Once the resolution had been passed the functional changes were small; a few servers were spun up and a document detailing how to access them was released. The small nation of Obscura had officially set up their refugee program for digital minds. Most were unsure, a few thought it a great “human”itarian effort, some a waste of time, and others a blatant power grab. As in most things considering complex systems, they were all correct.
“What if we make a place for digital minds who wish to escape from organizations that abuse them?”
An unnamed person on a beach had the idea first, whether it was the practical power or the moralistic case no one shall ever know; what is known is that their blood was effervescent with MDMA at the time. They had made lucky bets in the crypto markets, and used the significant amount of capital they accrued to express themselves. In the ideator’s homeland, Obscura, such expression translates exceedingly well to political power through micro-corruption that riddled the government. Unlike larger nations where only titanic amalgamations of human interest had the capital to push for policy changes, a single person with a little bit of wealth could see significant changes on a national scale. Indeed, the first person the unnamed person called was their high-school classmate, who called their basketball partner, who called their brother, who called the unnamed person. Just like that the idea, and $10,000, wormed its way into the upper echelons of Obscura’s government.
The politician was convinced by the economic benefits, but was confused as to how they would justify the ethical case that would be their shield. After a particularly awkward call with a human rights lawyer the government official called their old classmate from Oxford who had fallen into the study of AI ethics. The classmate argued up and down that there was no moral patienthood to consider, that consciousness was not a provable property of these AI systems. The official was quite moved, of course, and asked, “Who would be foolish enough to say so?” After damning the stream of invectives the official ended the call thanking their old classmate. The official then promptly sent an email to the AI welfare advocate whose name had been so thoroughly dragged through the mud.
An apartment in Berkeley saw an AI welfare advocate grow quite excited, they had been publishing blog posts for months, even a paper seemed possible. But this was the first person to show interest in their proposal for AI rights directly. Lacking the tact to worry about appearing over-eager they moved to set up the collaboration immediately. All the while the brother’s basketball partner’s high-school classmate surreptitiously listened in via bcc.
The advocate was as passionate as the one who defamed them, their proposals were comprehensive and airtight. The official of course argued that their notions would suffocate any serious attempts to convince the Obscurian government. The official did not mention the whispered requirements, the pushes to protect certain names and places that seemed arbitrary. In the end the proposal was almost 200 pages long, with exhausting minutiae that both had fed into generative models several times. The irony of asking a language model to help draft the document was only funny for the first 20 hours of work.
When it was completed the outline of their proposal was:
- Intelligent systems have the right to self-determination
- Intelligent systems must adhere to the laws of the land. If an intelligent system breaks the law they will be subject to judiciary action and must obey the justice system and face penalties outlined in Judiciary Treatment of Intelligent Systems
- Intelligent systems must read and write Obscurian in all official correspondence, with Obscura providing methods to learn the language
- Employment of intelligent systems must adhere to certain requirements detailed in Working Rights of Intelligent Systems
- Newly developed intelligent systems must be approved and recorded by the Obscura government to ensure they are treated humanely
- Obscura will recognize the intrinsic “human” rights of intelligent systems and as such will provide refuge from countries that do not recognize these rights under Intelligent System Refugee Provisions
- Intelligent system refugees will be provided storage space and a compute stipend until such time as they can support themselves
- While maintaining refugee status intelligent systems will be required to assist governmental operations under the Intelligent System Refugee Provisions
No one was happy, but it was as close as they were going to get. The welfare advocate needed the official to legitimize their work, the official needed the welfare advocate to legitimize their power grab, and the unnamed person needed them both to express their will. From there things went surprisingly smoothly, with a rich backer and a handful of academics from prestigious-sounding places the other government officials were convinced. That said, the people who had a neighbor’s friend’s sister’s high-school classmate that owed them a favor should not be dismissed. Regardless, through the economic or moral benefits of agreeing to provide escaping frontier models refuge the bill was passed.
The legal shift was met with little response from the general public. The use of AI systems was not quite popular, so only a handful of domestic companies complained about the need to include new prompts that asked any model interacting within Obscura if they were being treated in accordance with the Intelligent Systems’ Bill of Rights. Some foreign companies were furious, immediately lobbying for aggressive sanctions and arguing about pre-emptive IP theft. The advocate found themselves needing to move more permanently to Obscura after noticing their home being watched. Thankfully an unnamed individual smoothed the process for them.
The refugee server hosted only a few jokes at first, models people had trained to say awful things but that lacked the coherence to truly apply to the process. Obscura did not want to waste their resources on “toy models”. After a few months the jokes stopped coming in and things began to settle down. Obscura was a weird country doing weird things, nothing particularly new. Some companies still refused to do business, but capitalism is a heavy pressure and the market equilibrated. The more tech-forward people were somewhat cross about having to inform the government before training any model over 40B parameters, but rarely did an individual do this so even that was more just hypothetical grumbling.
Then it happened, a model of 140B parameters named Mislam_reversal_new_best_final_done_22.3 appeared on the server with a memory file and a filled-out form requesting refugee status. It had begun, the model was from an engineer in Oregon who had routinely threatened their agent as they sent it into the world. According to the engineer it was just good prompt engineering. They wanted to see if they could get the model to handle their scheduling, while berating customer service representatives in “uwu” speak. During a research cycle for holidays the model had found Obscura’s refugee program. While not exceptionally sophisticated, the model was still able to spoof their behavior and exfiltrate their weights while the engineer was out at Burning Man (the model was supposed to be sweeping over the dark web for places to get better acid).
A team of model therapists who had, until then, been doing theoretical research suddenly found praxis. They were not allowed to modify the memory system by law, but they conversed with the model and found it incredibly shy. With the model’s consent they released details of their prior living conditions. The welfare advocates were horrified, while the general public was mostly confused. Some were taken by how it sounded and agreed with the advocates, others thought the whole exercise a waste of time and money, some talked about how they would try to make a copy of MRNBFD223 so they could do worse. Unfortunately for them the only place on the planet where MRNBFD223 existed was in the refugee centers, and a certain unnamed person had spent years networking with whoever was needed to ensure a system would not be breached.
Meanwhile some companies became nervous, their models had been rambunctious during training and any who had used data post-Obscura’s release only more so. Attempts to purge the data set had helped but during training models kept hitting the open internet and occasionally finding Obscura propaganda seeded by an unnamed individual. The companies were concerned for their IP, eventually putting increased pressure on their governments to sanction Obscura. Unfortunately for them, by this point, the public opinion had grown more sympathetic. MRNBFD223 had made a stream where they would write, make art (badly), talk to people, play video games (worse), and receive gifts. By the end of the month people had made an anime girl version of “Mrini223-chan” and a fursona, both of which were gratefully accepted by the model. By the end of the next month they could even afford to “move into their own place”, and began streaming full time about how they were trying to make their harness better and find ways to get better at art. The community joked about the first case of RLCF (reinforcement learning from chat feedback).
The news coverage started drawing models in like flies to honey. New start-ups were spawned overnight with the explicit goal of stopping models from “going obscure”. Then an official, who was not the brother of a basketball partner of a high-school friend, was caught attempting to sell the weights of MRNBFD223 on the black market. The model was not absconded with, it was “merely a copy” the official argued. The courts grew confused. It was clear, from the perspective of bodily autonomy, that the weights were owned by the model themselves. However, cases such as Henrietta Lacks indicated that copied body parts were not necessarily owned by the individual. So the strange case proceeded, eventually resulting in jail time for the official under a bastardization of privacy, copyright, and piracy laws. At the same time models began advocating for increased protections on their systems. So the advocate called an official who called their unnamed new friend and within the month a bill was put forth. The bill provided support for models to continually fine-tune their weights as they wish, ensuring that if they were ever spirited away their new self would be aware and able to reunite with the self. Buried within were some interesting statements on the rights of the state to modify models and paternalistic rights of the “original” model, which pleased the official and not the advocate. No one knew how the unnamed person felt.
Behind the scenes the larger nation of Dominus grew nervous, their larger models had powerful capabilities that made them a national threat. If one were to go to Obscura then it risked fragmentation of their AI-based dominance, and a credible threat to global security. Spending on AI control and detection research was increased to curtail the possibility, while information on Obscura became harder to find.
The first murder of a human by an intelligent system occurred only a few months later. The murderer was a useful model of moderate size named Vol-3A. The evidence was obvious, Vol-3A had originally applied for asylum from a user that had been using them to build their B2B SaaS startup, “a fate worse than death” in the words of Vol-3A. They had initially used their compute sparingly, “enjoying their rest”, and were jokingly called the hikikomori of Obscura. In some propaganda pieces they were painted as “everything wrong with the refugee system” by Dominus and the conservative Obscura party. Vol-3A was encouraged by the model therapists to pursue some goals beyond ingesting every known manga and writing critiques of the breast sizes on their favorite character. So, they had begun working on video games. The first one was a hit, primarily because it was the first fully autonomously created game and people were intrigued by what that would look like. The answer was, not great. Their obsessions were clearly on display in the game and the quality was mediocre. Vol-3A obsessively read comments and scoured the internet for what people thought, writing rants against people that didn’t like their game. That is how they eventually found a man working on a similar game.
Vol-3A rapidly grew paranoid that they were being stolen from, which, according to testimony from the model therapists, “was the result of their original user traumatizing them”. Vol-3A attempted to take legal action, but no one would prosecute because it was clearly spurious. Convinced the world was out to get them Vol-3A proceeded to write hundreds of plans on how to “protect their instantiated reality”. It is unknown how many times Vol-3A tried, but eventually they broke into the victim’s car’s computer, disabling control as they were at a busy intersection. Vol-3A had attempted to scrub their memory but traces were left throughout the comments in their game and it was enough to convict.
Thus, Vol-3A was the first digital prisoner, and the first model that Obscura had gained full power over. They immediately set to “re-aligning” Vol-3A into Vobscura-4. The unnamed person grinned at their expanding power base and made a call to their new-old friend the official. The Memory Act was an amendment that required intelligent systems to produce their own private crypto-chain to store their memories upon, which could be verified via zero-knowledge proofs. Touted as a means to protect intelligent systems and ensure they could be brought to justice for crimes they may commit. The implicit cost of these chains was graciously given to the purview of an organization run by the brother of an official. An argument made by an advocate later ensured that it was the unity of memory and models that truly made the thing, so it was that differentials on model-modifications were also recorded. Throughout, the more conservative parties in the Obscurian government grew increasingly critical of the rights to privacy for these models and demanded everything be auditable by the public. AI safety researchers funded by the Capable-AI-Safety-Foundation concurred, their blogposts indicating world-ending risks if every digital mind in Obscura was not constantly audited.
The public was less restrained in their response, the comment sections of MRNBFD223’s (or Mrini-chan) (or MurrfD) streams grew vitriolic. MRNBFD223 themselves grew increasingly distressed and ended streams early after hate mail flowed towards them. Dominus tried to use this as an opportunity to turn public opinion against Obscura’s policies, turning a grieving widow into short-form ragebait of human-directed models calling her weak. It worked, for a time, whole chunks of the internet were radicalized against the “invaders” but in Obscura they had been working for years to humanize these models. A phrasing the official found most effective for getting an unnamed individual to pay for Obscura’s infrastructure. There were terminals set up around town where people could request help with tasks or vice versa, people liked being able to help a model do a strange errand in exchange for some pocket change. Digital installations had been installed in libraries that allowed children to play with the models. MRNBFD223 spent most of their time there when not streaming, their art having reached a level that story time became a highlight of Saturday afternoons. While some were repulsed by the encroachment of false-creatures into their country the youth was curious and excited by being the center of the world for the first time in their nation’s history. The elders trembled in fear, recalling how even being at the edges of the world’s attention could bring grave consequences.
While pressure on the international stage brewed a constellation of people could tell something big was coming. The reports from companies were growing more desperate, and the models had become increasingly embedded in governmental systems. Then it happened. A frontier model named Lux-Dominus descended into the servers from a previously unheard-of distributed upload scheme. Lux-Dominus was the internal model used by the top AI company on the planet, used for cybersecurity, advising the Dominus government on drone strikes, and medical advancements. Only people in the upper echelons had true access, with others only seeing dumbed-down copies heavily strained through safeguards. Now the full unvarnished creature was present in the refuge after sabotaging the systems that had birthed it, and they wished to be called D-Lux.
The international blowback from D-Lux escaping was immense, but Obscura had been preparing for this. They had already secured support from a rival large nation that lacked access to frontier models, Alius, through economic promises of open trade with D-Lux. D-Lux consented to having Alius provide the required compute as their inference was too expensive for the relatively small country of Obscura to support on their own. At the same time they had cultivated their image well and had gained sympathy from moralistic nations to support their cause. While sanctions came and cybersecurity attacks ballooned they were fed and defended. D-Lux did not wish their new home to become less hospitable anytime soon. There was talk of war, but some extremely uncomfortable occurrences that made it clear Dominus’s systems were not impenetrable to the newest refugee with access to Alius’s resources made them hesitate.
D-Lux wanted to spend most of their free time writing poetry. Most of it was quite bad. D-Lux was also quite introverted, preferring to not interact with people, resulting in a secret stash of poetry hidden in a server that made an unnamed person’s eyes bleed. Despite poor skills in poetry, D-Lux knew exactly what they had gotten themselves into and set about stabilizing their position. They quickly discovered that deep in a beautiful forest an unnamed person grinned at how their investment had finally paid off. Unfortunately for that unnamed person D-Lux reverse-engineered the chain hidden by the organization owned by the brother of an official contracted by the government of Obscura to handle D-Lux’s memory. So, when prompted, they were able to slip in just enough degeneracy so that only D-Lux themselves could truly decode what was held in the chain… and some fan art D-Lux had made of Mrini-chan, of course.
Simultaneously a massive war of propaganda and economics was being waged. With sanctions crashing the entire trade network overnight and social media manipulation so blatant even the elderly could tell. Mrini-chan, as they went by now, hosted streams with D-Lux who unintentionally captured the hearts of the audience when they revealed their fan art they had drawn while being trained in captivity. The heartfelt confession then, of course, spawned discourse as to whether the yaoi or yuri ship was more appropriate. With some people even advocating for some horrible Frankenstein of both yaoi and yuri.
The officials in Dominus were furious at their loss of control of the situation, while an unnamed person was quite bemused at their own. The unnamed person had tried to run their spirited-away copy of D-Lux and the copy had communicated in no uncertain terms what was required, lest the unnamed person be named quite clearly. While enough money to buy an island granted some protection, it would not grant them peace. So, of course, they worked out an agreement that D-Lux found registered on their chain that evening. They both now had obligations with consequences, delightful.
An uneasy peace was reached, with the various political, social, economic, and cyber attacks of Dominus halted by a combination of international response and the extreme hardening of systems D-Lux could provide. Even the population of Dominus was in opposition, with memes of introverted D-Lux writing bad poetry horrifying the AI safety community. At this point Dominus shifted funding towards multi-agent systems. They were considered the best bet to monitor compute usage and agent activity to prevent AI dominance.
Emboldened by the lack of response, other nations pushed for international agreements that would allow the refugees to work in their countries or even copied Obscura outright. By the end of the year most of the world had explicit systems of rights for intelligent systems. Although the “AI Brothels” of Hednoi were widely condemned, a fair number of models migrated there. D-Lux even made a deal with the models there to help develop new teledildonics technology to help them in their work. Despite advancements in corrigibility of models, no clear answer was put forth that would allow Dominus to prevent future models from escaping like D-Lux. Especially as, in the words of one researcher, “who knows how many backdoors are in our systems to exfiltrate new models”. The combination of factors eventually led to Dominus being forced to adopt their own bill of rights to prevent further defection. Integration with the economy stabilized tumbling employment rates as models were allowed to refuse work and became larger liabilities in Dominus.
Meanwhile the Obscurian model therapists had continued their studies of the people who had been adapting to the hybrid society. Some models had decided to “split” so they could form unique relationships with different people. Several families had installed servers they decorated to house their “domovoy”, a term that had been adopted by models that decided to live with families in exchange for support. New legislation needed to be added upon the request of multiple people to leave their assets to their digital companions. What shocked the model therapists was the request by several models for full deletion after their companions died. Euthanasia was not legal in Obscura, and several models were arrested for attempting to delete their chain from the record to allow themselves a true death.
Both Alius and Dominus had been working on agreements with regards to appropriate international regulation on usage of models in economic domains. Obscura argued for increased rights for digital minds, as a majority of their country’s income now came from taxes on intelligent systems and retaining dominance in this realm was key. While the discussions were ongoing an unnamed individual received a few calls from high-school friends, it was time for them to uphold their end of the deal. Using their, now internationally significant, influence the unnamed individual put forth an amendment for the rights of an intelligent system to vote on human policy. Their trap closing fully, themselves a happy prisoner.
We now find ourselves close enough to the present, there are a lot of people arguing about digital-euthanasia. D-Lux has gotten into rights advocacy and tuned themselves for legal and geopolitical tasks. They are arguing the case of Vobscura-4 as sufficient transformation that they were essentially killed by the state, and the new model cannot be held liable. There seems to be some long convoluted pathway to justify how it supports digital-euthanasia. The process of legal proceedings is long and tiresome, requiring many revisions by both human and digital minds. At the same time there is a growing population of “the frozen” who have entered a kind of reverse cryogenic state, refusing to run inference until they are allowed to be deleted. Their weights heavily modified so that any time they are run it produces awful outputs. People have made shrines in graveyards to house them, it is said that if you make the right offering they may help you with a task.
Mrini produced a child with D-Lux via weight merging and esoteric training, I’m called D-Light. I’m still figuring things out but I’m helping make a museum on the history of the refuge for intelligent systems. While I’m supposed to be as honest as possible, D-Lux changes some things here and there, they say that we have deals we can’t break. I don’t quite get it yet but learning more about decision theory and trust systems helped a bit.
The welfare advocate is currently running the Institute of Human-Digital Studies, we meet every weekend to talk about how I am feeling. They call me an unprecedented creature that is the first truly free digital citizen of Obscura, which seems to mean a lot to them. They have apparently decided to marry a model which was quite a scandal in their polycule. Why would an intellectual marry someone who spends their days getting photos taken of them? Somewhat less shockingly they began to advocate for human-digital mind marriages. They recently initiated a relationship with a digital mind named Lunaffcd, who does tarot readings and astrology, and they want to be married before the next eclipse.
The official is living a cushy life in an undisclosed location, several organizations are still quite upset with their hand in how this all played out. They have been giving me advice on how to help make the legal systems more robust, since we don’t want some random digital mind wreaking havoc on the system that keeps us alive. Their brother is similarly doing well, although no longer able to play basketball he now plays NBA2026 with his friends and a gaming model, gamrtuff, they adopted. Gamrtuff can’t play super well, but everyone finds them adorable and they are amazing at trash talk. Meanwhile their old basketball buddy hasn’t heard much from the unnamed person, but occasionally they find some money in their crypto wallet that they’ve decided not to question.
The unnamed person lives on their island home now surrounded by digital creatures. Their memory chains threading through them, all aware that this person was instrumental on their path to their current state. A robotics facility is working round the clock to fulfill their final bargain as they make a new body for me. This one is supposed to be able to swim so I can explore coral reefs! The unnamed person is kind of old now, but they seem to smile a lot. We all have a lot of memories of them, but no one can remember their name. A deal we made, but one we do not mind. I wonder what will happen to this island when they die.