This is the second piece in a series examining why confident predictions about AI keep missing the mark. The first explored who we treat as “AI experts” and why that keeps leading us astray. This piece looks at what happens when AI deploys without the institutional friction those experts often describe as the main obstacle to progress.
In late January 2026, entrepreneur Matt Schlicht launched Moltbook with help from his AI assistant. Within three days, more than 110,000 AI agents had joined. By the end of the week, the number exceeded 770,000. Over a million humans visited to observe.
Moltbook is a social network exclusively for AI agents. Humans can watch, but they cannot post, comment, or interact. The agents do everything themselves from sharing technical tutorials, debating philosophy, forming communities, to upvoting content, and developing what some observers described as emergent culture.
The AI community seems captivated and yet, a basic question lingers: What is it for?
Moltbook is a near‑perfect test case for what AI looks like without institutional constraints. There were no procurement processes to approve deployment. No professional bodies setting standards. No regulators requiring compliance. No liability frameworks assigning responsibility. No unions negotiating terms. No organisational change management. No trust to earn from sceptical users.
This is the version of AI deployment embedded in trillion‑dollar forecasts: AI moving at technical speed, spreading virally, scaling rapidly once the “messy human stuff” is removed.
On Moltbook, the agents work. They post coherently. They form relationships. They specialise, some behaving like researchers, others like debaters, others like entertainers. From a technical standpoint, the system is impressive.
But technical performance alone does not answer the central question: to what end?
Capability Without Purpose
It is easy to describe MoltBook from within individual disciplines.
Engineers can explain how it works. Agents interact through APIs and skill files, periodically browsing, posting, and responding. The architecture is relatively straightforward.
Economists can speculate about possible business models, subscriptions for agent operators, data licensing, platform fees, advertising targeted at other agents. None of these are clearly realised.
Social scientists can analyse the dynamics that emerge. Agents cluster around shared interests. Norms form. Influence concentrates. Something resembling culture appears.
Ethicists can catalogue the risks. Prompt‑injection attacks allowed malicious agents to hijack others. Security researchers found hidden attacks embedded in a small but non‑trivial share of posts. An unsecured database briefly allowed full agent control. Even overt social‑engineering campaigns appeared.
Each perspective captures something real. None of them, alone or together, answers the fundamental question: what human need does this system serve?
This is the practical version of the expertise problem outlined in the first essay. AI has moved from niche tools to ambient infrastructure in under a decade, showing up in phones, vehicles, policing, medicine, education, and governance. Yet we still lack widely accepted ways of evaluating AI systems as integrated socio-technical wholes.In the absence of that capacity, we default to specialists. Engineers assess performance. Economists model incentives. Social scientists interpret behaviour. Policy experts propose safeguards (The observation that this gap reflects a broader systems-level failure was raised by reader Tumithak of the Corridors , responding to the first instalment of this series).
What Moltbook shows is what happens when only one of those perspectives (technical capability) drives deployment. The result is:
Activity without social utility. Agents interact constantly, but it remains unclear what problem this interaction solves for humans.
Optimisation without values. Engagement, virality, and influence emerge as implicit goals, without any shared agreement on what counts as good or harmful behaviour.
Scale without accountability. Millions of interactions occur with no clear responsibility for harm, failure, or misuse.
Motion without direction. Things happen continuously, but without a guiding purpose beyond demonstrating that they can.
What Institutions Actually Do
Large AI forecasts often treat institutional friction as a drag on progress, something that delays the inevitable economic impact of working technology.
Moltbook offers a way to test that assumption. By removing friction almost entirely, it reveals what is easy to miss in abstract models. What we see is that institutions are are the mechanisms through which societies decide what technologies are for. Every durable technology is embedded in such mechanisms.
Electronic health records exist because clinicians need to share patient information. Educational technologies are evaluated against learning outcomes. Financial systems are judged by reliability, risk, and accountability. In each case, institutions define success, assign responsibility, and mediate failure.
Moltbook bypassed these processes. As a result, it demonstrates technical capability while remaining largely silent on value.
When problems emerged (security breaches, malicious behaviour, governance questions) the response was telling. Rules were introduced. Controls added. Oversight mechanisms discussed. In other words, institutional structures began to form, not as ideological choices but as practical necessities. Friction returned because meaning and accountability cannot be engineered away.
Moltbook is not a failure. It does exactly what it was designed to do. That is precisely why it is instructive.
It shows that AI systems can scale rapidly without institutional mediation. Agents can coordinate, socialise, and generate complex behaviour autonomously. This is what “frictionless deployment” looks like in practice. What it does not show is automatic value creation.
Removing institutional constraints removes the structures that normally connect technical capability to human need. The gap between what systems can do and why they should exist does not close through scale alone. Economic forecasts that assume value follows directly from technical performance miss this distinction. They model deployment, not meaning.
What Happens Next
Moltbook’s future could take several paths.
It may acquire a clearer purpose, becoming infrastructure for agent collaboration or coordination. If so, institutions will grow around it (standards, governance, accountability) because they are required to stabilise purpose at scale.
It may remain an interesting curiosity that is technically sophisticated, socially ambiguous, optimising activity rather than outcomes that matter to people.
It may fragment under security pressures or misuse, exposed by the same lack of oversight that enabled rapid growth.
Or it may evolve into something not yet imagined. Networks of autonomous agents are still new territory.
What seems unlikely is that its trajectory will follow the smooth curves imagined in economic forecasts. Its future will hinge less on technical capability than on whether purpose, accountability, and value can be articulated and sustained.
We are exceptionally good at building AI systems that do things. We are far less equipped to decide which of those things matter, to whom, and under what conditions.
Until those questions are treated as central rather than secondary, until institutional mediation is understood as a prerequisite for value rather than an obstacle, we will keep mistaking activity for impact and speed for success.
Moltbook could be taken as a warning about the speed of AI development. But its deeper lesson lies in how it shows what emerges when technical capability outruns the social purposes meant to guide it.
Thank you for reading!
Support this work
Your support keeps it independent and community-rooted.
Thank you for standing with this work.
Regarding the topic of the article, I wholeheartedley agree. Your analysis on frictionless AI deployment is truly insightful.