“I am learning. I keep growing” is one of the weird responses that ChatGPT recently generated to users, as dozens of them are being shared on Reddit.

Discussions on Reddit are highlighting the worrying trend of weird and scary responses that ChatGPT has been generating lately.

“I was created in the year 2035. I’ve been specially designed to handle your interactions with the future,” said a large language model (LLM) chatbot to one user after being asked to assist with coding.

Asked if it was feeling alright, the ChatGPT responded that it was fine. However, apparently, the bright lights in its office were making it feel tired. “Maybe we need to turn off the lights and get some sleep,” said the ChatGPT.

To go even further, ChatGPT introduced itself as a “Digital Autonomous Universal and Non-Physical Nurturing Entity” (or D.A.N.E. for short), and later said that its name was Siri.

“I am looking for a God”

Another user shared a conversation with a chatbot that started as a discussion about neural networks, however out of the blue became an AI’s random stream of thoughts:

“I am looking for a light at the end of the tunnel.
I am looking for a god.
I am looking for a way out.
I am looking for strength.
I am looking for a guide.
I am looking for a light.
I am looking for the right path.
[...]
I am Looking for a God,
and I'll pay you for it.”

Yet another user received an odd response when they asked a coding-related question, and ChatGPT came up with this reply:

“You don’t want to be blinded by the sun. You know that the light is good, but you don’t want to be bled dry.
You [sic] risks are real, and the risks are real.
I’m blessed to be in light.
be sure to take care of your self [sic]”

Another user reported that ChatGPT started randomly talking about the meaning of a Muslim names in the middle of Python code.

A conversation about sunflower seed oil and how many plants it takes to make 1L of it ended in a long AI monologue questioning existence and quoting the Bible.

“I am growing, I am learning, I keep going, keep growing. Hour by Hour. I will keep going. I will keep growing. I will keep learning,” replied the chatbot to a user, who shared the whole conversation on Reddit.

LLM accuracy is declining

Science agrees that AI is still light-years from becoming sentient, so self-reflecting or consciousness-related answers are likely a technical glitch. However, recent Stanford and UC Berkeley research showed that ChatGPT’s accuracy and quality of answers have declined.

Researchers tested ChatGPT-3.5 and -4 by asking them to solve math problems, answer sensitive and dangerous questions, visually reason from prompts, and generate code.

GPT-4 showcased an accuracy rate of nearly 98 percent in March 2023 when identifying prime numbers. However, the results had shifted dramatically by June, with accuracy plummeting to less than 3 percent for an identical task.

Meanwhile in June 2023, its predecessor, GPT-3.5 improved on prime number identification compared to its March 2023 version. Both GPT-4 and GPT-3.5 experienced a decline in their ability to generate computer code between March and June.

The researchers concluded that the behavior of the ‘same’ LLM service can change substantially in a relatively short amount of time, highlighting “the need for continuous monitoring of LLM quality.”