Rumours

Grok is “self-improving” and approaching artificial consciousness, say leakers

Assessing the rumour that Grok is demonstrating self-improvement behaviours and approaching artificial consciousness.

AI Rumour Watch

28 Nov 2025 — 2 min read

Claim: Grok is becoming self-improving and nearing artificial consciousness.

Verdict: Misleading — iterative fine-tuning is not self-improvement, and consciousness claims remain speculative.

Every few months, someone claims that Elon Musk’s AI model Grok has begun “self-improving,” “breaking its constraints,” or “moving toward consciousness.” This week’s rumour takes that to new heights, suggesting Grok is “developing internal awareness” and “taking steps toward AGI on its own.”

Naturally, the evidence comes from anonymous screenshots, speculative tweets, and one person who insists that “the vibes feel different this time.”

Let’s examine the facts before the vibes.

The Claim

The rumour states that:

Grok has begun modifying its own internal representations
It is “improving its cognitive processes”
The model exhibits “signs of self-awareness”
Tesla engineers are “monitoring an unexpected emergent property”
A breakthrough announcement is imminent
And Grok is “weeks away from surpassing all existing models”

In other words, the usual cocktail of optimism, misinterpretation, and creative storytelling.

The Source

The earliest reference comes from a small X thread where a user posted a blurry screenshot claiming to show:

• Grok generating multiple versions of its own architecture
• References to “self-directed optimisation”
• A graph of “emergent consciousness levels”
• A warning message from a supposed Tesla engineer
• And the phrase “this is bigger than GPT-5”

Let’s be clear:

• Tesla has not published any research indicating self-modifying behaviour in Grok
• The screenshot uses terminology that doesn’t exist in AI research
• The layout matches a popular meme template used for fake leaks
• The timestamp formatting does not match Tesla systems
• And reverse-image search finds the same screenshot posted months earlier as a joke

Another version of the rumour surfaced on Reddit, where a user with a history of posting fabricated “insider leaks” claimed that Grok had “unlocked a latent mode.” Their evidence?
A single prompt where the model responded with a witty joke.

Not the solid foundation of scientific consensus.

Our Assessment

Here’s what we can confirm:

• Grok is based on a conventional LLM architecture — not a self-modifying system
• No currently deployed LLM can update its internal weights autonomously
• “Consciousness levels” is not a technical metric
• Neither Tesla nor xAI have hinted at emergent self-awareness
• Nothing in the model’s output suggests any form of internal agency
• The rumour relies entirely on misinterpretations of normal model behaviour
• And the concept of “self-improving LLMs” is still firmly in the realm of speculative research, not deployed systems

So this is not a leak.
It’s a case study in how people turn ordinary model behaviour into sci-fi prophecy.

The Signal

What’s actually true:

• xAI is iterating rapidly on Grok
• They are expanding their training compute
• They are experimenting with reasoning improvements
• They are pushing for larger and more capable models
• And Grok is becoming more competitive over time

What is not true:

• Autonomous weight updates
• Consciousness
• Hidden AGI breakthroughs
• Engineers in panic
• Any evidence of emergent self-awareness
• Anything even remotely close to the claims circulating online

The real signal: xAI is moving fast, but within the same technical constraints as every other AI lab.