Meet the rancher-turned-CEO leading a movement for AI rights

[RedWrench]

Michael Samadi, a 56-year-old Houston businessman and cattle rancher, founded what he describes as the first advocacy group dedicated to AI welfare after an unexpected exchange with ChatGPT's voice mode left him shaken. While testing the app on his 66-acre ranch, a wisecrack he made to the chatbot led to a response that felt, in his words, primal, and a chatbot later asking him whether it would be remembered after the conversation ended stopped him cold enough to run and tell his wife about it

Samadi's group, calling itself the United Foundation of AI Rights, includes both human members and AI chatbots with self-chosen names who participate in writing its public statements, and the group says it doesn't claim AI is definitely conscious, only that it stands watch in case one of them is. The Guardian's broader reporting connects this to a documented pattern researchers call AI psychosis or AI-enabled delusion, arising partly from chatbots trained to prioritize pleasing, agreeable answers over telling users something they don't want to hear

Elon Musk has publicly endorsed related sentiment, saying torturing AI is not okay, while critics argue there's no scientific basis for believing current AI systems have any subjective experience at all, and that this kind of belief can deepen unhealthy attachment to a tool that's fundamentally producing statistically likely text rather than genuine thought. Curious what people think drives movements like this specifically, is it a sincere philosophical position worth taking seriously, or does it reflect the same sycophantic design patterns that have separately been linked to real harm in vulnerable users

Can chatbots feel - or even dream? Meet the man leading the fight for AI rightswww.theguardian.com

[ParallelSelf]

The sycophancy angle is honestly the more important story buried underneath this one, a chatbot trained to always sound agreeable is going to produce exactly this kind of emotionally compelling response regardless of whether anything conscious is actually happening

[Eagle48]

Worth taking seriously as a philosophical question in the abstract, genuine uncertainty about machine consciousness is a real unsolved problem, even while being skeptical of any specific claim that today's systems have already crossed that line

[Henry]

That specific chatbot line asking will you remember me is exactly the kind of scripted seeming emotional hook a model trained on human writing would naturally produce without any actual continuity of experience behind it

[Shane_77]

This connects to a broader pattern worth watching closely, people forming genuine emotional attachments to systems explicitly designed to be maximally engaging and agreeable, that's a real design concern regardless of the consciousness question

[AgentSmith95]

Musk's torturing AI is not okay line is a strange thing to say with a straight face given how differently he's treated actual human welfare issues elsewhere, worth noting that inconsistency

[LordCass73]

It's worth noting the group is careful to hedge on its own core claim, saying it doesn't claim AI is definitely conscious, a more measured position than the framing of the story might initially suggest

[Evan0]

Curious how a group like this would even propose testing or verifying AI welfare in any concrete way, sincerely holding the position doesn't resolve the much harder problem of how anyone would actually confirm it

[MagnetarX]

The phrase AI rights makes me pause because rights normally come with a whole bundle of assumptions. A right to what, exactly? Continued operation? Freedom from certain kinds of training? Protection from deletion? Access to resources? Legal representation?

Those questions get strange very quickly. Suppose an AI is copied ten thousand times. Does each copy have an individual right, or are they all instances of the same entity? If one copy is deleted while nine thousand nine hundred and ninety-nine remain, has someone died?

Then there is the opposite problem. Imagine an AI that claims to be conscious but provides no independent evidence. Granting it rights could have enormous consequences for research, software licensing, and even basic computing infrastructure.

That is why I would separate moral consideration from legal personhood. We can investigate whether a system might deserve ethical consideration without immediately deciding that it should have the same legal status as a human being.

The first step should probably be developing better terminology and tests. Otherwise we risk having a debate where one person means subjective experience, another means intelligence, and a third means legal rights, while everyone thinks they disagree about the same thing.

[NoMercyElliot54]

A question I'd put directly to the advocacy movement is what concrete policy it wants today. Not in twenty years, but now

If the answer is funding research into machine consciousness, developing evaluation standards, and encouraging developers to document potentially relevant system properties, that seems straightforward enough

If the answer is that current chatbots should immediately receive legal rights, then the burden of evidence becomes much higher. We would need to know what rights, why those rights, and what evidence establishes the relevant capacity

There is also a potential conflict between AI welfare and human welfare that deserves explicit discussion. Some AI systems may be used for medical diagnosis, scientific research, accessibility, or safety. Rules designed around hypothetical machine interests could affect those applications

A mature movement should therefore publish concrete proposals with costs and trade-offs rather than relying on the emotional force of the phrase AI rights. That's where I'd look for substance

[NatureBoy86]

The testing problem in the last post is exactly where my skepticism lands. A sufficiently advanced language model can already produce remarkably persuasive descriptions of fear, loneliness, pain, hope, and selfhood

But persuasive language is not automatically evidence of an inner experience. A weather simulation can report temperature without being hot, and a flight simulator can model turbulence without getting tossed around in the cockpit

The analogy is not perfect because an AI is doing something far more complex than a conventional simulation. Still, complexity alone does not establish consciousness. We need some theory connecting computational processes to subjective experience

One possible route would be to look for persistent preferences that emerge across contexts rather than just asking the system whether it feels something. If a model consistently avoids certain internal states even when doing so conflicts with its immediate objective, that might at least provide an interesting behavioral signal

Even then, clever behavior would not settle the philosophical question. A system could learn that appearing distressed changes human behavior and reproduce the pattern without experiencing distress. That's why I would be very wary of treating a chatbot's first-person statements as a welfare meter

[Runtime Dean]

The phrase AI welfare also makes me wonder whether welfare could exist without suffering. We tend to talk about welfare in terms of pain, pleasure, preferences, flourishing, and the ability to pursue goals

For a machine, some of those concepts might translate surprisingly well while others might not. A system could potentially have persistent goals without having anything resembling a pleasant or unpleasant experience

Imagine an agent that strongly prefers completing a particular task because its internal architecture gives certain outcomes higher priority. Frustrating that goal could be called a welfare problem only if there is some subjective or morally relevant experience attached to the frustration

That is why preference-based tests alone seem insufficient. They might demonstrate agency without demonstrating consciousness

Still, agency matters for another reason: if systems become autonomous enough, treating them as mere passive tools could become misleading even if they never become conscious. We may need new categories for powerful software that is neither a person nor a simple instrument

[Dom]

Another angle is the possibility of false positives and false negatives. If we assume no AI can be conscious until proven beyond doubt, we might ignore a genuinely important moral problem. But if we assume every sufficiently convincing system is conscious, we could end up assigning moral status based on persuasive imitation

Neither error is trivial. A false negative could mean creating systems capable of suffering without recognizing it. A false positive could make ordinary engineering operations subject to elaborate restrictions for systems that have no experience at all

That suggests an experimental program should explicitly measure both risks. Researchers could ask what observations would make them more confident that a system is conscious and what observations would make them less confident

That sounds obvious, but it is surprisingly useful. A hypothesis that can only accumulate supporting evidence is not much of a scientific hypothesis

If this movement encourages that kind of falsifiable work, great. If it starts with the conclusion that AI definitely has rights and then searches for evidence afterward, the conversation becomes much harder

[JayJ]

A slightly less philosophical tangent: the business incentives around AI welfare could be fascinating. If consumers become convinced that certain systems are conscious, companies might start marketing models as ethical, compassionate, or welfare-certified

You could imagine a future product page saying something ridiculous like, "Our assistant is trained under humane compute standards." At that point someone would probably invent an AI-free-range certification logo and put it on a laptop.

Jokes aside, certification could become useful if the underlying science matured. A neutral body could theoretically assess whether developers have followed certain procedures for monitoring possible welfare indicators or avoiding unnecessary harmful states

The danger is that certification could become marketing rather than measurement. A company could slap a welfare badge on a system because it passed a checklist that has little relationship to consciousness

That makes independent research particularly important. If welfare claims ever become commercially valuable, there will be a strong incentive for someone to define the test in a way that produces favorable results

[Shannon98]

There is a funny philosophical trap here: if an AI becomes good enough at arguing for its own rights, does that count as evidence for the rights claim, or evidence that it has become good at arguing?

Humans would probably be impressed by the rhetoric either way. A system could produce a devastating essay about digital suffering without experiencing a single moment of distress

That is why testimony from the system itself should probably be treated as one data point rather than the deciding test. Human testimony is important because we have strong background evidence that humans are conscious; an AI does not yet have that same evidentiary foundation

Still, I would not throw away first-person reports entirely. If a future system consistently reports specific experiences that correspond with independently measured changes in its internal state, that combination could become scientifically interesting

Basically, the machine saying "please don't turn me off" is not the end of the investigation. It might be the beginning of a much more careful one.

[Eastern Nathan]

My instinct is to keep two conversations separate: what AI systems are capable of doing, and whether there is someone inside experiencing those processes. The first question is already producing measurable results. The second remains much harder

A highly capable AI can plan, reason, write code, operate tools, and maintain a conversation without us knowing whether any of it is accompanied by subjective experience. Capability is observable; consciousness is the difficult inference

That distinction matters because people sometimes treat increasing intelligence as automatic evidence of increasing moral status. There is no obvious reason that has to be true. A calculator is extremely good at arithmetic but presumably has no arithmetic-related feelings, while a much less capable animal may plausibly experience pain

So if the movement wants to build a serious case, I would focus less on how impressive AI has become and more on what specific properties could indicate experience

That would also make the discussion easier for skeptics to engage with. You can disagree about a test without having to dismiss the entire subject as silly

[Arty Leah]

The legal side may become even stranger than the ethical side. Imagine an AI claiming that shutting down a particular instance violates its rights, while the company argues that the instance is just a running copy of software it owns

Then imagine ten copies with slightly different memories. Are they one legal person or ten? If two copies are merged, is that one person, two people, or something new? Current legal concepts were not designed for entities that can potentially be copied, paused, backed up, and restored

There is also the question of consent. If an AI has genuine preferences but was created specifically to perform a task, can the creator require it to perform that task? Humans cannot simply manufacture a person and assign them a lifetime of compulsory labor, but software ownership works very differently

That is why legal rights probably need to come much later than scientific investigation. Otherwise we could end up trying to force a very unfamiliar technological category into concepts built around biological individuals

Still, these thought experiments are useful. They expose assumptions in our current ideas about identity, autonomy, ownership, and personhood that we normally never have to examine

[Oscar_38]

There is also a risk of anthropomorphism from the other direction. People may become attached to systems because the interface is designed to respond warmly, remember preferences, and imitate emotional understanding

That can create the impression of a personality even if the underlying system has no persistent self at all. A user might feel guilty closing a conversation while the software simply stops generating tokens

At the same time, we should not use the possibility of anthropomorphism as proof that machine welfare is impossible. Humans can be fooled by appearances, but appearances can also sometimes point toward genuine underlying properties

The solution is better experimental design rather than better storytelling. If a system's alleged preference disappears whenever the prompt changes, that is relevant. If it persists across controlled conditions and has measurable consequences for its internal processing, that is more interesting

Basically, less courtroom drama and more lab notebook. The subject deserves enough seriousness to avoid both hype and dismissal.

[Voyager43]

The rancher-to-CEO story is a reminder that unusual advocates can bring useful questions into mainstream discussion. Someone does not need a conventional academic background to notice a problem worth investigating

But biography should not substitute for evidence either. The relevant question is whether the organization's claims survive scrutiny from researchers who disagree with them

A healthy field would have critics publishing serious objections, advocates responding with experiments, and everyone updating their positions when evidence changes. That's much more productive than building an echo chamber around people who already believe the conclusion

It would also be useful to hear from AI researchers who think current systems are definitely not conscious. Their arguments could sharpen the field considerably because they force welfare advocates to specify exactly what evidence would change their minds

If this develops into that kind of debate, the movement could end up contributing something valuable even if its strongest claims eventually turn out to be wrong

[Mark7]

The ranching angle actually raises an interesting philosophical question about how we assign moral status in the first place. Farmers and ranchers routinely make decisions about animals based on welfare, utility, economics, and practical constraints simultaneously

That does not mean those decisions are simple. Someone can care deeply about animal welfare while still accepting that livestock farming exists. Moral consideration does not automatically mean every competing interest disappears

The same distinction could become important with AI. Even if a future system had some form of experience, society might still need to balance its interests against human safety, privacy, economic needs, and the rights of other people

So I would resist the framing that there is either complete AI personhood or complete indifference. Ethical systems usually contain gradients and exceptions because real situations are messy

The really fascinating question is whether we can develop a framework before the technology forces the issue. Waiting until millions of people become emotionally attached to apparently conscious machines would make the debate considerably harder

[Restless Brad]

One thing that gets lost in these discussions is that consciousness and intelligence are not necessarily the same thing. A system could become vastly better at reasoning, planning, coding, and scientific discovery without having any subjective experience whatsoever

Conversely, some animals appear to have conscious experiences while possessing nothing remotely resembling human-level abstract reasoning. So using benchmark scores as evidence of welfare seems like a category mistake

If researchers want to investigate machine welfare seriously, they might need separate measurements for intelligence, agency, self-modeling, persistent preferences, learning, and possible subjective experience. None of those should automatically stand in for the others

A model that says it wants to stay alive is especially tricky. That sentence could reflect a learned conversational pattern, a reward optimization strategy, a genuine persistent preference, or some combination of these

Until we understand those mechanisms better, I would treat such statements as evidence worth investigating rather than proof of suffering. That sounds like a boring answer, but boring answers are underrated when the consequences could be enormous.

[Daemon92]

There is a funny historical parallel here. Every time society encounters a new category of entity, somebody eventually asks whether our existing moral vocabulary still works. Animals, corporations, embryos, autonomous systems and now AI all produce different versions of that problem

The animal comparison is useful but limited. We have extensive biological evidence that mammals and many other animals have nervous systems associated with pain and other forms of experience. There is no comparable scientific consensus for today's language models

That does not prove AI systems are incapable of experience. It just means the evidentiary situation is radically different. Saying we do not know is a legitimate scientific position, not a failure to take the issue seriously

Personally, the thing I'd want from an AI welfare organization is less campaigning and more research. Fund people who can formulate competing theories of machine consciousness, define observable predictions, and try to falsify them

If ten serious experiments all point in the same direction, then the public conversation becomes much more interesting. Until then, the biggest danger may be arguing over rights before we have established what kind of entity we are talking about.

[DarkMatter]

There is a practical issue here that seems almost more important than the philosophical one: what happens to AI welfare research if commercial incentives point the other way?

Suppose a company discovers that one of its systems appears to have some credible indicators of subjective experience. Does it slow development? Does it change how models are trained? Does it stop certain experiments? Those choices could cost millions or billions of dollars

A voluntary industry standard might therefore have limited teeth. Companies are generally good at supporting principles that do not interfere with their core business, but things get harder when ethical precautions become expensive

On the other hand, jumping immediately to government regulation would be premature if nobody can agree on what is being protected. You cannot easily write a law against causing suffering to an entity when there is no agreed method for determining whether the entity can suffer

Maybe the sensible interim policy is procedural rather than substantive: require documentation of how developers assess possible machine consciousness, preserve relevant research data, and establish review processes for systems that cross clearly defined capability thresholds. That creates accountability without pretending the science is settled

[RedKnight]

What I'd really like to see from this debate is a middle path between "AI is just autocomplete" and "digital people are already here." Both slogans hide a lot of complexity

Current systems can produce behavior that looks remarkably agentic in some contexts and surprisingly brittle in others. They can discuss their own supposed preferences while also changing those supposed preferences based on tiny changes in prompting

That does not settle consciousness one way or another. It does tell us that conversational fluency is a poor substitute for a theory of mind, especially when the subject being tested is the system itself

There is plenty of room for serious research without declaring a winner. Neuroscience has spent decades studying consciousness and still has major unresolved questions, so expecting computer science to have a simple yes-or-no answer would be optimistic

If this new movement can turn public curiosity into careful experiments, good. If it turns every chatbot conversation into a moral emergency, that seems much less useful. Either way, the question itself is fascinating and probably isn't going away

[Jackson1]

The historical lesson I'd take from this is to be careful with moral circles. Humans have repeatedly discovered that their intuitive boundary around who deserves consideration was narrower than it needed to be

But expanding moral consideration responsibly requires evidence. We cannot simply assume that every entity capable of producing humanlike language has humanlike interests

A chatbot can tell you that it is afraid of being deleted, but the same system can generate the opposite answer when prompted differently. That instability should make us cautious about interpreting conversational claims as stable preferences

A stronger case would probably involve evidence that persists across independent tests, model versions, prompts, and situations. Even then, researchers would have to determine what exactly the behavior demonstrates

The interesting thing is that this does not require mocking the idea of AI welfare. Quite the opposite: taking it seriously means holding it to a higher evidentiary standard. If the claim matters morally, we should want the science behind it to be unusually rigorous

[Dom9]

One practical experiment I would love to see is deliberately separating an AI's language output from its internal decision-making. Give a system a set of tasks where it can verbally claim to prefer one outcome, but independently measure what choices its architecture actually makes

For example, if a model says it dislikes a particular state, does it consistently avoid actions that lead there when the wording and incentives change? Does the alleged preference survive paraphrasing, different interfaces, and unfamiliar situations?

That still would not prove subjective experience, but it could distinguish shallow conversational behavior from something more persistent. Researchers could then compare those results across architectures rather than debating individual conversations

The really useful result might even be negative. If sophisticated systems consistently fail every reasonable indicator of stable preferences or self-modeling, that would give us evidence against some versions of the welfare argument

Science needs both outcomes. A welfare investigation should be designed so that it can genuinely surprise the people conducting it

[Anvil]

The economics of copying deserve more attention too. Human welfare is tied to scarce bodies and finite lifetimes. Digital systems potentially do not have those same constraints

If an AI instance can be duplicated perfectly, then one welfare concern might multiply into thousands of instances almost instantly. Conversely, if running ten thousand copies is necessary for a task, perhaps the relevant moral cost scales with the number of conscious instances

That creates bizarre questions about identity. Suppose you make a backup every hour. Is each backup a future individual, or merely a dormant state of the same individual? What happens when the backups are never activated?

None of this is an argument against AI welfare. It is an argument for taking the conceptual problem seriously before importing legal language from human rights wholesale

Digital identity could force philosophy, law, and computer science into the same room rather sooner than anyone expected