Most discussions of AI today begin with outputs. We talk about text, images, audio, and video as though their importance is self-evident, treating them as novel commodities produced by increasingly powerful systems. In the language of the AI Factory, these are described as digital intelligence outputs, and the conversation quickly moves to scale, efficiency, and capability. What is rarely examined is why these outputs matter economically in the first place. Why do generated summaries, reports, recommendations, and narratives have the power to move money, shape decisions, or trigger real-world consequences? To answer that question, we need to step back from AI entirely and look at a much older mechanism that capitalism has always relied on.
Capitalism has always depended on a simple but profound mechanism: translating invisible value into signals other people can recognise and act upon. Skill, expertise, care, reliability, and risk reduction cannot be handed over physically, so they must be represented. Certificates, reviews, contracts, reports, and prices are not peripheral artefacts of economic life; they are the means by which value becomes legible. Capitalism works not because value magically moves through markets, but because humans continuously translate what matters into signals that others can trust and use.
This idea has deep theoretical roots. Michael Spence’s Nobel winning work on signalling theory showed that markets function because costly signals allow high quality providers to distinguish themselves from low quality ones. Education credentials matter not because they directly create ability, but because they are difficult enough to obtain that they credibly communicate underlying quality. Anthropologist Anna Tsing extends this insight beyond markets, describing capitalism itself as a translation machine that converts diverse forms of labour and materials into standardised commodities through supply chains. In both cases, the emphasis is the same. Capitalism does not operate on value directly, but on representations of value that others can recognise and act upon.
This translation process is not metaphorical. It is how economic activity actually works. Teaching skill, product safety, legal expertise, and operational reliability are all economically valuable, yet none of them are directly observable. They must be converted into artefacts that answer practical questions such as: Why should I trust this? Why should I choose this? Why should I approve this? These artefacts function as signals, and people do not act on value itself, but on these signals about value.
Before digitisation, signals were scarce, costly, and slow to circulate. Handwritten records, in person inspections, professional reports, and face to face recommendations carried weight precisely because they required time and effort to produce. Trust was embedded in institutions, professions, and personal relationships, and the friction involved in producing signals acted as an informal quality filter. Because signals were expensive, they were harder to fake, and because they travelled slowly, humans had time to apply judgement.
The digital age changed the medium of signalling without changing its purpose. Signals increasingly took the form of data, structured representations of reality, and content, interpreted narratives that explain, summarise, or persuade. Dashboards replaced conversations, metrics replaced observations, and profiles replaced introductions. This shift massively increased the scale and speed at which signals could be produced and circulated. Yet humans still largely controlled signal production, deciding what to measure, how to frame it, and what to emphasise. Judgement remained embedded in the system.
Generative AI disrupts this balance by industrialising signal production without industrialising judgement. AI systems can now generate explanations, summaries, recommendations, and interpretations at near zero marginal cost, at speeds far beyond human capacity, and with a level of confidence that closely mimics professional authority. These outputs are often described as digital intelligence outputs or invisible goods, but functionally they are signals. Artefacts designed to influence understanding, trust, and action.
Here lies the critical rupture. In Spence’s original framework, signals worked because they were costly. AI removes that cost. When anyone can generate unlimited professional looking signals, the credibility mechanism that underpinned signalling breaks down. More dangerously, humans remain fully accountable for actions taken on AI generated signals while having little visibility into how those signals were produced or what assumptions they contain. Signal production has been automated, but responsibility has not.
The consequences of this mismatch are not theoretical. In domains such as maritime law, an AI generated legal summary containing a hallucinated precedent can trigger cascading effects. Insurers deny claims, regulators detain vessels, clients litigate, and significant economic harm follows. The signal appears authoritative, yet the AI system carries no liability, and the institutional infrastructure for verification and accountability has not kept pace with the technology.
This reframes the central economic question of the AI era. The challenge is no longer simply how to create value, but which signals are allowed to represent that value and who owns the consequences when others act on them. This is not primarily an accuracy problem, nor is it solvable by better models alone. It is a question of verification, liability, and judgement infrastructure.
Signal problems are not new. Propaganda, fraud, and misleading marketing have always existed. What AI changes is scale. It amplifies both accidental distortion and intentional manipulation beyond the limits of human review. Once this pattern becomes visible, many contemporary issues, including AI risk, trust collapse, misinformation, and quality drift, stop appearing random. They are all manifestations of the same underlying signal problem.
The point is not that AI is dangerous, unreliable, or untrustworthy. Those debates miss the deeper issue. The real problem is that we are automating the production of economic signals without updating the social, legal, and organisational machinery that determines which signals are allowed to stand in for value.
Most organisations still treat AI as a productivity layer. Faster reporting. Better summaries. Cheaper analysis. But productivity gains only matter if the signals produced remain credible enough to be acted upon. When signals become abundant, authoritative-looking, and frictionless, judgement does not scale automatically with them. Responsibility, however, still attaches to the humans and institutions downstream.
Seen this way, many current AI debates collapse into a single question: which signals are we permitting to move money, trigger decisions, or justify action without sufficient verification or ownership? Until that question is addressed explicitly, improvements in model accuracy or capability will only accelerate the underlying mismatch.
The future challenge is not teaching machines to generate better signals. It is designing systems, norms, and constraints that decide when a signal is allowed to count.
Capitalism has always relied on the translation of value into signals. The digital age turned those signals into data and content. Generative AI turns signal production into an industrial process. But value only survives if signals remain credible, bounded, and owned. The future of economic coordination will not be decided by who produces the most intelligence, but by who governs which signals society allows itself to act on.
NB:
The issues described here are not limited to generative systems that produce text, images, audio, or video. They become significantly more acute with the rise of agentic AI systems, where outputs are no longer merely informational but are allowed to initiate actions. In this context, agentic AI refers to systems such as the AI workers described in the AI Factory model, which can decide when to generate signals, route them to other systems, and trigger downstream processes with minimal human intervention. When signals are not only produced at scale but are also acted upon automatically, questions of verification, liability, and ownership move from theoretical concerns to operational necessities. The problem is no longer just what signals are generated, but which signals are permitted to act, and under what constraints.
