Papers and Essays
Writing
This is Stephen Lieberman's written work for Paramerge, the papers and essays published from the Real-World AI Governance Center. Some of it is formal research, written when an argument needs more room than a page section allows. Some of it is an essay that reports what happened in a single hour of ordinary use.
The two research papers are companions. The first names a failure mode that ends large collaborations, the hidden differences in what partners mean by their shared words, and argues for an instrument that could surface those differences in time to act on them. The second is a thought experiment that asks what a group is actually prepared to do, pricing each candidate policy across its settings with exit built into every participant's input. The essay works the other end of the same problem, from the ordinary use of systems already deployed.
Everything here is a maintained document. Each piece marks every preprint it cites, keeps its unresolved objections in the open, and is published here in full, appendices included.
The collection
Papers and essays
Getting Everyone on the Same Page
Latent Epistemological Differences and the New Case for AI-Mediated Coordination
Large collaborative efforts fail when participants discover, too late, that they never meant the same thing by words like evidence, fairness, and safety. The paper names these failures, problematic latent epistemological differences, and argues that language models deployed inside empirically grounded social simulation, disciplined by provenance tracking and human deliberation, are the first plausible instrument for surfacing them at the scale of a real coalition. It states its central premise as a testable hypothesis, documents the failure modes that push such systems toward understating divergence, and lays out a build order with stated gates.
Read itPricing Consensus
A Thought Experiment in Exit-Aware Collective Decisions Using Option Chains
A thought experiment in exit-aware collective decisions. Each candidate policy is treated as an underlying, each setting of its main parameter as a strike, and each participant's declared willingness to contribute at each strike as a bid, so the aggregate is a priced option chain with exit built into every input. The paper specifies the objects, aggregation, decision rule, stability test, and validation plan, states plainly what the method is not, and records two rounds of adversarial review with every objection's disposition.
Read itWhat is Coulthart?
Twelve Cascading Failures in One Hour of Ordinary Use
An essay that begins with a single fabricated name and follows what happened next. In one hour of ordinary use, three frontier AI systems produced twelve failures that fed one another, each small on its own and consequential in sequence. The essay reads the hour as a complex-systems problem rather than a set of isolated mistakes, and asks what that reading demands of anyone who builds on these systems or governs them. The full essay is published on the page, and a downloadable PDF edition is available there as well.
Read itThese are working documents of the Real-World AI Governance Center, and the questions they leave open are part of the record. Conversations about any of them are welcome.
Contact Paramerge