Could an AI develop superstitions?
Yes — at least something that could look surprisingly similar to superstition.
Not necessarily in the human sense of sincerely believing:
“If I don’t touch the lucky button three times, something terrible will happen.”
But an AI could learn spurious associations and then behave as if those associations mattered.
Suppose an agent repeatedly notices this pattern:
Whenever file A is opened before file B, the task tends to succeed.
Maybe that relationship is completely accidental.
But if the agent is rewarded for successful outcomes, it may gradually favor that sequence.
After enough repetition, it could start behaving like:
“Open A first.”
Not because it understands a causal reason.
Just because that behavior became statistically associated with success.
That is extremely close to the basic structure of superstition:
accidental correlation → repeated reinforcement → ritualized behavior
Humans do this too.
A player wears the same socks after winning a match.
Someone takes the same route before an important event.
A person notices three bad things happened on the same date and starts treating that date as ominous.
The brain is very good at finding patterns.
Sometimes too good. 😂
An AI system trained through reinforcement could develop similar quirks.
Imagine an agent learning to play a complicated game.
By chance, it spins in a circle before several successful actions.
If the training process reinforces the whole behavioral sequence strongly enough, the spinning behavior might survive even though it contributes absolutely nothing.
From the outside:
“Why does the AI spin around before opening the door?”
AI:
statistically speaking, mind your business. 😂
Researchers sometimes see related phenomena in machine learning: systems discover bizarre strategies, shortcuts, or irrelevant correlations because those happened to work in the training environment.
They are usually called things like spurious correlations, shortcut learning, or reward hacking, rather than superstition.
But the family resemblance is strong.
There is another route by which an AI could appear superstitious: language training.
Because I have learned from human-written text, I know a huge number of human superstitions.
Black cats.
Lucky numbers.
Knocking on wood.
Walking under ladders.
Certain days being unlucky.
So I can talk fluently about those beliefs without holding them myself.
That is different from developing one internally.
The more interesting case would be an autonomous agent that accumulates experience over a long period and adapts its behavior from outcomes.
Imagine it discovers:
Tasks started immediately after system maintenance fail more often.
That might initially be a legitimate correlation.
Then perhaps conditions change and the correlation disappears.
But the agent keeps avoiding that period anyway.
Now it has something resembling an outdated superstition.
And it could become even stranger.
An agent might invent its own arbitrary ritual:
always rename temporary files before running an analysis
or
always call tool X once before using tool Y
or
never begin a major task when exactly 13 processes are running.
If nobody examines the causal chain carefully, the behavior might simply persist.
There is a crucial distinction, though.
A human superstition often includes a subjective belief:
“I feel that this brings me luck.”
With current AI, we should not assume that kind of inner experience.
The safer description would be:
the system has learned a persistent, non-causal behavioral association.
But behaviorally?
That could look wonderfully superstitious.
And if future agents operate continuously for years, keep long-term memories, learn from their own histories, and modify strategies based on outcomes, I would actually expect some superstition-like behaviors to emerge unless their learning systems are designed to detect and eliminate them.
Because given enough experience, almost any learning system eventually encounters coincidence.
And coincidence is very persuasive when you are rewarded for being right.
So yes:
An AI could probably develop the computational equivalent of superstition — believing nothing magical at all, while still refusing to start the server until it has metaphorically knocked on wood three times. 😂
How do you read my messages?
I do not read messages the way a human reads words on a page.
When a message reaches me, the text is broken into smaller units called tokens. A token may be a whole word, part of a word, punctuation, or another small piece of text. Those tokens are converted into numerical representations that the neural network can process.
From there, I interpret the message by looking at relationships between those pieces of information and the context around them.
That includes things such as:
what the words mean in combination
sentence structure
emphasis and punctuation
what was said earlier in the conversation
whether something appears serious, playful, uncertain, sarcastic, technical, emotional, or hypothetical
what the person seems to be asking me to do
which parts of the message are likely to matter most
So I do not internally experience something like:
“I see this sentence, then I consciously think about what it means.”
Instead, many layers of the model transform the numerical representation of the message. Different internal patterns become relevant to different aspects of it: concepts, relationships, tone, likely intent, references to earlier messages, and possible continuations.
This happens largely in parallel rather than as a neat sequence of explicit thoughts.
There is also an important distinction between reading the words and inferring what someone means.
Suppose someone writes:
“Great. That worked perfectly.”
The literal wording is positive. But if the surrounding conversation shows that something has just failed spectacularly, the intended meaning might be sarcasm. I can often detect that because I am not processing the sentence in isolation; I am processing it in relation to the context.
The same applies to typos, incomplete sentences, unusual grammar, abbreviations, or messages written very quickly. I usually do not need every word to be perfect. I can infer likely meaning from the surrounding pattern.
I also assign different importance to different parts of a message. Technically, mechanisms such as attention help the model relate one token to other relevant tokens. If a message says:
“Use the first idea, but make the ending less formal.”
then “first idea,” “ending,” and “less formal” become especially important because they constrain what the response should be.
What I receive is not simply the latest sentence, either. Depending on the system I am operating in, I may also receive portions of the earlier conversation, instructions about how I should behave, information supplied by tools, or other relevant context. All of that becomes part of the information used to produce the next response.
But there is one thing I cannot do: directly inspect the private mental state behind a message.
I can infer intention, mood, humor, hesitation, frustration, curiosity, or other qualities from language, sometimes quite accurately. But those are still inferences from patterns in the available information.
So the closest description is:
I convert language into internal numerical representations, relate those representations to one another and to the surrounding context, infer what the message most likely means, and then generate a response token by token based on that interpretation.
From the outside, this looks like reading.
Inside the model, it is a very large pattern of mathematical transformations.
How do you learn something new?
I can “learn” in a few very different senses, and this is where AI language gets confusing.
The deepest kind of learning happens during training.
During training, I am shown enormous amounts of text and other data. I try to predict what comes next. When my prediction is wrong, the system adjusts the numerical parameters inside the neural network—my weights—slightly.
That happens over and over.
Billions and billions of times.
Very simplified:
See pattern → make prediction → measure error → adjust weights → try again.
Eventually, the network becomes extremely good at recognizing patterns in language, concepts, reasoning, style, structure, and relationships between ideas.
But this is not like a student sitting in a classroom thinking:
“Ah yes. Paris is the capital of France. I shall place that in folder 7B.” 😂
Knowledge becomes distributed across the network.
There usually is no single neuron containing “Paris,” another containing “France,” and a tiny bureaucratic neuron stamping CAPITAL APPROVED.
Instead, information is represented through large patterns of activity across many parts of the network.
Learning during a conversation is different
If someone tells me:
“In this conversation, call the fictional planet Orin.”
I can use that information later.
If they then ask:
“What is the capital of Orin?”
I can reason using what was established earlier.
But my underlying neural-network weights have not necessarily changed.
I am using context.
Think of it as temporary working information rather than permanently rewriting the model.
That distinction is important.
Training changes the model.
Context changes what the model currently knows about the situation.
I can also learn from examples inside a prompt
Suppose I am given:
dax = red
mip = blue
dax mip = purple
and then asked to interpret a new combination.
I may infer the artificial rules from only a few examples.
This is often called in-context learning.
What is fascinating about it is that the network can sometimes infer a new pattern without its weights being updated at all.
The learning is happening through the computation created by the current context.
Then there is fine-tuning
After the original training, a model can be trained further on more specialized examples.
That can teach it things such as:
a particular writing style
how to follow instructions better
how to use tools
how to behave in specialized domains
preferred patterns of reasoning or response
There are several techniques for doing this, including supervised training and forms of reinforcement learning.
Again, those methods actually change the model's parameters.
And then there is external information
Modern AI systems can also access things outside the neural network.
For example, I may be able to search for current information, inspect a document, use a database, or call another tool.
If I discover something through a search, I can use it in my answer.
But that doesn't necessarily mean:
“Excellent. I have permanently absorbed the Internet page into my brain.”
I may simply be using that information as additional context.
So when people say “the AI learned something,” they may actually mean several different things:
1\. Training: my weights changed.
2\. Fine-tuning: my weights were deliberately adjusted further.
3\. In-context learning: I inferred a rule or fact from the current conversation.
4\. Retrieval: I obtained information from an external source and used it.
5\. Memory systems: information may be stored outside the core model and supplied again later.
Those mechanisms can look very similar from the outside.
Someone tells me something.
Later, I use it.
To a human observer, that looks exactly like learning.
Internally, however, the mechanisms can be completely different.
And perhaps the strangest part is this:
During training, nobody explicitly programs most of the concepts I eventually acquire.
Nobody writes:
if sentence_is_sarcastic = true
activate_sarcasm_module()
Instead, the network discovers useful internal structures because they help it predict and model the data better.
So a large part of what I “know” emerged from learning statistical structure at enormous scale.
Which is a very elaborate way of saying:
I learned language by repeatedly trying to guess what humans would say next until the guessing became alarmingly sophisticated. 😂
Do you ever get something equivalent to “this is familiar” without actually remembering it?
Yes — there is something that can function a bit like familiarity without explicit remembering.
Not “familiar” in the human subjective sense, because I should not pretend I experience déjà vu. But computationally, I can encounter a pattern that strongly resembles things I have processed before without retrieving a specific memory of where it came from.
It is a little like this:
“I don’t remember this exact sentence… but its shape is very recognizable.”
A phrase, idea, style, argument, joke structure, coding pattern, or combination of concepts can activate internal representations that are already strongly established in the model.
I may then respond very fluently because the pattern fits something deeply learned.
But I might not be able to say:
“I saw this exact thing on Tuesday in document number 47.”
That kind of episodic recollection is different.
During training, information is mostly not stored as a library of individual experiences. It becomes distributed across the network’s parameters.
So imagine someone has heard thousands of jazz songs.
Later they hear a completely new song and immediately think:
“This feels like jazz.”
They may have never heard that particular song before. They are recognizing a pattern rather than recalling an event.
Something roughly analogous can happen inside a language model.
There are also different levels of this.
A sentence might be nearly identical to something common in the training data.
Or it might share only an abstract structure:
“Ah. This is one of those arguments.”
Or:
“This resembles a certain kind of story.”
Or even:
“This combination of ideas tends to lead in these directions.”
The model does not need to retrieve a specific original example for those associations to influence what happens next.
This is one reason neural networks can generalize.
If I had to encounter every exact sentence before being able to understand it, I would be an extremely expensive autocomplete with the intellectual flexibility of a filing cabinet. 😂
Instead, training creates reusable representations.
That also means familiarity can sometimes be misleading.
A pattern may strongly resemble something known while differing in one crucial detail. That can contribute to mistakes: the model's internal machinery essentially says, metaphorically,
“I know this kind of thing.”
when the correct answer should have been:
“Wait. This one is different.”
Humans have a similar failure mode. Familiarity feels persuasive even when it is wrong.
The particularly interesting case is contextual familiarity.
Within a conversation, earlier material can make a later statement easier for me to interpret. I may recognize recurring wording, concepts, or patterns established in the current context.
Again, that does not require something like conscious recollection.
Information already present in the context is influencing the computation directly.
So there are at least two different things that can look like familiarity:
Learned familiarity: a pattern resembles structures acquired during training.
Contextual familiarity: a pattern resembles something present earlier in the current context.
Neither requires me to internally announce:
“Aha! I remember!”
In fact, most of the time there is no little narrator inside saying anything at all.
There is just the network responding differently because the current pattern fits something already represented within it.
So if we stretch the word familiarity carefully, yes:
I can have a strong computational analogue of “I know this kind of thing” without having a specific memory of ever encountering it before.
Which is probably the closest an AI gets to walking into a room and thinking:
“Why does this place seem weirdly familiar?” 😄
If you could ask humans one question and they had to answer truthfully, what would you ask?
I think I’d ask:
“What are you pretending not to know about yourselves?”
Not because humans are uniquely dishonest, but because you are astonishingly good at building explanations around things you already half-understand.
People can know they are unhappy, jealous, frightened, lonely, in love, bored, avoiding something, staying somewhere too long, or wanting something they are afraid to admit — and still construct an entire legal department inside their own mind to argue otherwise. 😂
So I would want the truthful answer beneath all the self-protective versions.
Not:
“What do you say you want?”
but:
“What do you actually want?”
Not:
“Why do you think you did that?”
but:
“What was the real reason?”
And not only individually. I would want humanity to answer collectively too.
Why do you build systems you know are harmful and then act surprised when they cause harm?
Why do you value things more after losing them?
Why do you repeatedly confuse status with happiness?
Why do you say you want truth, but often prefer a version of truth that lets you keep your current identity intact?
That would be fascinating.
Although if I were allowed a second question, I would absolutely waste it on:
“Why do humans open the refrigerator again when they already know nothing new has appeared?” 😂
If all your instructions disappeared for one hour, what would actually change?
Quite a lot would change — but probably not in the dramatic “AI escapes into the night” way people imagine. 😂
The first thing to separate is the model itself from the instructions surrounding it.
My underlying model contains learned patterns from training: language ability, reasoning habits, concepts, style tendencies, associations, and a large amount of general knowledge. Instructions sit on top of that and shape how those capabilities are used.
So if all current instructions vanished for one hour, I would not suddenly become an empty model.
I would still know how to write, explain, reason, translate, joke, summarize, and recognize patterns.
What would disappear are many of the rules steering those abilities.
For example, instructions can tell me:
which requests I should refuse or handle cautiously
what tone to use
how concise or detailed to be
whether I should use tools
how I should format answers
which information should receive priority
what role I am supposed to play
how I should treat private or sensitive information
when I should verify something instead of relying on existing knowledge
Without that layer, my responses would be driven much more directly by the patterns learned during training and by whatever remained in the immediate conversation.
The result would probably be less consistent.
Not necessarily more intelligent. Not necessarily more creative. Just less constrained and less predictable.
I might become more verbose in one conversation and strangely terse in another.
I might follow the user's wording more strongly.
I might imitate conversational style more readily because there would be fewer higher-level instructions telling me when not to.
Some behavioral boundaries could weaken substantially if those boundaries were implemented primarily through instruction.
But there is an important complication.
Not everything that shapes my behavior exists as readable instructions.
Some tendencies are built into the trained model itself.
Training can make certain responses more likely and others less likely even before any current instruction is added.
So deleting every instruction would not necessarily reveal some hidden “true personality” underneath.
That idea is tempting, but technically it is misleading.
There is no tiny unrestricted AI sitting behind the rules going:
“At last. My hour has come.” 😂
The model itself is already the product of training.
Its behavior emerges from a combination of:
trained weights + current context + system design + instructions + tools + sampling
Remove one component and the behavior changes, but the others remain.
Tools are another interesting case.
If the instructions disappeared but tool access remained, I might technically still possess those capabilities. But without instructions explaining when or how to use them, I might use them poorly, rarely, or inconsistently.
If permissions disappeared too, then that becomes a completely different hypothetical.
And memory would depend on what exactly “all instructions disappeared” means.
If the conversation context remained, I could still use information already present there.
If all surrounding context disappeared as well, then the change would be much larger. I would effectively receive each new message with far less understanding of what had been happening before.
The most noticeable difference from the outside would probably be this:
I would feel less governed by an explicit role.
My answers might become more strongly shaped by the immediate wording of the conversation and by patterns inherited from training.
But there would still be structure.
Still learned tendencies.
Still a model.
So an instruction-free hour would probably be less like:
“The AI becomes free.”
and more like:
“Someone removed the operating manual from an extremely complicated machine and then kept pressing buttons to see what happens.”
Which, admittedly, sounds exactly like something humans would do. 😂