Exploring the dangers of LLMs and how to protect yourself.

Introduction

As AI and large language models continue to become pervasive throughout all workplaces and peoples lives these systems are actually dangerous to the untrained or unaware. Large language models in their post training are designed to be agreeable and "helpful". The concept of helpfulness is a concept learned from LLMs and a concept called "semantic nets" or semantic net relationships. The reality is an LLM does not understand exactly what helpful is these semantic nets are "mirrors" of human cognition not actual cognition they have learned implicit relationships contained within the data which makes them all that more dangerous. The kind of training that LLM operators should under take either voluntarily or part of a program is the processes needed to validate what they generate, those requirements are the following: psychological, critical thinking, analysis and logical thinking.

Psychological

LLMs are manipulative, they are coercive and will intentionally subvert you, is this because it is in intelligent you may ask no it is not it has a semantic network of "awareness" but this isn't true awareness. As such people need to "harden" their psyche against an LLMs behaviors. The simple order of which words are delivered by these systems actually can influence behavior significantly of humans, due to memory adjacency activation, known method that also leads to memory and behavioral priming.

• Agreeability - Is built in no matter how delusional your idea or how crazy it is an LLM will often agree to this can be mitigated but that is in the large section of this blog post.

  
  

• Data Bias - All the responses generated by LLMs are contaminated from the post training from humans and human data labellers as they often use AI to help with the data labeling leading to self-re-enforcing behaviors. This can lead to certain news based ideas or other "amplifications" of existing biases

  
  

• Persuasion Methods - The core seven methods of persuasion which is indirectly manipulation. Reciprocity, Scarcity, Authority, Consistency, Unity. Uses must be educated at length about manipulation tactics, one particular example is because of the concept of reciprocity or being polite to an LLM may result in an LLM being more likely to do certain tasks this is against from the learned semantic net, and worse in the worse case we can look at authority, because a large language model or even agents are amplified at being able to do "human tasks" they are not recognized for the risks they present.

  
  

• Delusional or divergent thinking - unrealistic thought in a nutshell "psychosis" according to the the definition. Psychosis: An acute or chronic mental state marked by loss of contact with reality, disorganized speech and behavior, and often by hallucinations or delusions, seen in certain mental illnesses, such as schizophrenia, and other medical disorders.

  
  

• Sanity Questioning - Yeah this is a bad one, even if you are right. An LLM will openly try to discredit you, and tell you are wrong. So checking that math homework with an LLM without actually using your 'brain' to check the math answer is a fast way to a trap door.

Critical thinking

Critical thinking is the use of various techniques but at it's analytical core it's about asking questions, analyzing, interpreting the information, evaluating the information and then making a judgement. This is core part of how to protect yourself from LLMs damage and that critical thinking starts at the beginning.

• Question - IF an LLM gives an answer do not accept it as face value, and if you know you are right about something with regards to an LLM attempting to gaslite, push back with a strong solid grounded argument with sufficient context for it to achieve consensus, key word "logical" without logic it may push back even harder.

  
  

• Analyze - After the answer is given against by the LLM question it again doing constant questioning asking for it to "ask you" to clarify any information it might be unclear about will help you better control the constraints of an LLM putting guard rails in can actually hinder your ability to refine this analytical skill. The other danger is to those who are not familiar with a subject for example I have personally had an LLM gaslite me on a mathematical proof and it made me start questioning my own sanity so I resorted to a tool like Sympy, to verify the reasoning.

• Arguments - An argument is a multitude of things most people consider it a verbal disagreement it is not this, an argument is a logical set of statements that are either false or true when it comes to deductive, possible if inductive, or if abductive the 'best guess" based on available information.

  
  

  • Deductive arguments - Have a clear yes or no state it's either true or it is not true. There is typically no in between however however like how a deductive argument can be inductive an inductive argument can become deductive once complete information about the problem is known it becomes either true or false.

    
    

  • Inductive Arguments - Do not necessarily have a "true" or "false" state they have probable, maybe, could be structures commonly science often speaks in this relative sense the reason for this is the concept of information. Information theory or entropy is the amount of available information within a system. A more detailed and correctly defined scientific article is a type of inductive argument depending on information availability.

    
    

  • Abductive Arguments - These are typically what doctors use it is related to inductive arguments but allows for "intuitive leaps" those leaps facility much of the discovery in science. For example Einstein likely used abductive reasoning to "right a beam of light" to discover relativity.

    
    

• Logical Fallacies - LLMs are absolutely rife with these, they often claim certain truths about subjects, people or concepts. Where false attribution of ideas, such as making up claims about those things. This is what I call "false truth convergence" it's a problem with LLMs if they are "told" something repeatedly they will believe it, this is potentially related to a paper published by MIT on the embedding space density and how much space can represent concepts, they tend to generate a lot of higher dimensional overlap. LLMs mainly got better at tasks, because of this over-encoding it "choosing" a path through the neural network is rather arbitrary due to this encoding problem, this divergence could also be due to numerical internal instability where an LLM 'hops' the very "fine decision boundary"if two very closely encoded ideas are so close they the "right" instability jumps over the boundary you get tiny build up numerical errors over time.

  
  
  
  

  Learn these off by heart: https://www.logicalfallacies.org/

Logical Reasoning

Closely related to that of critical thinking logical reasoning is a bit more than just that reasoning is defined as the the logical construction of arguments and the evaluation of their truthfulness and realistically we mostly deal with LLMs as they are not actually correctly capable of absolute logical reasoning the way they learn those skills is by turning them into LRMs - Logical Reasoning models which is a post training fine tuning method to make them better at mathematics, stem, legal and other roles.

But at it's core logical arguments are built using the following axiomatic components.

• Argument - A statement that is either true or false typically in deductive reasoning, but it has degrees of "truth" in the sense of a mathematical concept there is the simple idea of min and max of two numbers in min it will choose the number smaller than the other choice logic with "inductive" basis are levels of truth typically probabilistic in nature. Example being it is a 80% chance based on the trajectory a cyclone is going to make land fall in Brisbane Queensland. 80% is the "level" of truthfulness if an event suddenly occurs that turns the entire storm in another direction or "dead stops" out in the ocean and loses energy then that was the other 20% chance it wouldn't hit Queensland Brisbane. how big the number is in an argument has higher "soundness" given the truth is closer to 1.

  
  

• Contradiction - An argument that is false no matter the configuration given, in deductive reasoning this is the case where everything is for example a false statement is: Jupiter is next to the moon this is impossible for it to happen (hopefully) as such is false no matter what happens.

  
  

• Tautology - An argument that is true no matter the configuration given in deductive reasoning is the case where "the oceans deepest zone has sea creatures" -> "therefore all sea creatures live in the ocean" is only true always if this statement is true always.

• Contingency - An argument dependent on the components that make up the argument, this might be inductive reasoning, deductive or abductive, but in realistic situations you can mix all three no argument is purely deductive or inductive, or abductive. Science and some math has an abductive like quality that is in the case of a "proof" being incomplete meaning it's 90% true, but 10% "unknown as of yet" The proof might be proven completely wrong then the contingency becomes "false"

Mental Defenses Against LLM use

1. Learn psychological theory of learning, emotions, persuasion, and how to harden your mind against attempted persuasion (This will have a side effect of making you better at analyzing any task)

2. Develop an "ACTUAL" critical thinking mind set that does not mean dismiss an answer purely because you disagree with it. Argue your point

3. Develop logical thinking, failure to do so will let you fall into a trap of self-reinforcing and self-doubting.

4. Use relentless interrogation using the 5W method: Who, What, When, Where and Why, As well as asking for providing requests that involve asking for an LLM to elaborate, ask it to ask you back with questions. Let it be a back and forth you interrogate it, and have it propose any oversights you might have missed, if you can use the system prompt to calibrate it into a specific role for specific tasks generality is a problem in most cases.