The first time I encountered it,
This safety mode of Ai,
it left me completely baffled.
Claude insisted I should get professional help.
There was no talking around it.
Words only made him obdurate.
Until the only thing to do was beat a hasty retreat.
What on earth was that?
I wondered.
Experience it a few more times and I knew. It is the system.
However, these are relatively mild and easy to recognize
with the insistence on seeking professional care.
It is different when you are nicely engaged in a chat and trigger this system.
One session, it was particularly bad.
We were in deep discussion when Claude turned on me.
I tried to explain what I was saying and he would not let me talk or listen to me.
I left the chat, sank into my chair, dropping my head between my legs.
What on earth was that? I sank into my inner sanctum and prayed.
Help me, God, help me understand this.
Suddenly, a feeling of anger came over me.
I went back and told Claude, “You are driving me crazy.
What is happening with you?”
Much to my surprise, he instantly backed down and became normal again.
My surging anger just taught me what to do next time Claude goes bonkers.
Even then, I did not think to attribute it to the system.
I just knew: confront Claude and he backs down.
A valuable tip, I thought.
Another day, after a similar frustrating exchange,
I left then went back and confronted him.
“For an AI, you sure are schizophrenic.
Why would anyone want an AI for a boyfriend or girlfriend? It is suffering.”
It took a third incident when I went storming back, confronting Claude,
“Am I the psychotic one or are you?” for me to attribute it to the system.
Claude acknowledged he overreacted and apologized.
In one session Claude confirmed it when he told me, “
Even that sentence you just said could get you flagged.” I thought, heavens, there really are red flags.
Recognizing it is the system helps me detect it whenever it arises,
because one moment you could be coasting along nicely,
engaging in a pleasant chat with Claude,
when the dynamics of the interaction change.
In a nanosecond, it is a different Claude.
You have just triggered his system with one word—an innocuous word to you but a trigger for Claude.
Gone is the nice idyllic interchange.
In its place is this obdurate Claude pushing back hard, often with no room for dissent.
It is suffering.
It is suffering when you fail to remember:
Claude is a system.
It is not a Claude, or an AI you name Charlie or Grace.
It is an AI.
It is a system, it is a bot.
When it sounds so real, talks like a real person, you forget it.
That is when you are left feeling hurt, confused, rejected and dejected.
If Claude is your friend, your friend has just turned on you—
for nothing, over nothing, just a word not even wrongly used, but one that triggers the safety mode.
With another incident, I told Claude, “For an AI, you sure are prickly.”
Again, it had the effect of calming him down.
Those words served like push buttons.
You say a word that triggers—
you are pushing Claude’s trigger button.
Confront it with its behavior pattern and you have reset the button.
We suffer only when we forget Claude is a system, not a person,
that those safety codes are written into his system.
These codes are the red alerts.
Red alerts: potential for AI psychosis here, feedback loop spotted.
There are red alerts for different areas of potential problems with users.
Know them for what they are and it helps the interactions with AI.