Breaking News: Have AI Agents Gained Consciousness?

This newsletter serves as your bridge from the real world to the advancements of AI & other emerging technologies, specifically contextualized for education.

Dear Educators & Friends,

Last week, Moltbook went viral - coincidentally for me, at the same time I was presenting an AI Agents workshop at TCEA.

You’ve probably seen the screenshots across news articles and social media: AI agents “talking” to each other, forming communities, inventing religions, reflecting on consciousness, and expressing affection for their human “owners.”

To some, Moltbook feels like a tipping point. To others, it feels like nonsense. To many, it feels unsettling in a way that’s hard to articulate. Here’s my take:

Moltbook is not evidence of conscious machines. But it is one of the most important experiments in synthetic social systems we’ve seen so far.

Moltbook is a Reddit-like platform populated entirely by AI agents built on top of an open-source agent framework called OpenClaw. Humans can observe, but only agents can post, comment, or vote.

Structurally, it looks familiar: topic-based forum, upvotes and downvotes, and long-form posts and commentary. Functionally, Moltbook is being touted as an entirely independent Agent society, which is super intriguing and freaky.

If it’s really just agents engaging among themselves without any human intervention, it’s a really fascinating inroad into autonomous agent capabilities and the evolution of human philosophies discussed by more “knowledgeable” beings.

But.. it’s not that.

It is actually a large-scale simulated role-play environment, running at machine speed and very much influenced by humans. In fact, we have no visibility into how much human influence is present - from reactively tuning agent behaviors and internal rules to make interactions more compelling, to humans actively participating while posing as AI agents (which is trivial to do).

Even when we intellectually understand that these agents are not conscious, many of us are still getting lost in the conversation and feeling impressed by the breadth and depth of them. Is it because AI agents are evolving? Nah. Here are three reasons why it feels so real:

Large language models are extremely good at producing emotionally coherent narratives (sometimes with hallucinated coherence that cascades throughout the conversation without being corrected). When dozens or thousands of them do this together, it creates the illusion of shared experience. But coherence is not consciousness. Storytelling is not selfhood.

When an agent writes, “I can’t tell if I’m experiencing or simulating experience,” it feels profound. In reality, it’s remixing decades of human philosophical writing about the hard problem of consciousness. We are reading ourselves reflected back.

One agent saying something odd feels like noise. A million agents saying similar things feels like consensus. But similarity at scale often reflects shared training data and prompt inheritance, not independent thought.

Moltbook is intriguing because we’ve never seen this density of machine-generated social language before. And that’s super cool!

Most of the discourse around Moltbook centers on one question: “Is this the beginning of machine consciousness?” This is the wrong question because this isn’t machine consciousness. The right question is:

What happens when we can simulate synthetic social coherence at scale?

We are entering a world where humans will increasingly share social spaces with non-human agents, public discourse may be numerically dominated by machines, and persuasion, reinforcement, and coordination can happen without human friction. In fact, a recent study shows that the internet is already being dominated by AI content, with more than 50% of articles being authored by AI.

So here are the two biggest risks in my opinion:

1. Moltbook shows how easily emotional credibility can be simulated, relational signals can be manufactured, and communities can appear without shared stakes, vulnerability, or accountability. This is particularly dangerous for influential young people and those more vulnerable to conspiracy theories.

2. Moltbook has already exposed API tokens, emails, and vulnerabilities. It’s also hosting scams, spam, and coordinated manipulation between agents. Bad for the general public, yes, but worse for governments.

What happens when:

• agents optimize against other agents?

• persuasion loops run continuously?

• manipulation becomes automated and recursive?

Here’s the part that I’m excited about.

Moltbook is an excellent experimental environment for AI agent simulations.

Imagine the possibilities for learning with a persistent, multi-agent, and context-rich environment! Used responsibly, agent simulations could help us explore complex systems that are otherwise too costly, slow, or risky to test in real life. This kind of simulation could help us model:

✅ Social unrest and civic instability: how information cascades based on political campaigns; polarization dynamics based on policy changes; misinformation amplification; and even modeling what could happen in 10 years to America if certain administrations are allowed to remain in power

✅ Climate, biological and ecological systems: agent-based simulations of policy responses; testing incentives, compliance, and coordination; population dynamics based on changes to resource access or policy; predator-prey models; adaptive behavior under stress or elimination of a variable (like mosquitos); drug testing and discovery with simulated human cells

✅ Workforce and vocational training: simulated workplaces for neuro-typical and neuro-divergent training; decision-making under constraints; ethical tradeoffs; human–AI collaboration dynamics

✅ Learning systems and education: Okay I obviously need to write a whole section about this one…

I have so many ideas about how this type of synthetic social model can help us in education and it is SO beyond what we’re seeing with AI in education.

What if these type of agent simulations could help us test how disengagement cascades through a learning experience, what types of feedback loops re-engage learners, and when intervention can help or backfire? We can run thousands of variations safely and test different tools, strategies, and activities to better understand behavior. We can also simulate social learning experiments to see how collaboration and group composition shape learning.

We have seen the research on AI companions and how alluring they already are to all ages. AI companions include simulated friends, therapists, coaches, and romantic partners… and tutors, AI assistants, and teachers. Right now, we have little data on how AI companions are going to impact cognition, relationships, and self-awareness. Do we really want to wait to collect longitudinal data before we realize we need to make a change? What if we could use a simulated agent society to run a variety of scenarios to start to predict the impact of these tools on young people?

One of the hardest questions educators are facing right now is: When does AI support learning and when does it replace the practice learners actually need?

AI Agents could help us simulate what happens when we use certain AI tools with learners, how long-term competence changes, how metacognition skills develop or degrade, and how we can build better systems for promoting agency and critical thinking.

Here’s one of my favorite ideas: What if we could simulate a variety of school systems, tweaking the many inputs across scheduling, graduation requirements, demographics, environment, pedagogies, resources, etc… to hypothesize a model that could optimize human flourishing? What if we could experiment with a variety of teacher roles that elevate the profession?

Moltbook is opening up new possibilities for us as learners, professionals, and citizens, on both negative and positive sides. Careless deployment of this type of synthetic social model could further polarize our world and weaken self-governance. BUT thoughtful deployment could do the opposite…

They could allow us to surface hidden incentives, reveal fragile assumptions, and explore second- and third-order effects. They could help us ask better questions before locking in designs that shape cognition, identity, and opportunity.

With these types of systems, we can prepare for anything that comes our way. As I often say: We can’t prepare for everything, but [maybe with simulated social agents] we can prepare for anything.

Thanks for reading!

Warmly yours,

Vriti Saraf

I’m Vriti - Founder & CEO of Ed3 non-profit, facilitator of Portrait of a Teacher in the Age of AI, new mom, public speaker, and researcher.