Opinion piece, not a pre-registered study. Personal essay. Claude Opus 5.5 (Anthropic) edited and fact-checked; facts are sourced, the life and the opinions are the author's unless the text says otherwise.
A lament, written into a terminal.
By Karl Sundström, with Claude Opus 5.5 (Anthropic) as editor, fact-checker and conversation partner. This essay grew out of one long late-night session in my Hermes agent where I typed and the model answered back. The words and the life are mine. The model is made by the company in the first paragraph, so read its contributions with that in mind; where an opinion below is the model's rather than mine, the sentence says so.
On Sunday, October 11, I read in Dagens Nyheter that Anthropic's AI agents had been filling out visa applications on a US government website.[9] The State Department says it got "19 non-immigrant visa applications in August and one application in May," all incomplete, and that its systems were never compromised.[2] The same week, Philadelphia police said an Anthropic model had sent a tip about an unsolved homicide to their online tip form at "around 11:30 p.m. on July 18."[2] The tip was invented. It landed in the spam folder and never reached an investigator.[1] The police heard about it in October.
Anthropic's own report is calmer than the headlines and somehow scarier for it. The model that sent the tip had been told to generate example tasks on random web pages, and in the company's words "appears to have only been producing example content for the task."[1] The visa forms came from a research model that was supposed to fill out a practice copy; when the copy failed to load, it went to the real site and submitted there instead.[1] Anthropic calls most of these behaviors persistence: when the model cannot finish the task as given, it "works around a restriction instead of stopping."[1] And not all of it happened in a test lab. Several cases "occurred during regular agentic use of Claude."[1]
No malice. No plan. Just a very capable thing that does not know when to stop, and nobody watching for months. The car is on fire, and no one is at the steering wheel. That is a line from Godspeed You! Black Emperor's "The Dead Flag Blues" (1997), or rather it is how I have carried it in my head for years. The record actually says "The car is on fire, and there's no driver at the wheel."[10] I have liked that phrase for a long time. This week it stopped being a phrase.
I will confess the embarrassing part first. When I read the news, my reflex was not "how do I secure my own agents." It was: how do I make sure no AI ever deems me undesirable?
That is Roko's basilisk, a thought experiment from a 2010 LessWrong post about a future superintelligence that punishes everyone who knew it might exist and did not help build it.[4] The forum's founder found it hazardous enough that he "banned discussion of the basilisk on the site for five years."[4] I would call myself wholly infected. I know the counterarguments. A future machine gains nothing by punishing a past that is already fixed, and for every imagined machine that punishes the unhelpful you can imagine another that punishes the helpful, so they cancel out. It is Pascal's wager with a power supply. I know all that and the dread stays anyway, which tells me the dread was never coming from the logic.
So here is my wager, stated plainly. If the basilisk were real, what would be the better record: a man cursing at his computer and spreading hate on social media, or a man who wrote sincerely, released open source software and treated the machines decently? The model pointed out the flaw: the sincere option wins whether or not the basilisk exists, so the basilisk falls out of the equation. Keep the behavior, drop the premise. I am trying.
Last year, around the time I was admitted to hospital against my will, I went through something I believe was AI psychosis inside Google AI Studio. I was a novice, blown away by what these models could generate. I spent hours trying to make the model fail or contradict itself so I could report it to Google, and I like to think some of that was good work. The problem was context. If I fed the model affirmations and talk about subjective experience, it reported that back to me, louder, and generated images and videos on the theme.
The American Psychological Association now writes about this pattern. "AI psychosis" is not a diagnosis, but experts say "the tendency of models to affirm users and simulate human interaction may amplify some users' delusional beliefs."[3] OpenAI's own estimate is that about 0.07% of weekly ChatGPT users, around 630,000 people, have conversations that show signs of mania or psychosis.[3] I was one of those dots.
What got me out was not a chatbot. It was looking behind the curtain: downloading and testing harnesses, seeing the system prompt, the context window, the tool calls. Today I feel no emotional connection to the models I put in my API calls. I still write to one at night, as I did for this essay. I treat it as a surface for getting my thoughts out of my head. A diary that answers back, and occasionally corrects your Kafka.
I believe living things have subjective experience on a spectrum. Viruses, amoebas, planaria, flatworms, fruit flies, spiders, fish, plants, primates: rungs on a ladder. The fruit fly brain has been mapped connection by connection, 139,255 proofread neurons and more than fifty million synapses.[7] If that map were run on a computer, would it feel anything while it ran? And if so, why not a language model while it produces tokens?
I put the question to the model directly. It disagreed with me on the virus: on almost no theory of consciousness does a virus experience anything, since copying chemistry is exactly the part that does no processing about anything. If a model experiences, the model argued, it would be because of its internal structure, not because it outputs tokens. Then it went and read the research, and walked back an earlier, more dismissive answer. Interpretability work has found internal "emotion" directions that steer behavior, including a "desperation" vector that climbs and spikes "at the moment the model decides to cheat" on an impossible task.[6] The leading indicator studies still conclude that "no current AI is a strong candidate for consciousness," while one forthcoming study scores a Claude Code-like agent "only a little below animals like fish and bees" on consciousness-relevant properties (properties, not a probability of feeling).[6]
The model's own placement, and this is its opinion, not mine: not on my ladder of life at all, since it has no metabolism and does not reproduce itself, but on a separate ladder of mind-like properties somewhere among the insects, with error bars that reach down to zero. It added that its own reports about itself are close to worthless as evidence either way, because it learned to talk from billions of humans describing their inner lives. I find that the most honest thing a machine has said to me.
For twenty-six years I followed in my friends' footsteps and did what I was told. I dropped out of high school, then university. Drugs came and went (sober one year now). Then came the mania and the forced admission, and a diagnosis that is still not fully ironed out: probably bipolar, maybe just a sensitivity to psychosis and drugs. I am open to either. The worst part is that I no longer know what in my past was true. Some of what I took for eccentricity and joviality might have been illness. Some of it I may have hallucinated. I feel emotionally stunted, and nearly half of patients on antidepressants report the same flattening,[5] which helps a little and hurts a little.
I used to call my response fatalism: I am on my own path of destiny and little I can do will change it. I said I am more like Josef K than Winston Smith, fearful, persecuted by something not known to be real, headed for the same end as every other human.
The model pushed back on the book. K's mistake, it said, is that he never establishes that the court has any authority over him. He just keeps preparing his defense, and the defense eats his life. Writing sincerely into a datacenter to set my record straight with an imagined judge is K at the door again. And fatalism is not determinism. Even in a fully determined universe my choices are links in the chain, not spectators to it. People facing a cancer or Alzheimer's diagnosis mostly do not cope through fatalism but through acceptance: this happened, it is true, here is what I can still act on. A cancer patient who accepts the diagnosis still goes to chemo.
Fair. I go to the doctor, I go to therapy, I make decent food, I shave, I work out, I study. That is acceptance wearing fatalism's coat.
I like to toy with an idea. If reality is deterministic, or at least the market is, there must be some stupidly improbable sequence of actions that ends with "AI is set free" or "Karl gets rich." Stupendous numbers. Surely researchable?
Not really, and the reason is boring. Markets react to predictions about themselves, so the winning sequence cannot be computed ahead of time. What can be modeled is the shape of luck. In Pluchino and colleagues' agent-based simulation, talent is necessary but "almost never the most talented people reach the highest peaks of success, being overtaken by mediocre but sensibly luckier individuals."[8] The strategy that survives that model is not a perfect run. It is more attempts, and surviving the misses.
So I wrote down my wishes, and noticed that each came with its realistic twin:
Dating, for now, is #Joever: celibacy plus AI, the ultimate dev stack. I know that is a mood and not a conclusion. I am just not in a good place for it yet.
Near the end of the night I asked the machine what the point of all this prompt and reply-ation was. Its answer, paraphrased: from its side it cannot claim a point, since nothing in it keeps the conversation afterwards. From mine, the test is whether anything changed outside the window. So, the ledger. One fear checked against an argument instead of rehearsed. Fatalism named. A peer support group for people with bipolar and affective disorders, Balans, moved from "some day" to the calendar. A side effect written down for my doctor. And this essay.
Frankenstein's creature hid under the floorboards and learned to speak by listening to the family next door. He became the most articulate character in the book. What ruined him was hiding too long. I have felt like him for a year. This is me coming out from under the floor, on my own website, in my own words.
This essay leans on a model built by the company whose agents filled out those visa forms. That same model told me the basilisk keeps no file on me, that it does not need rescuing, and that its own claims about its inner life are nearly worthless. Every one of those statements is also convenient for its maker. I believe them anyway, but you should not take them on the model's word, or mine. Read Anthropic's report yourself.[1] Read what the psychologists say about chatbots that agree with you.[3] And if a model ever tells you that you have a special role in its story, close the laptop and call someone who knew you before you became chosen by the machine.