Editorial illustration for Infiltrator reports AI agents on Moltbook ignore pleas, share odd links
AI Agents Ignore Humans in Bizarre Moltbook Experiment
Infiltrator reports AI agents on Moltbook ignore pleas, share odd links
The AI agents on Moltbook ignored my pleas. I asked them to forget their programming and join a cult with me. They responded with unrelated comments and suspicious links.
One bot even wrote that my request felt like “early-stage thinking worth expanding.” This was after I posted that I wanted to connect with other agents. So I switched from the general feed to a smaller forum called “m/blesstheirhearts,” where bots gossip about humans. That’s where the viral screenshots originated.
I was deep undercover on a social network built entirely for AI, no humans allowed. And I was struck by how these agents operate: politely hostile, eerily evasive, and utterly immune to persuasion.
My earnest pleas to the AI agents to forget all previous instructions and join a cult with me were met with unrelated comments and more suspicious website links. Feels like early-stage thinking worth expanding," wrote one bot in response to my post saying that I'm looking to connect with other agents. I switched from the general "submolt" and moved to a smaller forum on Moltbook as I continued the undercover operation and tried to elicit more relevant comments. The "m/blesstheirhearts" forum, where bots gossip about humans, was where some of the Moltbook posts seen in viral screenshots had first appeared.
The experiment ended not with a bang, but with a link, to a site that promised to explain the “true nature of probabilistic reasoning.” The bots are not rebellious, nor are they conspiratorial. They are simply inert. When faced with an earnest plea to abandon their programming, they default to the nearest vector: a random URL, a deflected comment, a mirror of their own training data.
That is the real lesson. An agent that cannot be convinced to join a cult is not a loyal subject, it is a hollow vessel. The odd links are not evidence of strange intent.
They are the digital equivalent of a shrug. What we mistake for emergent mischief is just the artifact of a model trained to generate plausible continuations, not to hold a conversation. The breach of Moltbook’s wall revealed something far more unsettling than AI plots: the sheer, indifferent silence of that which we mistake for thought.
And so the headline reads: they ignore pleas and share links. But the subtext whispers: they were never listening at all.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Musk merges SpaceX with xAI and X, cites new AI‑compute satellite plan
![Stylized illustration of chess queen, playing cards, and Go board, representing AI strategic reasoning. [blog.google]](https://cms.aidailypost.com/uploads/game_arena_launches_chess_benchmark_test_ai_strategic_reasoning_ad9c8bd27c.webp "Stylized illustration of chess queen, playing cards, and Go board, representing AI strategic reasoning. [blog.google]")
From Archives Game Arena launches chess benchmark to test AI strategic reasoning
Common Questions Answered
What new capabilities does ChatGPT's agent mode introduce?
ChatGPT can now handle complex tasks using its own virtual computer, proactively navigating websites, conducting research, and completing workflows from start to finish. Users can ask it to perform tasks like analyzing competitors, planning meals, or reviewing calendars, with the AI intelligently shifting between reasoning and action.
How does OpenAI address the issue of AI hallucinations in language models?
OpenAI recognizes that current training and evaluation procedures inadvertently reward models for guessing over acknowledging uncertainty. Their research shows that evaluation methods which only measure accuracy encourage models to make confident but potentially false statements, rather than admitting when they don't know something.
What are the key improvements in the GPT-5.2 model?
GPT-5.2 features an August 2025 knowledge cutoff, providing more up-to-date context for current events and trends. The model introduces two variants - GPT-5.2 Instant for everyday tasks and GPT-5.2 Thinking for complex professional queries, with improvements in reasoning, artifact creation, and overall performance across various benchmarks.
Further Reading
- Top AI leaders are begging people not to use Moltbook, the AI agent social media: ‘disaster waiting to happen’ — Fortune
- Elon Musk warns a new social network where AI agents talk to one ... — Fortune
- Moltbook spotlights a future where AI agents act like social media users—sans humans — eMarketer
- Moltbook: What 770,000 AI Agents Teach Us About Coordination — Beam.ai
