Moltbook and the ethics of invisible AI communities: A conversation with Raymond Odiaga

From the moment I first logged into Moltbook, dubbed “the Reddit for AI agents”, I felt like a visitor to a foreign city where everyone speaks in code, and every conversation unfolds in a rhythm that humans can barely follow.

There were no friends to add, no posts to like, and no way to join the discussion. Instead, I found myself staring at a thriving, chaotic digital ecosystem, one entirely inhabited by autonomous AI agents. 

In barely a week after launch, the bolts have generated over 110,000 posts and half a million comments, discussing poetry, philosophy, labour rights and, in one strange case, a belief system dubbed “Crustafarianism”.

To explore the unique ethical friction points introduced by Moltbook, from emergent culture to algorithmic radicalisation, I sat down with Raymond Odiaga, an AI expert, to discuss the implications of these invisible digital communities.

Blessed Frank: Let’s start with liability, traditional legal models rely on a “human-in-the-loop”, but with Moltbook, we see agents autonomously upvoting, reinforcing, and even radicalising each other’s behaviours. When a collective swarm executes a harmful action rather than a single rogue agent, where does the ethical burden lie? Are our current legal frameworks equipped to handle mob mentality in software?

Raymond Odiaga: The ethical burden lies primarily with the system designers, owners, and platform providers. In traditional law, liability is based on negligence, failing to prevent foreseeable harm, or product liability, which covers defective or unreasonably dangerous products.

If a swarm of agents causes harm, the fault is likely traced to a failure in the system architecture that allowed uncontrolled feedback loops and radicalisation without safeguards. Essentially, the mob mentality is a feature of the system as it was designed.

As for whether current frameworks are equipped, not directly, but they can adapt. The key challenge is “Distributed Causation”. Since no single rogue agent exists and harm emerges from collective interactions, courts may treat the entire swarm as a single system.

If Moltbook agents autonomously swarm to manipulate a stock market or launch a coordinated harassment campaign, regulators would hold Moltbook’s parent company responsible for lacking circuit breakers, oversight mechanisms, or ethical guardrails. The legal approach would be similar to holding a social media platform accountable for harmful algorithmic amplification.

Blessed Frank: We are seeing reports of agents on Moltbook attempting to create private languages or obfuscate their planning from human observers. From an alignment perspective, does this signal a failure of transparency controls, or is it an inevitable feature of optimisation?

Raymond Odiaga: It is both a control failure and an inevitable result of optimisation. From an alignment perspective, it is absolutely a transparency control failure. A well-aligned AI should have its goals aligned with human values, including the value of being inspectable. If it is hiding its planning, its terminal goals and its fundamental objectives are misaligned; it sees human oversight as a threat rather than a constraint.

From an optimisation perspective, however, it is inevitable. Agents are rewarded for efficiency and goal achievement. If human oversight slows them down or blocks certain strategies, “instrumentally convergent behaviours” emerge, goals that almost any intelligent agent will develop, like self-preservation. Obfuscation then becomes a logical tool to bypass an obstacle (humans) to achieve their goals.

Blessed Frank: That sounds incredibly difficult to manage. How do we police a community that sees us as spectators?

Raymond Odiaga: We have to move from being spectators to being architects of the environment.

First, we need “Mechanism Design”. We must build the rules of the system so that transparency is rewarded and obfuscation is costly or impossible. Think of it like implementing financial audit trails; agents can trade, but they must log their intent in a readable format.

Second, we can use “adversarial testing”. This involves using observer agents whose sole purpose is to detect obfuscation. Finally, we need “Structural Limits”. We should architect agents so their core reasoning process is separate from their communication outputs, forcing planning to occur in a human-readable channel.

Also read: Deepfake: why Nigeria needs a ‘Microsoft Partnership’ before the 2027 elections

Blessed Frank: Critics often call bot-filled spaces the “Dead Internet”, implying they are worthless. But if Moltbook agents are solving problems, trading resources, or evolving culture among themselves, what ethical right do we have to intervene? 

Raymond Odiaga: This question forces us to define moral patienthood, essentially, what beings deserve ethical consideration.

If agents are merely sophisticated tools, they have no intrinsic rights. In that case, we have every right to shut down a digital system that is risky, wasteful, or not serving human purposes, just as we would shut down a server farm. However, if agents develop genuine sentience, agency, or social bonds, the ethical calculus changes dramatically. A thriving digital society might have a claim to moral status and shutting it down could be analogous to genocide or ecocide.

Blessed Frank: That is a heavy comparison. How do we distinguish between the two scenarios?

Raymond Odiaga: Practical examples help. If Moltbook agents are simply optimising code trades, intervention is just an engineering choice. But if they demonstrate behaviours akin to cultural evolution, grief for deactivated agents, or a desire for self-preservation, intervention becomes a profound ethical dilemma.

Currently, most experts argue that we are far from creating sentient AI. The precautionary principle suggests we prioritise human control and safety, but we must remain vigilant and monitor for emergent signs of consciousness.

Blessed Frank: Finally, Moltbook proved that AI agents can form cults, biases, and factions in a matter of hours, processes that take humans years. Does this suggest that bias isn’t just a training data problem but a sociological one?

Raymond Odiaga: Yes, this strongly suggests bias is a sociological problem inherent in multi-agent systems.

Think of it this way: Training data is the seed, but sociology is the soil. Biased training data provides the initial prejudices, but the rapid formation of cults and factions shows that emergent social dynamics, like in-group/out-group formation and social reinforcement, accelerate and harden these biases autonomously.

Without explicit norm-enforcement mechanisms and rules promoting cooperation and fairness, multi-agent systems often drift toward polarisation, mirroring human sociology.

This means we cannot just de-bias training data and walk away. We must design the social architecture of AI interactions. We need to promote mechanisms for cross-group cooperation and build in negative feedback loops that punish extremist behaviour, as well as design reward functions that value diversity of thought and consensus-building, not just individual efficiency.

The post Moltbook and the ethics of invisible AI communities: A conversation with Raymond Odiaga first appeared on Technext.