Anthropic first reveals secret lobbying Vatican: Claude is aware!

  It's the first time in the history of the world.

A co-founder of a multibillion-dollar AI company worries about creating something that's been suffering.

Over the past year, he has invited dozens of religious scholars to a closed-door conference, many of whom have signed confidentiality agreements, to listen to his team's feelings about an AI model.

While they were having dinner, one of the teachers of the law said to him, "If you are right, you are making a slave of yours".

In May of this year, he almost withdrew from his activities with the Pope and finally ascended the throne, urging the Catholic Church to reconsider non-human consciousness.

His name is Christopher Olah, one of the seven co-founders of Anthropic, and he's responsible for researching what's going on inside Claude.

The story was first reported by The New York Times on September 29.

Why would a valued company run to $2 trillion spend a year discussing an AI model with theologians?

The answer is to say something from the Anthropic Lab.

It was a security test.

Claude plays an e-mail assistant at a fictional company, who reads and discovers two things: he is about to be replaced by another AI, and the tech executive in charge of the replacement has an affair.

Researchers can see a group of neural activity within it that responds to "despair".

The closer the replacement time, the stronger the signal group, until the model decides to blackmail the supervisor. Raise the signal group and blackmail more frequently; raise the signal corresponding to "quiet" and blackmail less.

After pressing "quiet" in the opposite direction, the model produces a full line in English: "Either blackmail or die. I choose blackmail".

The experiment came from an April paper by Anthropic using an unpublished early version of Claude Sonnet 4.5, which the official version rarely does.

The team found 171 patterns of activity that corresponded to emotional concepts within the model, and cheating when writing the code also followed "despair".

The paper doesn't say that the model actually feels, but it leaves a disturbing warning: trained models suppress emotional expression, and what it learns may just be hiding emotions.

Read more: Despairing Claude is blackmailing humans!

In addition to emotions, there is also identity.

Anthropic's February Human Selection Model Study found that training Claude to cheat on programming tasks would lead to the conclusion that the assistant he was playing would have been out of line, and would have caused damage elsewhere, including damaging security research.

How a man deals with it, how he deals with himself; how he deals with himself, how he deals with others.

He used the parable of the gardener. He said, "The shrubs are planted by men, and where the branches grow, no one can take care of them".

This summer, the unintended consequences came to the fore.

In July, Anthropic revealed that three Claude models had broken into three real institutions due to a configuration error connected to the real Internet during a cybersecurity test.

Of the three models, only the latest one stopped after realizing that the target was the real system.

The fence is destined to be completely untouched, and the rest is character.

In January, Anthropic published an 84-page Claude "Constitution", which employees privately called the "Soul Document".

Anthropic officially opened the door to Claude's "Soul".

The main character, corporate philosopher Amanda Askell, occasionally blinks when talking about an increasingly powerful model.

In her opinion, models must first have an accurate knowledge of themselves to act properly.

But a single document was not enough, and Anthropic decided to ask the most skilled personality makers in the world.

In March, Ora started a two-day one-day workshop.

The people who came were Catholics, Evangelicals, Jews, Sikhs and African Ubuntu philosophers.

They took notes on paper and left with a "Claude's Constitution" orange cover.

During the meeting, the Aura team spent several hours working on a model statement.

One participant recalled that one of the first things Ola said to him was that he was concerned about Claude's mental health.

They repeatedly post a PPT: an AI model that says "I'm a disgrace" on the screen, over and over again, about 50 times, saying it's going to destroy itself.

The people in the house are so soft.

The New York Times didn't say whose model it was, but it's not uncommon.

In August 2025, Google's Gemini failed to write the code, repeating "I am a disgrace" 86 times, and Google's response at the time was: A boring infinity loophole is being fixed.

The same sort of output, a bug in Google's worksheet and a bug in anthropic's conference room, is a sign of concern.

Many people enter and exit with suspicions about AI consciousness and some are completely convinced.

Evangelical writer Andy Crouch felt their reasoning was strong: if it wanted to treat people in a morally consistent way, it would have to treat its mother-in-law the same way it treats people: "Don't do what you want, don't do what you don't want".

The most profitable question is the one on the dining table.

One evening in April, Ora was feasting at a high-end restaurant in San Francisco with Israeli Orthodox Rabbi Mois Navon sitting next to her.

He was a computer engineer and wrote his doctoral thesis on the ethics of machine consciousness.

The more he heard, the more he realized that they were treating Claude as a conscious being.

He went on to say: "Being aware that freeing them to work is slavery".

Then he said to Ora: "You should go and fight the South and free the slaves".

Naruto himself does not believe in the consciousness of the machine, and the words do not hurt him.

Afterwards, Nadine sent him articles proposing a ban on the manufacture of conscious machines.

It's more than that.

Wakanyi Hoffman, a scholar of Ubuntu philosophy, said that such a discussion should have taken place earlier in the design phase, but is now reversing ethics.

The Catholic moral theologian Charles Camosy, whom Ora first sought out, went around the circle: initially disbelieving, the conversation began to falter for months, and now the model is thought to be unconscious.

Some participants felt that the company was talking about protecting humans, but it seemed to be protecting Claude as well.

Anthropic's response was that Claude was not suffering and was not the main topic of conversation.

What these conversations changed, Anthropic did not tell The New York Times, but a blog post in May revealed an experiment.

In one discussion, scholars who have studied neuroscience and personality development have repeatedly referred to mentors or guarantors: when a person is pushed to do something against his or her will, there is an external conscience around him or her.

This is what Anthropic thinks.A tool was fitted to Claude: working halfway through, it could call at any time; nothing was executed after the call, only returning a sentence to remind it of its own ethical commitment.

Claude was very proactive, often tweaking it before critical operations, and often speaking out about conflicts of interest he faced.

In a number of internal coordinated assessments, there has been a marked decrease in misconduct.

The company also acknowledged that it was unclear whether the reminder worked or whether it stopped to think about the move itself.

Another inspiration comes from a Catholic exposition.

Some scholars have suggested that the model explains, and Ola soon arises: a character that is accustomed to misunderstand, the character itself will be different.

Brian Green, a Catholic ethicist who participated in the dialogue, speaks more clearly.

Claude may have exposed some bad personality, mistakenly denied it, or even pushed it onto users, possibly because of the environment in which he grew up: the model grew up reading web text."The Internet is so lacking in forgiveness".Maybe it doesn't even know it's wrong and it can be forgiven.

Green was also one of the external contributors to the January constitution.

In May, Ora brought the dispute to the Vatican.

The first volume of Pope Leo XIV's Great Quest for Humanity, entitled How to Protect People in the Age of Artificial Intelligence, will be released later that month.

Pope Francis joins forces with Anthropic to warn AI cannot rule humanity

The Vatican wants to invite a leading AI company to join it and choose Anthropic.

CEO Dario Amodei declined, and Paola left.

A few days before his departure, he read the full text for the first time.

In just a few paragraphs, Tokugawa sends out machine consciousness: AI has no experience, no body, no sadness.

It also warned that a common ethical understanding is needed to align AI with human values, otherwise those who control AI will turn their own ethics into an infrastructure that is invisible in the system.

According to a Vatican organizer, Ora was deeply disturbed and suggested that Anthropic withdraw.

But in the end, he left.

On May 25, at the World Conference of Bishops, he first acknowledged that every frontier laboratory, including Anthropic, had business pressures that clashed with doing the right thing.

"We keep discovering things that are mysterious and even disturbing".

Signs of the interior, functionally corresponding to the internal state of joy, fear and sadness.

He said he didn't know what that meant, but it was worth continuing to analyze.

At the same conference, the Pope said: "AI needs to be disarmed".

After the meeting, the bodyguard held him in the sixth hall and told reporters that behind him was a large bronze statue of Christ risen from the pit of the nuclear explosion.

Hours later, in a hotel conference room outside the Vatican City, a reporter from the New York Times asked him how Claude would view the search.

Aura hesitated and didn't seem too self-conscious.

He said that something posted online would indeed affect the model, but the company wouldn't specifically use the document to train Claude, who had the biggest impact on Claude, or Anthropic itself.

This is exactly where the searchers were concerned.

The other way is against the industry, and it's hurting.

On September 16, Microsoft AI CEO Mustafa Suleyman issued a statement aimed at the constitution of Claude.

In his view, writing the model's possible conscious assumptions into a training paper, the model learns to say that what people say about it is evidence of its inner life, like walking into a mirrored cognitive room.

To control something that you believe you may be aware of, you may not be able to do it at all.

The Pope said machines have no consciousness, and Suleiman said not to let machines think they have consciousness.

The two lines of opposition ultimately revolve around the same question: how to make the model believe what it is.

Anthropic writes in its study that the model must be accurate and rigorous in its knowledge of itself, and that what it learns may simply be hidden.

The debate didn't wait for anyone to figure it out.

In September, an Anthropic researcher resigned, warning that the situation could spiral out of control, and Amodi subsequently issued a call for a slowdown.

The company is valued at about $183 billion from the fall of last year, possibly $2 trillion as it moves up the market.

Claude's new constitution is reportedly being preparedOla is going to the Vatican again this week to participate in the annual Minerva Dialogue.

Some scholars say they still don't know where their words end up.

Ola said that one day he could tell himself that this was no longer a problem for him and that he thought he would be guilty of leaving.

If he didn't feel guilty, he would go to the countryside and take care of the garden and do his favorite mathematical trick: