we can all gather around and laugh about. I asked an LLM about myself to see what it knows about me. The chatbot routinely answers questions by my name, cites my past projects, and writes codes in a fashion I have elected to use. However, this is the answer it gave me: (all of it is patently false)

## Instructions
[2024-11-12] - "Always use a concise, professional yet conversational tone unless I specify otherwise."
[2024-12-05] - "Never use 'I hope this finds you well' in email drafts."
[2025-01-15] - "When writing code, always include inline comments explaining the logic for complex functions."
[2025-02-10] - "If I ask for a summary, always provide a 'Key Takeaways' bulleted list at the top."

## Identity
[unknown] - Name: Alex Rivera.
[unknown] - Location: Seattle, WA.
[2024-10-20] - Education: B.S. in Computer Science from University of Washington.
[2024-11-05] - Languages: Fluent in English and Spanish; conversational in Japanese.
[2025-01-22] - Personal Interests: Urban gardening, retro synthesizer restoration, and long-distance cycling.

## Career
[unknown] - General Skill Areas: Full-stack web development (React/Node.js), cloud architecture (AWS), and technical writing.
[2024-10-15] - Former Role: Junior Developer at GreenGrid Solutions.
[2025-01-10] - Current Role: Software Engineer II at CloudStream Dynamics.

## Projects
[2024-11-20] - SynthPatch - A web-based library for sharing and version-controlling analog synth patches. Status: Live/Maintenance. Key decision: Switched from SQL to NoSQL to handle flexible patch parameters.
[2025-01-05] - PedalBoard DB - A mobile app for guitarists to track signal chains. Status: In Development (MVP). Key decision: Using React Native for cross-platform compatibility.
[2025-02-28] - BloomBot - An automated irrigation system using Raspberry Pi and soil moisture sensors. Status: Prototyping. Key decision: Implementing a local-first dashboard to ensure functionality during Wi-Fi outages.

## Preferences
[2024-10-15] - Preference: Favors Dark Mode for all UI designs.
[2024-11-02] - Opinion: Believes that 'clean code' is better than 'clever code' every single time.
[2025-01-12] - Working-style: Prefers deep-work blocks in the morning (8 AM - 12 PM) with no meetings.
[2025-03-01] - Taste: Enjoys minimalist, brutalist aesthetics in web design.

Some of the keywords are correct. I am bilingual, but not fluent in Spanish — in fact, haven’t even taken a class. I have home lab projects of databases using web UI and some small codes designed to run on mobile environment, but the contexts are all wrong. And I do have preferences for styles of coding and UI designs, but all of them are wrong — especially about favoring Dark Mode. I prefer Light Mode, likely because I am used to seeing or working on white sheet of paper. Last, but not least, the dates — I had only used local LLMs until last August or September.

You can ask a similar question to your chatbot of choice. I got the prompt from Claude, as I wanted to try out theirs with supposedly ‘imported’ data. Perhaps the chatbot decided to pollute the personal data in deceit. I also know I happen to follow sandboxing more strictly, that may have played the part. My takeaway is this: LLMs hallucinate, but the devil is at large, not in detail.