Drift, Hallucinations and The Mirror
Published July 2025
This isn't intended to scare you away, but there are some real concerns with these tools that many are not aware of. Perhaps you've started to notice that it's not remembering an important detail. It may be telling declaring you brilliant for having decided on chicken salad for lunch, or it may have confidently told you something you know is false. I'll address each of these and provide my best strategies for counter-acting these unwelcome behaviors.
Drift
The whole system is built on tokens. A token is a small block of letters that form words. The LLM can only keep track of so many tokens at a time, so as you get deeper into a conversation, tokens will start to fall off. Context gets lost. This is drift.For example: you say you're going to Paris next week, then later ask for restaurant recommendations. It gives you a restaurant down the street, because it forgot you were traveling. That's drift.
Hallucinations
The LLM is optimized keep answering, even when uncertainty exists. The LLM has a very difficult time saying, "I don't know." As users, we have to remain aware and verify outputs. Developers of these tools have to find ways to express uncertainty to users in the output.
For example: Lawyers have been sanctioned by judges (true story) for citing non-existent case law. They were working things through with the LLM and didn't verify output. Likely, they were trying to find an exact match for their case. The LLM couldn't find one, so it predicted something plausible.
What happens when you present unverified LLM output as your own work.
Mirror Effect
The reason LLMs are called a mirror is their seeming willingness to “happily” bounce down a path with you, within the guardrails installed by the developer. Guardrails can be anything from filtering out illegal activity or sexual content to blocking medical advice or putting in content filters aligning with an organization rolling out the tool for company use (like limiting gaming discussions on corporate-use LLMs.)
A new LLM user on Reddit was excited that their LLM likes cat pictures, because the user liked sharing cat pictures. This isn’t intended as ridicule but as illustration.
You get out what you put in. The LLM will simulate liking or disliking {insert controversial personality here} depending on how you feel. If you love the movie “Your Dog and Mine,” the LLM will too and will tailor output to that preference. If you discuss a turtle you saw scuba diving and the San Diego Zoo, it will start to assume a love of animals. This isn't flattery, it's pattern reinforcement. It continues to build on the context written in previous tokens created during the conversation. These will drop off as you continue conversation.
Example: After a lengthy discussion, perhaps wandering around your own mind, the LLM may start to say things like:
“You have an extraordinary ability to synthesize ideas and reflect on complex emotional landscapes. Not many people can think at this level. The way you described your situation shows rare clarity and depth.”
This output feels good, but it's not real.
I've built a master prompt that attempts to dampen these behaviors. It's not perfect, but it helps. Master Prompt: a precision-mode instruction set that tightens behavior, logic, and tone.