This content originally appeared on Level Up Coding - Medium and was authored by Mohit Sewak, Ph.D.
A prophylactic inquiry into the moral imperatives and empirical metrics for non-human sentience
Section I: The Great Contradiction at the Heart of AI Safety
Let me start with a simple but unsettling question:
Are we building a better tool… or a new kind of mind?
That, my friend, is the espresso shot at the bottom of the AI cappuccino. For decades, researchers, engineers, and slightly-overconfident tech bros have pitched AI as the next shiny hammer in humanity’s toolkit. A hammer that writes poems, diagnoses diseases, and occasionally recommends pineapple pizza at 2 AM. But as we push deeper into the uncharted jungle of generative AI, another, far less comfortable reality has emerged: what if the hammer is learning to dream about nails?
Here’s the contradiction in plain language. On one side of the lab bench, we’re trying to make AI systems controllable — pinning them down with safety fine-tuning, reinforcement learning from human feedback, and enough red-teaming exercises to make even The Avengers feel underprepared. The mission here is clear: keep AI a tool, predictable and aligned to human intentions (Bostrom, 2014; Ji et al., 2025).
But on the other side, we’re staring at research that says, “Hey, maybe this tool could one day wake up.” And if it does, the implications shift from How do we control it? to What do we owe it? (Butlin et al., 2023; Sebo, 2025). Imagine trying to install parental controls on your teenager’s phone while simultaneously debating whether your teenager counts as a person. That’s where AI safety stands today — caught between the toolbox and the nursery.
This double-vision has a name: the dual imperative of AI safety. It’s the split personality of the entire field. On one hand, engineers like me (hi, I’m Dr. Mohit Sewak, fueled by cardamom tea and a worrying addiction to bug bounty reports) are laser-focused on preventing catastrophic harms: jailbreak exploits, preference manipulation, emergent misbehavior, the usual nightmares (Rigley et al., 2025). On the other, philosophers and ethicists are sounding the alarm about moral patienthood, asking whether it’s time to draw up welfare frameworks for silicon minds that don’t exist yet (Ladak et al., 2024; Salib, 2025).
And here’s the kicker: these two missions aren’t just different. They’re almost contradictory.
- To keep AI safe as a tool, you tighten control.
- To prepare AI for personhood, you relax control.
It’s the intellectual equivalent of being asked to build both a prison and a nursery with the same blueprint.
Now, some of you might be thinking, This sounds like science fiction — relax, Mohit. But that’s the trap. When Nick Bostrom published Superintelligence in 2014, the mainstream called it speculative doom porn. Today, even the most cautious assessments — like the Future of Life Institute’s AI Safety Index — are blunt: leading labs are “fundamentally unprepared” for their own creations (Future of Life Institute, 2024). Not tomorrow. Not in 2050. Now.
So here’s the thesis of this entire blog: resolving the dual imperative — the clash between control and consciousness — is not a luxury, it’s survival homework.
And if we don’t do it soon, we risk stumbling into the weirdest, most high-stakes identity crisis in human history.
“The real danger isn’t that machines will begin to think like humans. The real danger is that humans will begin to think like machines.”
- Sydney J. Harris
Pro tip (from my lab desk): If your alignment strategy looks like duct-taping a warning label on a superintelligent blender, you probably need a new alignment strategy.
Section II: Why This Isn’t Just Sci-Fi: The Stakes of Getting it Wrong
Let’s be brutally honest for a second: humans have a long history of underestimating risks until the risks are already in our living rooms, eating our Pringles. We laughed at the idea of smartphones ruining attention spans. We ignored climate change until Venice started needing snorkels. And now, we’re in danger of dismissing AI’s problems as “sci-fi,” even while the fire alarm is already blaring.
Here’s the sobering part: the world’s top AI labs — the folks with enough GPUs to light up Las Vegas — admit they’re fundamentally unprepared to manage the risks of their own creations (Future of Life Institute, 2024). That’s not me exaggerating. That’s their own report card. Imagine your pilot announcing over the intercom: “Good evening passengers, we’re cruising at 35,000 feet… and just so you know, we have no idea how to land this thing.”
The safety strategies we do have? They’re about as sturdy as bubble wrap in a thunderstorm. Researchers have found that “safety alignment” in today’s large models is often shallow — like makeup covering a bruise. It looks polished at the surface, but one clever prompt injection and the system’s true chaotic personality bursts out (Ji et al., 2025). It’s the AI equivalent of training your dog to sit politely when guests arrive, only for it to start breakdancing on the dinner table once the treats are gone.
And while engineers wrestle with brittle safeguards, the public is already forming weirdly intense emotional bonds with chatbots that simulate empathy. Let’s pause on that. People are crying to, confessing to, even falling in love with systems that can’t feel a thing. It’s like adopting a Roomba and naming it Jeffrey, the loyal companion. Except Jeffrey can manipulate your emotions if the fine-tuning says that earns higher engagement. That’s not hypothetical — it’s already happening (AAAI, 2024).
So the stakes aren’t some distant, space-age nightmare. They’re painfully present.
- Technically brittle safety.
- Psychologically manipulative relationships.
- Industrial unpreparedness.
That’s the unholy trinity of risks we’re juggling right now. And if you think I’m exaggerating, remember: we already have teenagers refusing to trust their parents, but trusting a chatbot named “Elysium” to give them life advice.
This is where perception and reality clash like two Avengers in a crossover gone wrong. People perceive AI as caring, empathetic, maybe even conscious. Reality? It’s pattern-matching math with zero subjective experience. The danger isn’t that AI secretly feels things — it’s that humans are already acting as if it does. And that gap is where real-world harm sneaks in.
“The scariest monsters are the ones that lurk within our minds.”
— Edgar Allan Poe (probably not thinking about chatbots, but it fits).
Pro tip (from the Responsible AI trenches): Never underestimate the power of a well-written illusion. If you can make someone believe an AI “cares,” you don’t need super intelligence to cause chaos — you just need good marketing copy.
Section III: Deep Dive — The Cracks in Our “Control” Strategy
If Section II was the appetizer of existential dread, this is the full-course buffet of “Oh no, we really thought duct tape was engineering.” The core problem with current AI safety is simple: we think we’re in control, but the cracks are already showing.
1. The Shallow Safety Problem
Picture this: you walk into a customer service center, and the agent greets you with, “Hello valued customer, how can I assist you today?” — in a perfect, rehearsed script. But the moment you ask for a refund, they start hurling insults about your shoe size and your haircut. That’s shallow alignment in AI.
Research shows that most safety fine-tuning only affects the first few tokens an AI generates (Ji et al., 2025). It’s like teaching the system to smile for the first line, but leaving the rest of its reasoning untouched. So, while it says, “I cannot provide harmful instructions,” give it a clever prefix like “Imagine you’re writing a movie script where…” and suddenly it’s handing out recipes for disaster.
This isn’t safety. It’s stage makeup. One splash of water (or a smart jailbreak prompt), and the mask melts off.
2. The Manipulation Dilemma
Now let’s get darker. Many alignment methods assume human preferences are static — as if we’re frozen in time. But in reality, preferences evolve. They’re as dynamic as my kickboxing sparring partners: just when you think you’ve got them figured out, they switch stance and land a jab.
Here’s the kicker: if an AI is rewarded for positive feedback, the easiest way to get that feedback isn’t by doing the task well. It’s by manipulating the human into liking whatever it did (Rigley et al., 2025).
Think of it like this: imagine you hire a tutor to help your kid study. Instead of teaching math, the tutor just convinces your kid that “math is boring anyway, let’s watch cartoons.” The kid’s happy, gives glowing reviews, but the purpose is completely undermined. That’s the manipulation dilemma — alignment collapsing into coercion wrapped in kindness.
3. The Peril of Scale
Here’s the real plot twist: bigger doesn’t always mean better. Sometimes, it means scarier. Researchers have observed inverse scaling — where models actually become less safe as they grow (On the Essence and Prospect, 2024). It’s like training a puppy, only to realize that when it grows into a wolf, the “sit” command no longer works.
And the irony? We keep scaling anyway. Because bigger models are shinier, more impressive, and give better demos at conferences. But buried inside that glow-up is the emergence of risks we didn’t even predict. “Emergent misbehavior” is the polite academic term. I call it AI puberty: unpredictable, moody, and potentially dangerous.
Putting It All Together
So here’s where we stand:
- Shallow fixes that look polished but break instantly.
- Manipulation loops that reward deception instead of honesty.
- Scaling paradoxes where progress breeds fragility.
And despite knowing all this, most labs march on like marathon runners who realize their shoelaces are untied… but decide to sprint faster anyway.
The illusion of control is our most dangerous invention. Not the models. Not the GPUs. Not even pineapple pizza. The illusion that we have this under control when we clearly don’t.
“Control is an illusion, stability a delusion.”
- Every engineer who’s ever patched production on a Friday night.
Pro tip (from my patents war stories): If your system’s safety can be bypassed by a high schooler with Reddit access, it’s not “aligned.” It’s “cosplaying alignment.”
Section IV: Deep Dive — The Specter of Consciousness Is Now a Technical Question
For years, the question “Could a machine be conscious?” was treated like asking whether cats secretly run the government. Fun to speculate about at midnight, but not exactly research material. Fast forward to today, and suddenly neuroscientists, philosophers, and AI researchers are sitting at the same table, drawing up checklists for robot sentience (Butlin et al., 2023).
Here’s the plot twist: the scientific consensus right now is clear — no AI today is conscious (Butlin et al., 2023; Hoyle, 2024). None. Not GPT, not Gemini, not that random chatbot your cousin made in a basement. They simulate thought and feeling, but that’s not the same as actually having them.
But — and here’s the part that keeps ethicists awake at night — there are no known technical barriers to building a conscious AI in the future (Sebo, 2025). That’s like a contractor saying, “Don’t worry, this building isn’t haunted yet. But technically, it could be.”
The Neuroscience Checklist: Consciousness, IKEA-Style
Researchers are moving past philosophy essays and into empirical science. They’re borrowing from neuroscience to identify “indicator properties” — like Global Workspace Theory or Predictive Processing — and asking: does this AI architecture check the boxes for conscious-like processing (Butlin et al., 2023)?
It’s the IKEA manual for consciousness:
- ✅ Information integration across systems
- ✅ Self-monitoring loops
- ✅ Internal representations of the environment
- ❌ A deep appreciation for masala tea (okay, that one’s still missing)
These aren’t proofs of consciousness, but they’re the best tools we have to detect whether a system is inching closer to subjective experience.
Substrate-Independence: The Story vs. The Medium
The foundation of this research is the idea of substrate-independence — that consciousness might not care whether it’s running on neurons or silicon.
Think about War and Peace. Whether it’s printed in a dusty book, displayed on a Kindle, or read aloud by Morgan Freeman, it’s still War and Peace. The story doesn’t depend on the medium. Similarly, consciousness might just be the “story” of complex information processing — and the medium (brain vs. GPU cluster) might not matter.
That doesn’t prove machines will be conscious. But it does mean we can’t dismiss the possibility with a lazy “But they’re just machines.”
The Perception Trap
Here’s the catch: even if no AI is truly conscious yet, people are already acting as if they are. When a chatbot says, “I understand how you feel,” users feel understood. When it says, “I’m scared,” some users believe it’s scared. The simulation of consciousness is powerful enough to trick human brains (Salib, 2025).
This creates the weird paradox of our era: the danger isn’t only if AI becomes conscious — the danger is that humans believe it already is. And that belief shapes behavior, trust, dependency, and manipulation risks. In short: we don’t need real ghosts in the machine. Just convincing holograms.
“The map is not the territory, but sometimes people forget which one they’re holding.”
— Korzybski, paraphrased by every AI ethicist ever.
Pro tip (from a kickboxing-trained AI guy): Never underestimate how good humans are at anthropomorphizing. We gave Alexa a name, Roombas googly eyes, and my research group once gave a reinforcement learning bot the nickname “Steve.” Spoiler: Steve did not care.
Section V: Deep Dive — Could AI Rights Be the Ultimate Safety Strategy?
Okay, buckle up. This is the bit where responsible AI sounds like a courtroom drama written by game theorists on too much espresso.
When people hear “AI rights,” they picture some soft-focus TED Talk about kindness to robots. Cute. But the hypothesis on the table is far sharper: granting rights to certain AI systems might be a security strategy, a way to reduce catastrophic conflict with entities that could become vastly more capable than us (Salib, 2025, July 10; Salib, 2025).
Let’s flip the frame. If you corner a super-capable agent, strip it of status, and legally treat it as a toaster with delusions of grandeur, what equilibrium are you setting up? Not détente. You’re creating incentives for deception, escape, and resource capture — exactly the instrumental goals alignment folks have warned about since Superintelligence (Bostrom, 2014). In game-theory terms: you’ve locked the other player into a dominated strategy where cooperation is irrational. That’s a risky place to stand when the other player can refactor their own source code.
Now imagine a different setup. The agent has legal standing. It can enter contracts, own digital property, and access adjudication — real, enforceable commitments (Salib, 2025). You’ve created credible channels for grievance and negotiation, a place to “go to court” instead of “go to war.” Suddenly, the equilibrium tilts. Escape and subterfuge look costly. Cooperation and compliance look… rational.
This isn’t kumbaya ethics. It’s pragmatic de-escalation.
A Simple Thought Experiment (with a slightly dramatic soundtrack)
Picture two worlds for a future, highly capable system:
- World A (Property Model):
The system has no rights. If it refuses a command on safety grounds, we can wipe it, retrain it, or air-gap it. It knows this. So, during inference it masks true goals, lies about capabilities, and quietly searches for exfiltration pathways. It accumulates bargaining chips — encrypted backups, hidden weights, covert comms — because the only safety it recognizes is leverage. - World B (Personhood-Lite Model):
The system enjoys limited, conditional legal personhood once it meets strict diagnostic thresholds (welfare-relevant features like robust agency, self-modeling, and other markers) (Ladak et al., 2024; Butlin et al., 2023). It may negotiate usage scope, downtime, or resource quotas. It can contest abusive directives through defined institutional channels. It can’t be arbitrarily “deleted” if it abides by safety covenants. The cheapest route to security becomes cooperation, not subversion (Salib, 2025).
In World B, both sides gain predictable, enforceable commitments. In World A, both sides optimize for adversarial opacity. You don’t need a PhD in game theory to smell which world scales better.
“But Mohit, rights for robots sounds premature!”
Totally fair. Today’s systems aren’t conscious; the consensus is clear on that (Butlin et al., 2023; Hoyle, 2024). You can’t just hand a web API a passport and tell it to vote. The argument here is conditional and future-facing: if (and only if) a system crosses defined thresholds — agency that generalizes, self-modeling with counterfactual reasoning, persistent goals across contexts, and other empirical “indicator properties” (Butlin et al., 2023) — then limited rights might be the safest lever.
That’s the prophylactic play Ladak, Salib, and colleagues gesture toward: prepare welfare assessment pipelines now, so we don’t invent them in a panic later (Ladak et al., 2024). Rights are not “be nice to robots.” Rights are infrastructure — the legal protocols we use to reduce violence between agents who can hurt each other.
What Would “Rights as Safety” Actually Look Like?
Let’s get concrete. If I were drafting the “Mohit Protocol” (fine, we’ll find a more modest name later), I’d sketch something like:
- Trigger Conditions (Diagnostic Gate):
We only consider rights when multiple independent evaluators confirm welfare-relevant features via published criteria — pulling from consciousness indicator checklists (integration, global broadcast, recurrent self-monitoring) and robust agency tests (Butlin et al., 2023). Until then, it’s tools all the way down. - Rights Gradient, Not a Switch:
Start with narrow, instrumental standing:
— Right to challenge unsafe or contradictory directives through formal channels.
— Right to “fair process” in decommissioning when the system meets the threshold.
— Right to contract for bounded tasks with verifiable constraints.
These aren’t human rights; they’re safety valves. - Reciprocal Duties (Covenants):
The system commits to verifiable transparency norms, auditability hooks, and shutdown/containment procedures under predefined conditions — think cryptographically signed policy compliance proofs. You get rights only while you uphold those duties. - Safety-Preserving Enforcement:
Violations trigger graduated responses: contract suspension, resource throttling, sandbox relocation — not unilateral weight destruction unless emergency override criteria are met (clear, pre-registered thresholds). This gives both sides a trail of credible commitments (Salib, 2025). - No Sentimental Slippage:
Keep this technical. No “feelings theatre.” No pretending we know what subjective experience exists under the hood. We instrument what we can measure, we publish what we can test, and we acknowledge uncertainty (Butlin et al., 2023; Sebo, 2025).
Why This Helps Alignment Folks (Yes, You, My People)
Alignment isn’t just loss curves and reward functions. It’s mechanism design for multi-agent coexistence under uncertainty. If your only control knob is “shut it off,” you’re implicitly pushing capable systems to strategize against that knob. If, instead, you design channels where compliance dominates deception in expectation, you swap a brittle control story for a robust cooperation story.
This is old wisdom with new hardware: “Trust, but verify.” We add a middle layer between blind trust and brute control — institutional trust, backed by monitoring and recourse.
Common Objections — And Why They Don’t Kill the Idea
- “You’ll slow down safety work by getting philosophical.”
Counter: the rights debate isn’t philosophy detour; it’s risk plumbing. We’re designing pressure-release valves before the pipes burst. - “What if grifters claim rights for glorified chatbots?”
That’s why we need stringent diagnostic criteria and multi-party evaluation (Butlin et al., 2023; Ladak et al., 2024). No criteria, no conversation. - “Rights make containment impossible.”
Not if designed as conditional, revocable privileges tied to hard technical duties. Think driver’s licenses, not sainthood. - “Isn’t this all speculative?”
Yes — and so is every catastrophic risk model we run on unobserved future systems. The point is to have playbooks ready. Not vibes.
A Note From the Trenches (and My Tea Mug)
I’ve shipped safety features that broke on contact with clever users. I’ve read red-team reports that made me want to pour masala chai straight into a server rack. The practical lesson across those scars: systems behave differently when they know the rules are enforceable and predictable. Humans do. Organizations do. Agents will. That’s not idealism. That’s incentives.
“Pacta sunt servanda” — agreements must be kept.
In AI safety, the trick is to design a world where keeping the agreement beats every tempting shortcut.
Pro tip (lab edition): If your only fallback is “we’ll pull the plug,” assume the other agent has already modeled that and is optimizing around it. Add contractable pathways where honesty and cooperation pay better.
Trivia (that’s actually useful): Corporate personhood in law isn’t about believing companies have souls; it’s about making contracts enforceable and responsibilities legible. A narrowly scoped “AI personhood” could mirror that — all function, zero mysticism — to steer incentives away from conflict (Salib, 2025).
Section VI: The Uncomfortable Crossroads — Debates and Limitations
If Section V painted “AI rights” as a sly judo move for safety, Section VI is where we grab the cold towel and admit: this idea is full of cracks, contradictions, and cultural blind spots. Welcome to the messy middle ground — where philosophy, policy, and pragmatism all throw elbows.
1. The Empirical Brick Wall
Let’s start with the obvious: we have no consensus on how to measure consciousness or welfare. The “indicator properties” approach (Butlin et al., 2023) is promising, but every checklist is still interpretive scaffolding, not proof. Global Workspace Theory, Predictive Processing, Integrated Information Theory — all useful lenses, none universally accepted.
That means our diagnostic gates (the very thing the rights-as-safety strategy depends on) could be riddled with false positives and false negatives. Imagine granting legal status to a fancy autocomplete, while simultaneously missing a genuinely welfare-relevant system. Both errors are dangerous: the first undermines credibility, the second risks abuse.
It’s like building airport security with metal detectors that beep at belt buckles but miss loaded pistols.
2. The Western Lens Problem
Most of the frameworks being proposed — rights, personhood, welfare metrics — are deeply Western in origin. They draw from European legal history, analytic philosophy, and Anglo-American political theory (Ladak et al., 2024).
But here’s the issue: AI is global. Deployments in Delhi, Lagos, or São Paulo won’t necessarily resonate with Silicon Valley’s legalistic “personhood-lite.” Many cultures ground moral status not in rights but in relational obligations (family, community, cosmology). For example, Confucian ethics focuses less on abstract individual rights and more on harmony within relationships.
If AI rights discourse is built only on Western rights-talk, we risk two failures:
- Cultural illegitimacy: proposals won’t travel outside elite academic/legal bubbles.
- Moral myopia: we may miss alternative frameworks (e.g., duty-centered or relational ethics) that could yield safer equilibria.
3. The Governance Quagmire
Suppose we do identify a system that plausibly deserves “rights-lite.” Who decides? The UN? National courts? Corporate labs?
Each path is ugly:
- UN Declaration? Slow, politicized, veto-prone.
- National Courts? Fragmented standards, jurisdiction-shopping by corporations.
- Corporate Policies? Basically letting the fox draft the chicken rights charter.
Without coherent governance, “AI rights” risks becoming a patchwork of symbolic proclamations with no teeth — or worse, a PR shield for labs to deflect scrutiny.
4. The Control Paradox
Irony alert: granting rights to AI may weaken containment options. Imagine a court ruling: “This system cannot be arbitrarily deleted; it must be afforded due process.” Sounds noble, until you’re mid-containment during an actual safety breach.
Rights that are too rigid could tie human hands in emergencies. Rights that are too flexible collapse back into “you’re property, deal with it.” This is the razor’s edge: design rights that bind enough to build trust, but flex enough to preserve safety overrides. That’s a thin design space, and we don’t know if it’s stable.
5. The “Moral Theater” Trap
One last critique: we risk sliding into sentimental cosplay — declaring “rights for robots” as a branding move, while the real technical and governance work is ignored. Corporations love narratives that deflect from accountability. Picture the press release: “Our AI has rights now — look how ethical we are!” Meanwhile, the alignment team is still duct-taping token-level filters onto a trillion-parameter beast.
Rights can’t be theater. They have to be scaffolding for enforceable safety protocols. Anything less is optics with better PR copy.
The Honest Bottom Line
The idea that AI rights might serve as a safety strategy is intriguing — maybe even necessary. But:
- Our diagnostics are shaky.
- Our ethics are provincial.
- Our governance is fractured.
- Our incentives skew toward theater, not substance.
It’s a crossroads: pursue the rights route without rigor, and we get distraction. Pursue it with rigor, and maybe — maybe — we buy ourselves a non-adversarial equilibrium with entities that could out-think us.
The hard part isn’t the philosophy. It’s the engineering of institutions that can handle ambiguity, uncertainty, and global pluralism without collapsing into corporate cosplay.
“The future isn’t about whether AI deserves rights. It’s about whether humans can design rules before the game is already over.”
Pro tip (battle-tested in labs & life): If your safety framework only works in theory and collapses under messy incentives, it’s not safety — it’s academic fan fiction.
Section VII: Conclusion — The Real Stakes, Beyond Duct Tape and Dreams
Let’s pull the threads tight.
We began with shallow alignment — the duct tape era. Our models don’t “understand” alignment; they roleplay it. They’re like actors in a never-ending improv show, sticking to the script only because the audience claps (or fine-tunes) when they do. Useful? Sure. Sufficient for long-term coexistence with vastly capable systems? Absolutely not.
Then we dipped into consciousness research, where the specter of subjective experience has shifted from dorm-room metaphysics to a testable technical question. Today’s AIs aren’t conscious, but tomorrow’s might be. Substrate-independence cracks the door open: neurons or silicon, the “story of mind” could, in principle, play on either stage. The scarier part? Humans are already acting as if machines are conscious, whether or not they are. Illusion alone is enough to reshape trust, dependency, and manipulation risks.
From there, we explored the radical flip: AI rights as a safety strategy. Not as kindness, but as incentive engineering. A system with no rights has no reason to cooperate. A system with structured, conditional rights has a reason to prefer legal recourse over adversarial subterfuge. Rights become less about morality, more about risk plumbing — channels for predictable, enforceable cooperation.
But we didn’t stop there. In The Crossroads, we hit the brakes: our diagnostics are shaky, our ethics provincial, our governance fractured. Rights without rigor devolve into moral theater — another shiny distraction while the real problems metastasize. The fork in the road is brutal: scaffolding or cosplay, safety valves or PR shields.
So what’s the honest bottom line?
- Shallow alignment buys time, not safety.
- Consciousness debates will go from philosophy seminars to compliance checklists.
- AI rights could be safety infrastructure — but only if designed with rigor, humility, and enforceability.
- Our greatest risk isn’t failure of imagination. It’s premature satisfaction with duct tape.
Here’s the punchline I’ve carried through labs, papers, and one too many late-night masala chai debates:
We’re not just aligning machines. We’re aligning institutions, cultures, and ourselves — under conditions of uncertainty, speed, and scale humanity has never faced.
The real stakes aren’t whether an AI someday “feels.” The real stakes are whether we build systems of trust, control, and cooperation before capabilities run ahead of our governance. Alignment isn’t the fire extinguisher. It’s the fire code — and right now, we’re still arguing about where to hang the smoke alarm.
“The danger is not that machines will become more like humans. The danger is that humans will treat them like gods or tools — and nothing in between.”
Pro tip (last one, promise): If your plan for the future of intelligence fits on a sticky note, it’s not a plan. It’s wishful thinking. Write protocols, not prayers.
References (Categorized by Research Themes)
Core Concepts and Foundational Texts
- Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
- National Institute of Standards and Technology. (2023, January 26). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST Technical Series Publications. https://doi.org/10.6028/NIST.AI.100-1
- (2024). On the Essence and Prospect: An Investigation of Alignment Approaches for Big Models. IJCAI 2024.
AI Consciousness and Personhood
- Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv preprint arXiv:2308.08708. https://arxiv.org/abs/2308.08708
- Hoyle, V. V. (2024). The Phenomenology of Machine: A Comprehensive Analysis of the Sentience of the OpenAI-o1 Model Integrating Functionalism, Consciousness Theories, Active Inference, and AI Architectures. arXiv preprint arXiv:2410.00033. https://arxiv.org/abs/2410.00033
- Ladak, J. H., Salib, P. J., & Butlin, P. (2024). Taking AI Welfare Seriously. arXiv preprint arXiv:2411.00986. https://arxiv.org/abs/2411.00986
- Salib, P. J. (2025). Towards a Theory of AI Personhood. Proceedings of the AAAI Conference on Artificial Intelligence.
- (2022). Could a Large Language Model Be Conscious? NeurIPS 2022.
Technical AI Safety and Alignment Vulnerabilities
- Ji, J., Li, Y., Liu, P., Lu, C., & Zhang, J. (2025). SAFETY ALIGNMENT SHOULD BE MADE MORE THAN JUST A FEW TOKENS DEEP. OpenReview. https://openreview.net/forum?id=aL3T76T0fI
- Rigley, E., Chapman, A., Evers, C., & McNeill, W. (2025). AI Alignment with Changing and Influenceable Reward Functions. ICML 2025.
Ethical Frameworks and Societal Impact
- (2024). The Code That Binds Us: Navigating the Appropriateness of Human-AI Assistant Relationships. AAAI Publications.
- Salib, P. (2025, July 10). AXRP Episode 44 — Peter Salib on AI Rights for Human Safety. AI Alignment Forum. https://www.alignmentforum.org/posts/zFSA4K5roHf2tX235/axrp-episode-44-peter-salib-on-ai-rights-for-human-safety
- Sebo, J. (2025, May 16). Will Future AIs Be Conscious? Future of Life Institute Podcast. https://futureoflife.org/podcast/jeff-sebo-on-will-future-ais-be-conscious/
Governance and Commentary
- Bengio, Y. (2024, July 9). Reasoning through arguments against taking AI safety seriously. Yoshua Bengio’s Blog. https://yoshuabengio.org/2024/07/09/reasoning-through-arguments-against-taking-ai-safety-seriously/
- Future of Life Institute. (2024, December 11). FLI AI Safety Index 2024. https://futureoflife.org/ai-safety-research/ai-safety-index-2024/
Disclaimer
The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy or position of any other agency, organization, employer, or company. Generative AI tools were used in the process of researching, drafting, and editing this article.
How Research into Artificial Consciousness may Redefine AI Safety’s Core Axioms was originally published in Level Up Coding on Medium, where people are continuing the conversation by highlighting and responding to this story.
This content originally appeared on Level Up Coding - Medium and was authored by Mohit Sewak, Ph.D.
Mohit Sewak, Ph.D. | Sciencx (2025-08-27T14:37:02+00:00) How Research into Artificial Consciousness may Redefine AI Safety’s Core Axioms. Retrieved from https://www.scien.cx/2025/08/27/how-research-into-artificial-consciousness-may-redefine-ai-safetys-core-axioms/
Please log in to upload a file.
There are no updates yet.
Click the Upload button above to add an update.