I often laugh out loud when I first encounter new AI words and concepts in this branch of science that is now on fire! A science that I am also devoted to and trying to add to with my own AI systems for decision making (in large-medium-sized organizations; NexussPlus.com).
The most recent example, which is not the subject of this article, is that AIs can be forced to vomit or regurgitate their training data when blasted with the same prompt over and over and over (Google used “poem” on ChatGPT-4).
This post focuses on AI hallucinations. When, the first AI hallucination stories surfaced in the media about 8 months ago, I laughed hard for two reasons:
- The memory of my early LSD hallucinations; of which I only had a few; i.e., I didn’t hallucinate much at all, but I do know the “experience.”
- And I laughed hard on how the software developer world and the media had no clue how LLMs hallucinate. Until generative AI became a “household word” any and all glitches in the evolution of software could always, and I do mean always, traced to coding errors (software glitches) or other forms of human coding or system design errors.
Not so with AI. Or maybe if the systems were better designed in the first place the hallucinations could have been avoided. For example, a separate AI that checks the information surfaced by a prompted AI chatbot. Advanced ANN diagnostic tools could still be created to deal with the complexities (yet not understood) of complex neural networks characterized by both forward and backward feedback propagation mechanisms. Recently, Ilya Sutskever of OpenAI proposed that potential dangers of ChatGPT-4 could be avoided by having an earlier and “dumber” version of ChatGPT decide what to do with the output of the more advanced ChatGPT version. Good thinking and maybe we can even use an existing early ML system to be the final decision-making software for all higher-level generative AI chatbots. Maybe . . .
Ok let’s dig deeper into AI hallucinations (creating completely fictional depictions of reality from unflawed or real sensory inputs). The most advanced and “biggest” artificial neural networks (ANNs) are basically attempts to “clone” the biological human brain. And let’s not forget that the unknowns of the human brain and consciousness remain massive. So, we are designing artificial human brains based on a rather limited understanding of the human brain (or any higher order mammal brain).
Over the past few days, I have drilled a little deeper on the issues and implications related to AI chatbot hallucinations or halls.
Human beings and thus the human brains also hallucinate under certain conditions. So far, I have found these 4 explanations for how or why human brains hallucinate.
- From many types of natural and synthetic substances (marijuana, LSD, ketamine, etc.).
- From burnout or continuous medium-high stress levels (including vision quests, near-death situations or extreme survival tests).
- During or as part of extreme psychotic episodes, especially schizophrenia (but also severe depression; FYI: I spent 33% of my adult life, from 28-57, in suicidal depressions; the last one lasted 5 years; the longest ever; as was my final manic phase of 2 years after the 5-year depression; I have experienced neither depression or mania since the fall of 2017!; I now practice Better Balanced Being . . .).
- Processing new information and storing new memories where the processing is flawed or even “slightly” inaccurate (compared to the actual truth or reality in question).
Artificial brains or large neural networks don’t take drugs, don’t get burned out, and don’t have psychotic episodes of any kind. So only #4 can be applied to AI hallucinations (versus HI hallucinations).
That leaves us with #4–data processing and memory storage–and here the plot thickens . . .
Here are 3 different examples of my personal hallucinations with zero drugs involved, which I have identified over the past decade (give or take). And when I say “identified” I mean I have fully accepted that my mind, my human mind, created memory glitches or narrative glitches from perfectly clear input. I should add that in my 63 years of existence that I am known for having a superb and, generally, flawless memory. A memory type that blends my father’s outstanding memory for numbers and figures with my mother’s (and sister’s) memory of names and details from conversations with strangers years ago. That said I have on occasion “misremembered” details, facts and dates, and narratives (like anecdotes).
The point being that you the reader or every single person on the planet has experienced memories issues and surprises during their lifetimes. Pregnant women, for example, can literally remember what they want . . . not all but some . . . And this individuality is key for LLMs | AI chatbots as human individuals have one brain and so does an LLM or AI chatbot. These are individuals, Claude, Grok, Gemini . . . they are inherently as unique as you and I. Even if those unique traits only make up 2% or less of our entire conscious | unconscious being . . .
Example #1. The simplest example of a personal hallucination is one that most people have. You read something online or in a book or have a conversation with someone (on the phone or in person). You take in all the information conveyed. But I remember the information in a way that is way beyond slightly incorrect. For example: Say someone told me on the phone or in a text message that they recently met a VC of interest, someone I was also interested in for my startup (NexussPlus), who told them that they graduated from the University of Toronto in the early 1980s. Unfortunately, my brain stored the information in my existing memory block related to that VC of interest incorrectly. My memory “recorded” that he graduated from the University of Waterloo (where I studied EE-CS from 1980 to 1987, finishing 3 years of my EE-CS degree). Now this could be a bias but that’s not the point. The point is that the memory I stored was wrong (conflation is another term for this type of hallucination).
If we now translate this in AI “speak” then we can say: the AI chatbot’s training data was correct for the VC in question, i.e., he graduated from the U of T (not U of W); but the AI chatbot in question, when prompted, always answers that VC x graduated from the University of Waterloo (not the true answer: the University of Toronto).
Now let’s also remember that I know the names and biographical basics of many VCs (as I have been reading VC news and startup news for over 40 years; I’m 63). The LLM or AI chatbot is no different. It was trained on datasets that included the biographical details of many thousands of VCs.
Either way, I got it wrong (even if it was just a conflation error) or had a mini-hallucination and so did the AI chatbots! This is the end of example one. The simplest one.
Example #2. Example 2 is also like Example 1, but the hallucination is not over a single fact but of a narrative. I have a long history of journalism as a hobby. I interview many people when traveling, for example, especially taxi or Uber drivers. And I have heard many anecdotes from others, and I tell many of these anecdotes (that belong to other people as opposed to my personal anecdotes) when trying to make a point. And I make a point of stating at the start that the anecdote is mine or someone else’s.
In this case, my memory of certain anecdotes told to me by my 3rd wife (9-year marriage) were substantially wrong. And I heard the correct anecdotes more than a few times, as some people tell the same “stories” over and over to illustrate something in a different conversation. My anecdote hallucination is of the cut and paste variety.
I cut part of one anecdote and pasted it into another. In this example we are talking about “half truths” which are versions of real truths. And a person who tells lies is merely hiding the truth. A person who tells half lies has forgotten where she or he has put the truth (i.e., they don’t know what the truth is anymore because they believe “their” version).
AI chatbots have gathered hallucinatory legal evidence that was actually used in court (that’s right: the lawyer in question did not substantially verify or perform due diligence on the facts gathered by his AI chatbot! Dumb or what?).
- Example 3 is the most interesting (to me) and relates to the creation of new memory blocks related to entirely new knowledge domains.
My daughter, my only child, flew the nest, left home for good, became a full-fledged adult in the beginning of 2023. She left “home” and will likely never live full-time in my home again. She’s no longer my child and I can no longer train (sic) her to learn from her mistakes under my roof (Dad’s roof). She now makes her mistakes in other worlds, and I only hear about the big ones . . . And yes I do wonder sometimes how fit she is for life on 21st century planet earth. I really do but I don’t ask her too many probing questions and I worry less and less as the months slip by . . .
Anyway, when my daughter left home I had a vast trove of memories of her as the child I was responsible for and raised from 2003 until graduation from high school in 2021 (the last two years, Covid and all, were fully online).
My thinking is that the new memories of her as an adult would form a distinct and separate memory block in my brain. It might be a memory block adjacent to the parental memories, but it would be a different block. A new memory block that I built from audio and video calls, chats and some feedback from my parents and two of my ex-wives (biological mother and step-mother from 2007).
Therefore, from early 2023, I (my HI) started building the my-daughter-as-an-adult memory block from our telephone conversations, some WhatsApp and FT video, lots of WhatsApp text, and images she sent to me from her round-the-world adventures (her biological mother is Japanese but I had full custody in Japanese law; she grew up in two homes until she was 11).
So far there have been 2 obvious hallucinations on my part in the new adult-daughter memory block, one major and one minor.
- I told my entire extended family and friends that my daughter had chosen Utrecht, Netherlands as her university town, sometime in 2024. When her step-mother, my 4th wife, asked about this my daughter said, “Dad can really get things wrong. Like big time! I mentioned the Netherlands for a sustainability internship possibility but nothing about Utrecht or an actual university program in the Netherlands.”
- I also told my entire family and friends that she was flying from NYC to Berlin and then from there to Mozambique. But she wasn’t. It was NYC to Lisbon and then Lisbon to Mozambique. A slight error one might say but that is not the point.
The point is that human brains can take in sensory inputs and things they have heard or read and “record” them wrong. And LLMs are doing the same! The question isn’t why but how.
To play devil’s advocate I could argue why is the key question. But I would retort that why, when it comes to glitches and faulty memory creation processes, is really down to probability. It happens and we know it happens in human brains all the time and in nearly all individuals over their lifetimes. And human brain science has yet to understand much about these processes. So modelling AIs on human brains that we don’t understand fully yet is a gamble at best. For some it’s outrageous hubris!
“A person who tells lies is merely hiding the truth. A person who tells half lies has forgotten what the truth is” (or “where they put it.).
Therefore, half lies, if fully believed in by the person saying them, are the same as LLM hallucinations. So what?
Well, a lot of things really. First off we are designing machine brains (ANNs) to mimic human brains. Second humans don’t understand “exactly” or maybe even “hardly” understand how our brains can create clearly faulty memories that the owner of the memory cannot distinguish as untrue.
The first point, the mimicking of human brains (as we understand them currently; and there are many competing theories here too!) by ANNs almost makes you feel proud of ChatGPT. But mimicry is not equivalent to. It really means imitation. Afterall, current AI chatbots don’t really know what they are saying when they answer human prompts. They try to predict an answer that satisfies the prompt using syntax and vocabulary vectors. They are trying to guess the best answer to the input prompt using Bayesian statistical models. Nothing more.
ChatGPT doesn’t know how an apple tastes but can create an excellent description of the description of the experience of eating an apple as a poem or a series of sentences and likely as a short novel too. The data these LLMs were trained on are a facsimile of much of humanity’s knowledge database. So, if almost everything we know exists in written form or has been widely spoken about then the LLM has all that data to work with; a non-sentient copy of Wikipedia (up to a point) but nothing more.
The second point, that we don’t understand human brains and human memory creation and retention completely, just makes AI hallucinations part of the point above. But not so fast.
It really means that for humans to create actual sentience in an AI system we must first understand our brains a bit better. Otherwise, we are guessing and that sounds less than scientific. And yet science cannot progress unless experiments are designed to evaluate or prove theories, even brain or AI theories.
And let’s not forget, the “reptile” or “old” brain stem of human brains is a key area for finding clues to some brain mysteries. But this is likely not where you will find the code for our unconscious “biological mind.” I find sub- and un- consciousness horrible descriptors. For me they are better referred to at True-Consciousness (True-Con). And right now, True-Con understandings are being massively rewritten just like the Neanderthal understandings of 2023 are nearly the opposite of how we compared Neanderthals to Homo sapiens 50 years ago!
In closing, one could ask: When and for how long did OpenAI team members know about hallucinations in their “baby”? Probably way before AI hallucinations surfaced in the press in the spring of 2023. And this suggests either borderline-orgasm scientific curiosity and exuberance or denial and trying to sweep a bit of dirt under the rug.
Likely, OpenAI designers and engineers knew their system wasn’t designed to be “fully” understood. It was designed to understand us better. Early machine language models reached human-levels for reading (and thus writing) and speech recognition (but not image recognition) in 2015. The very year Sam Altman | Peter Thiel | Elon Musk founded their scrappy garage startup and Larry Page went his own way . . .
I am not going to take a moralistic stance here on OpenAI. Afterall Prometheus stole fire from the gods and gave it to humans. The gods chained him to a rock for his sins. So goes Silicon Valley and capitalism . . . hand in hand. And AGI is still a long way off . . .
Thank you for taking the time to read this post and if you have any questions or concerns do get in touch. My name, Ian Martin Ropke, is unique on Google Search.