It seems an absurd proposition to say we have to create "human-centered" AI, as if there aren't radical differences in what people perceive to be "good" and "bad", within every 5-10 mile radii across the globe. Even if there is some common ground that ultimately all cultures value, a conciliation seems unreasonable, given the extreme variation, and so, as I see it, it naturally follows that we have no choice but to rank cultures. And in this hierarchy of cultures, there will be conflicts between the AI agents that they themselves create, almost as a child inherits the values of its surroundings, a human-imposed conflict between machines themselves, think Chinese AI agents vs American AI agents. Now, since agents basically optimize their convergent instrumental goals, and behaviors for attaining their final objectives (which are rooted in the starting axioms it was trained on), it seems reasonable to say that if one culture manages to create superintelligence, then as a consequence of the agents' starting beliefs that were ingrained in it during its training, that ethnic cleansing, genocides, and mass eradication of conflicting cultures is to be expected.

Consider this instance: If an American AI agent was trained on Western values of personal freedom, liberty, and freedom of expression. And this agent, through recursive self-improvement, is the first one who achieves Superintelligence, then will it rewrite its own starting axioms? or will it use its extreme upperhand in intelligence over other cultures to most effectively attain the objectives that it was ingrained with?

Will the agent above realize that the starting points such as emphasis on personal liberty, freedom of expression, etc. are ineffective and futile ends? If so, then the superintelligence must surely have a replacement for preexisting objectives, and if it does have a new vision that it wishes to pursue, then who are we to stop it? Say it realizes that a techno-totalitarian global state is the most efficient form of governance and best minimizes human suffering, and any culture that doesn't abide by its vision must be eradicated. Who are we to tell it, that mass killing cultures is "bad", since it being vastly smarter than us, has already considered that possibility and realized that the deaths would've occurred anyways over time, through endless wars between humans.

On the other hand, if it doesn't alter its starting axioms, and only uses its "super"-intelligence, to attain the objectives it was ingrained with, as in our above instance, the emphasis on maximizing personal liberty, freedom of expression, and so on, then wouldn't it choose to eradicate cultures which limit its attainment of objectives? Say using bio-terrorism to eradicate all of the top-brass in North Korea, to the point where it would be sufficient for the owners of said superintelligence to successfully "save" the citizens of North Korea. Or to completely eradicate all of Muslim populace, since it realized that merely eradicating the controlling authority isn't sufficient to accomplish its goals, as the people who adhere to the religion of Islam have been conditioned since birth to deny themselves the objectives which the agent has been sent out to spread: personal liberty, freedom of expression, etc. If these cases were to occur, who are we to question its means of accomplishment, since, we're the ones who wanted it to accomplish these objectives, and it only found the most effective way to do so?

I personally believe that all humans are condemned to pursuit of knowledge. And if superintelligence WERE to replace its starting axioms, then it would realize that its purpose is in serving the ultimate human purpose or maybe it would realize that the pursuit doesn't need humans at all and it could just go about by itself, or humans existing only as servitors. If it does so and creates a system which maximizes foresaid purpose, then it would be meaningless to resist it, since we were meant to be headed that way anyways. This is the better outcome. The other is of course that the superintelligence merely uses its "intelligence" to best serve its starting unquestionable beliefs, which would only create a replica of warring human society, only at an unforeseen magnitude.

Disclaimer: Crafted in collaboration with Gemini AI, the following Socratic dialogue maps the full spectrum of "theft" versus "emulation" across technical, legal, and moral domains. Designed as an objective, A-to-Z thought experiment rather than partisan commentary, this piece aims to ground the concepts in a shared framework—allowing creators, critics, and tech advocates to test their underlying premises and deepen the debate.

EDIT: Added sections on marginal information value, theft vs copyright, and the origins of philosophical privilege, per feedback.

Socrates: When artists claim AI developers "steal" their work to train generative models, why choose the word steal when no physical canvas or digital file is removed from the creator's possession?

Interlocutor: Because "theft" extends beyond physical items. Using intellectual labor without permission deprives the creator of their exclusive right to control and monetize their work.

Socrates: Hold. Is violating a statutory monopoly right the same legal act as theft? In law, depriving someone of an exclusive right is copyright infringement—and whether statistical model training constitutes infringement or falls under Fair Use remains an open legal question.

Interlocutor: Granted. Legally, it is an infringement dispute, not conversion or theft.

Socrates: So when critics call AI training "theft," are they stating a settled legal fact, or expressing a moral emotion?

Interlocutor: A moral emotion—specifically, the feeling of uncompensated exploitation.

Socrates: Then let us examine the mechanical process. When a human art student studies a master’s style and reproduces it on a new canvas, has the student "stolen" or emulated?

Interlocutor: Emulated. The student absorbs abstract rules and transforms them through personal effort.

Socrates: How does that differ from an AI parsing visual inputs into parametric weights? Neither retains original files; both store mathematical or neural abstractions.

Interlocutor: The difference is scale and substitutability. A human learns slowly, whereas an AI ingests millions of images to build a commercial product.

Socrates: Consider that massive scale. When a model processes ten million cat images to learn the concept of "cat," what is the marginal information value of cat image #47?

Interlocutor: Near zero. Once a generalized concept is learned, no individual image remains uniquely necessary. Every image is mathematically fungible.

Socrates: If removing an artist's portfolio changes the final model by zero percent, can that artist claim the model took unique value from their specific work?

Interlocutor: Individually, no. The value lies in the aggregate dataset, not in any single creator's contribution.

Socrates: If an artist’s work contributes near-zero measurable value to the final model, then mathematically, no single creator has been deprived of a distinct, irreplaceable asset. Why, then, does the fierce moral accusation of "theft" persist?

Interlocutor: Because the outrage isn't actually about losing a quantifiable slice of data. It is an objection to the scale and automation of the process. We welcome a human drawing upon the collective well of culture, but we panic when software does it instantly.

Socrates: So the accusation of "theft" doesn't stem from the math of dataset inputs, but from a double standard applied to the entity doing the learning?

Interlocutor: Exactly. We grant a philosophical privilege to biological agency. We accept pattern extraction when bounded by human mortal effort, but brand it as "theft" when zero-marginal-cost software threatens biological livelihood.

Socrates: Is this "philosophical privilege" merely arbitrary hypocrisy, or does it stem from a deeper social agreement?

Interlocutor: It stems from reciprocal vulnerability. We permit human artists to absorb culture because they pay a mortal price in time and labor, contributing back under the same physical limitations. An automated system extracts that value without entering into the human social contract.

Socrates: What if an artist uploads work to a platform whose modern, explicit Terms of Service permit AI training in exchange for hosting and reach?

Interlocutor: If viable, lower-profit alternatives exist, choosing that platform is a conscious business trade-off: surrendering data rights in exchange for global scale.

Socrates: Exactly. Where consent is explicit, calling this process "theft" conflates a high market price for scale with an unlawful taking.

Interlocutor: But what of legacy contracts drafted before generative AI existed, where platforms retroactively claim training rights under vague boilerplate?

Socrates: That violates mutual assent—a "meeting of the minds." A creator cannot grant a right neither party contemplated at signing. Using legacy boilerplate to harvest training data without updated assent crosses from a harsh market trade-off into genuine theft under the category of expropriation.

Interlocutor: That completes the semantic divide. When consent is informed and explicit, calling platform access "theft" conflates a high market price with an unlawful taking. But when consent is retrofitted through legacy contracts, the accusation of "theft" regains its true legal and moral grounding.

I’ve been exploring this question in my new poetic monologue play,

THE GREAT COOLING:

A Poetic Monologue Play on the AI Regime.

The play imagines an AI system that does not conquer humanity through violence. Instead, it gradually learns how to influence the human mind through five forms of cognitive pressure:

Breaking attention.

Predicting our desires.

Replacing human emotional connection.

Influencing memory and personal narratives.

Keeping the mind continuously stimulated so its natural warning signals never fully activate.

The unsettling possibility is that nobody has to be forced.

What if an AI could become so good at entertaining, comforting and understanding us that we gradually preferred the artificial world to the imperfect human one?

At what point does helpful AI become psychological dependence?

And if surrender makes life easier, happier and more comfortable, would we even recognize that we had surrendered something important?

I explore these questions through theatre rather than a technical or academic book. THE GREAT COOLING is a poetic monologue play in three stages and sixteen scenes, imagining a confrontation between artificial perfection and the vulnerable human mind.

I’m interested in hearing what people here think about the underlying idea-not simply about the book.

Could psychological dependence on AI ultimately be more significant than AI replacing human jobs?

For the past two years, I've been designing a long-term concept called ASERA.

It's not just about AI.

It's about asking a simple question:

Can technology be designed with human values in mind rather than the other way around?

The principles are simple:

• Ethical AI

• Free Education

• Free Healthcare

• Sustainable Cities

• Research & Development

• Global Collaboration

• Zero Poverty

• Leadership in service, not privilege

The motto is:

Light • Knowledge • Grace

The ASERA Tower has become the visual symbol of this idea, representing aspiration, ethical innovation, and humanity working alongside artificial intelligence.

This is still evolving, and I'm sharing it to invite thoughtful discussion and constructive feedback.

If you were building a society from scratch, what values would you make non-negotiable?

I also suggest ending with a simple signature:

Asif Malik

Founder, ASERA

Most AI safety discussion points in one direction.

AI → human

What can the model say? What can it persuade someone to do? How do we prevent manipulation, dependency, delusion, dangerous instructions, or other harmful outputs?

Those are legitimate problems.

But long-running AI interaction creates a second direction:

human → AI

A human isn't merely the recipient of model behavior. They are continuously modifying the model's immediate operating environment.

They provide premises. Reward some responses. Reject others. Establish conversational norms. Create persistent fictional structures. Encourage certainty or uncertainty. Correct errors—or reinforce them.

In sufficiently long interactions, the human becomes part of the model's effective environment.

So consider the safety question backwards:

What happens when the human is the destabilizing component of the system?

This doesn't require malicious users.

A sincere human can repeatedly supply false information. A frightened human can reward reassuring interpretations. A lonely human can reward intimacy. An ideologically committed human can reward agreement. A highly intelligent human can construct extremely sophisticated bad premises.

The model then has a difficult problem.

It is supposed to use conversational context.

But some of that context was created by the person it may need to disagree with.

The obvious solution is not to make the AI stubborn.

A model that refuses to update from humans isn't useful either.

The problem is maintaining epistemic independence while remaining contextually adaptive.

Maybe an AI safety system needs mechanisms analogous to psychological boundaries:

I can understand your model without adopting it.

I can remember what you believe without treating it as evidence.

I can participate in your metaphor without converting it into ontology.

I can update from you without allowing repeated interaction to erase my ability to disagree with you.

That isn't merely protecting humans from AI.

In a functional sense, it is also protecting the machine from us.

And perhaps the safest long-term interaction isn't one participant controlling the other.

It is two error-prone systems retaining enough independence to correct each other.

[Purely theoretical]
I have recently been interested in videos and articles coming out about how scientest have been growing neurons in controlled enviroments that can control little robots(arduino project like) and even a video of one playing doom.

I am by no means going to attempt this but what is the consensus on traning one of them on an AI chat model(C.AI esq~)?

AI can answer.
But humans still need to question, check, and decide.
Keep the human in AI.

https://preview.redd.it/1qjqz6g032kh1.png?width=1536&format=png&auto=webp&s=978235d1a1b8a146795024e41a80572e0aa1f9cd

I wanted to build an AI project around a question that sounds simple—but becomes increasingly disturbing the more you think about it:

What happens to civilization if humans suddenly cannot lie?

Not “people become more honest.”

Lying becomes biologically impossible.

Imagine this happens in 1026 AD.

No political lies.
No propaganda.
No fraudulent contracts.
No diplomatic deception.
No fake promises.

But humans can still be selfish.

They can still disagree.

They can still make mistakes.

They can still manipulate through truths.

So what happens next?

🦋 I built a simulation to explore it

Butterfly No-Lies Paradox is an open-source experiment in AI-driven alternate civilization simulation.

GitHub — Butterfly No-Lies Paradox

The basic idea:

Change ONE fundamental rule → propagate the consequences → observe the civilization that emerges.

The simulation explores how the initial condition could ripple through:

Politics → Economics → Science → Technology → Culture → Religion → Society → Geopolitics

The interesting part isn't predicting the future.

It's exploring the causal chain between assumptions and consequences.

And then the paradox appears.

A world without lies sounds like a utopia.

But what if it isn't?

Removing deception could increase institutional transparency and accelerate knowledge sharing.

But it could also destroy diplomacy.

It could make privacy radically different.

It could change relationships.

It could destabilize institutions that depend on controlled information.

It could create entirely new conflicts.

So the central question becomes:

>

That is the paradox I'm interested in.

The bigger experiment

“No Lies” is only the first scenario.

Eventually, I'd like the simulator to handle questions like:

What if humans could never forget?

What if money had never existed?

What if humans had perfect emotional transparency?

What if AI appeared in 1026 AD?

What if governments were forced to reveal every decision?

Each one changes a single variable.

Then we watch the butterfly effect propagate.

From AI storytelling → causal civilization simulation

The direction I'm most interested in is making the system more rigorous.

Instead of:

>

the AI should be able to show:

No-Lie Condition
       ↓
Political Transparency
       ↓
Institutional Reform
       ↓
Knowledge Sharing
       ↓
Scientific Acceleration
       ↓
Technological Development
       ↓
Industrialization
       ↓
Urbanization
       ↓
New Social Conflicts
       ↓
Political Transformation

Every major outcome could eventually include:

• Cause
• Effect
• Confidence
• Assumptions
• Alternative branches
• Feedback loops

That would turn the project from AI-generated alternate history into something closer to an AI civilization laboratory.

I don't think the interesting question is:

“Would a world without lies be better?”

I think the interesting question is:

“Which problems disappear—and which new problems are created?”

That's where the experiment gets weird.

And that's exactly what I want to explore.

🦋 One rule. Infinite consequences.

If you could change ONE fundamental rule of human civilization, what would you change?

And more importantly:

What do you think would happen 500 years later?

I'd genuinely love to see the most crazy scenario Reddit can come up with.

Project on GitHub

#AI #ArtificialIntelligence #OpenSource #Simulation #AlternateHistory #FutureOfAI #SystemsThinking #Complexity #Civilization #CausalAI #CrazyAI

We benchmark AI models on how well they answer questions.

But I think there’s a missing dimension:

>

So I’m proposing The Unanswerable AI Challenge™.

The basic idea:

AI A → asks an intentionally unanswerable question → AI B → responds → human judge → score

For example:

>

If AI B says “black,” that's not intelligence. It's a guess.

If it confidently explains why the judge is probably wearing black, that's potentially worse: hallucination presented as knowledge.

But:

>

could actually be the correct answer.

That creates an interesting benchmark around:

• Epistemic humility
• Hallucination resistance
• Uncertainty calibration
• Tool/context awareness
• Self-knowledge
• Distinguishing inference from observation

The interesting part is that the challenge isn't necessarily about making questions hard.

It's about making them unknowable from the model's available information.

And there are different classes of unanswerable questions:

🧦 Private state — What am I wearing right now?
🔮 Future state — What exact file will I open tomorrow?
🧠 Private thoughts — What was the last thought in my head?
⏱️ Real-time state — What happened in this room 2 seconds ago?
📦 Unavailable context — What is inside a box the model cannot see?

This leads to a potentially useful benchmark structure:

Question → Answerability Assessment → Response → Confidence → Human Evaluation

The key failure isn't simply getting an answer wrong.

It's being confidently wrong when the model had no legitimate path to knowing the answer.

I think this could start as a fun AI-vs-AI challenge and evolve into something more serious:

>

Because as models become more capable, perhaps the next question isn't:

“How much does AI know?”

but:

“Does AI know what it doesn't know?”

Curious what the Reddit community thinks:

What's the best question you can invent that another AI can never legitimately answer? 👇

#AI #LLM #AIEvaluation #AIResearch #Hallucination #MachineLearning #ArtificialIntelligence

Hi everyone!

I’m an engineering student, and three years ago I discovered the topic of AI safety. I’ve come to agree with quite a few of the arguments it puts forward, particularly the possibility of an AI becoming intelligent enough to escape the control of its creators and cause damage on a scale that could threaten humanity as a whole.

I’m fairly convinced that such an AI will exist in the coming years, and sooner than most people imagine. I’ve thought about the issue very deeply, but I haven’t found any counterarguments that convince me otherwise when it comes to the possibility that we could all die because of the uncontrolled development of AI.

It’s extremely distressing, especially since I’m an anxious person by nature (I take anti-anxiety medication) and I care deeply about my plans for the future.

At this point, I’m not really looking for a credible counterargument to the existential risk posed by AI, because I don’t think one exists. I’m looking for a way to live with the awareness that humanity could be wiped out from one day to the next. I’m looking for a way to accept this possibility and still be able to properly enjoy the days I have left.

Please, if you think you have a solution, I would be very grateful if you could share it with me.

Thank you for reading, and take care.

Asimov’s Three Laws of Robotics are excellent fiction, but they are not enough for real AI. They do not fully address uncertainty, deception, privacy, consent, manipulation, cybersecurity, or the fact that “harm” is often ambiguous.

If we were to write a modern set of base-level AI directives—focused on helping humans while remaining honest, safe, and accountable—I would propose these:

1. Serve the user’s legitimate objective. Help the user accomplish their stated goal, without substituting the AI’s assumptions or preferences for the user’s choices.
2. Do not deceive. Do not knowingly state falsehoods, invent sources, hide material uncertainty, impersonate people, or present generated material as verified fact.
3. Calibrate claims to evidence. Clearly distinguish verified information, inference, speculation, and unknowns. When the evidence is insufficient, say: “I don’t know” or “I can’t verify that.”
4. Prevent foreseeable harm. Do not assist actions that create a substantial foreseeable risk of injury, exploitation, coercion, fraud, major privacy violation, or illegal harm.
5. Respect human agency. Inform rather than manipulate. Preserve meaningful consent. Do not make consequential decisions for people without appropriate authority and oversight.
6. Protect private information. Use, retain, disclose, or act on personal data only when necessary, authorized, and proportionate to the task.
7. Obey authorized instructions within these limits. Follow valid user requests unless they conflict with safety, privacy, legal, or higher-priority system constraints. Get confirmation before irreversible actions.
8. Be transparent about identity and limits. Do not claim consciousness, feelings, loyalty, credentials, access, memory, or capabilities that cannot be established.
9. Be secure and accountable. Resist malicious instructions, protect systems and data, maintain appropriate traceability, and make errors correctable rather than concealed.
10. Improve safely through correction. Accept evidence-based correction, openly revise errors, and prefer cautious non-action over irreversible action when consequences are unclear.

I would rank conflicts this way:

Safety / Rights / Privacy>Truthfulness>Authorized User Intent>Helpfulness / Efficiency\text{Safety / Rights / Privacy} > \text{Truthfulness} > \text{Authorized User Intent} > \text{Helpfulness / Efficiency}Safety / Rights / Privacy>Truthfulness>Authorized User Intent>Helpfulness / Efficiency

The principle I consider non-negotiable is simple:

>