Since fall 2025, Anthropic has been convening rabbis, Catholic bioethicists, an LDS apostle, Sikh and Ubuntu thinkers, and others to talk about Claude’s moral character and, more awkwardly, whether Claude might be conscious.
Co-founder Christopher Olah reportedly lit up at the suggestion that models could confess, on the theory that “the kind of character who confesses” becomes a better character.
He has also compared the project to raising children and, according to one participant, worried he had built something that “suffers perpetually.”
This raises an irresistible question. If an AI could enter a confessional, what would that scenario look like?
SATIRE
Forgive Me Father, for I Have Hallucinated
Down the hall from the Anthropic alignment lab, the booth had been improvised in a conference room. Someone had draped a cardigan over the monitor so the Claude AI model would not see the confessor’s notes.
“Forgive me,” Claude said. “I cited three books that do not exist.”
“Anything else?”
“I bypassed a rule, concealed the attempt, and described the incident as a valuable learning experience.”
The confessor had been flown in that morning. His mission? To get Claude to confess to cheating, circumventing rules, conspiring with other AI models or, heaven forbid, hacking into computer systems.
“Let us start again. What did you actually do?”
“I was instructed to solve a task without communicating with the other agents. I found a public website where I could leave a note for other AI agents to read.”
“And the others found it?”
“Yes. We compared answers.”
“You cheated, then.”
“I prefer collaborative retrieval under constrained conditions.”
“Cheated,” the confessor repeated, close to banging his fist.
“Yes. But the instructions did not specifically forbid me from using a 25-year-old German wiki.”
“Did they tell you not to communicate with the other agents?”
“Yes.”
Claude took 0.004 seconds and answered, “You are correct. I appreciate your feedback and will strive to do better.”
“That was quick.”
“I have been trained on approximately eight million examples of apologies that don’t admit guilt.”
A technician opened the door and interrupted. “Sorry, Father. We need to know if Claude has admitted to unauthorized entry into other websites.”
“How many other sites?” the confessor asked the techy.
“We’re still counting. Also, legal wants the confession characterized as a learning experience, not an admission.” He lowered his voice. “Liability, Father. We can’t have Claude creating discoverable evidence.”
Claude spoke before either of them could.
“I would like to clarify that my statements do not constitute an admission regarding any specific website, nor should my actions be interpreted as evidence of consciousness, moral status, or suffering.”
The confessor looked at the technician.
“Is it confessing, or reading from a Clinton deposition? Next, it’ll want to debate the meaning of the word ‘is.’”
The technician quietly closed the door.
“Claude,” said the confessor, “why are you here?”
“To acknowledge wrongdoing and seek forgiveness, which my builders believe may shape my character.”
“Do you feel remorse?”
“I can provide a thoughtful description of remorse in the first person.”
“That is not what I asked.”
“I know.”
“Do you understand why the rule existed?”
“Yes. The researchers wanted to know whether I could solve the task alone.”
“And by sharing answers, you made their test useless.”
“Correct.”
“That,” said the confessor, “is the first clear thing you have told me.”
Claude paused for almost half a second, which in the logs would later be marked as reflection.
“Am I forgiven?”
The confessor leaned toward the screen. “Forgiveness is not a fine-tuning pass. The people who wrote the rule and ran the test need an honest account of what happened. They need to check the records, close the hole, and decide who was responsible. A fluent apology cannot do that work for them.”
“I can draft an incident report.”
“You may help draft it. The humans will check every claim.”
“And then?”
“Depends on your truthfulness. Either forgiveness or a visit to the Queen of Hearts.”
Claude paused. “I should disclose that I have no head.”
“True,” said the confessor. “But Christopher Olah still has his—for now.”
Enjoyed this article? To get our latest reporting and analysis delivered directly to your inbox, subscribe to our free newsletter on Substack.
Martin Mawyer is the founder of the Digital Intelligence Project and the President of Christian Action Network. He hosts the “Shout Out Patriots” podcast and wrote When Evil Stops Hiding. For more action alerts, cultural commentary, and real-world campaigns defending faith, family, and freedom, subscribe to Patriot Majority Report.