Economic centralization does not stay economic. Brynjolfsson & Hitzig close their paper
by tracing where it goes next — and what, if anything, might hold it.
Channel one — agenda-setting and lobbying
The first channel is the familiar one: money buys political access. "Economic
concentration can increase agenda-setting and lobbying capacity, a classic prediction of
political-economy models in which organized interests leverage concentrated rents"
(Grossman and Helpman, 1994). The empirical backdrop is stark — "the responsiveness of US
policy appears more closely aligned with economic elites and organized interests than with
average citizens…" (Gilens and Page, 2014). If AI concentrates rents the way the paper's
model predicts, this channel concentrates further with it.
Brynjolfsson & Hitzig, 2025, §8
Channel two — gatekeepers of public discourse
The second channel is new. "When firms that concentrate economic power also serve as
information intermediaries, there are new avenues for political power. As large AI systems
shape search, summarization, and content curation, they become gatekeepers of public
discourse." The same firms whose processing capacity K̄ makes centralization economically
attractive are, increasingly, the firms deciding what the public reads, sees, and is told
is true.
Brynjolfsson & Hitzig, 2025, §8
Channel three — the human-capital channel
The third channel is the longest-run and, the authors suggest, the deepest. As decision
rights concentrate, "human agents may see diminished bargaining power and weaker incentives
to invest in human capital." Education and civic skill are "robustly associated with
democratic stability and participation" — so a falling incentive to invest in either is not
only an economic loss. The paper states the trade plainly:
“If the 20th century was a century of rising human capital, the 21st may be one of
increasing machine expertise.”
Brynjolfsson & Hitzig, 2025, §9
Not technology alone
The paper is careful here: "the political consequences turn on governance, not technology
alone." It does not endorse a fix, but it does log the strands of institutional design
already in the literature — deliberative institutions such as citizens' assemblies; data
governance covering rights to access, portability, and control; proposals to treat data
contribution as compensable labor; and distributional proposals such as social wealth funds
or universal dividends, aimed at broadening who holds a claim on AI-concentrated rents. The
paper catalogues these without picking one, and this page follows it in that restraint.
The arc, in one breath
Hayek's problem: the knowledge an economy runs on is dispersed, local, and tacit, so no
single mind can gather it. The marvel: a price is a pointer, one number that lets a
dispersed system coordinate without anyone needing the whole picture. The machine that
changes the terms: AI codifies what was tacit and raises how much a center can process,
so the calculation that once favored the edge increasingly favors the center. The limits:
latency, long tails, and turbulence still force some autonomy back outward, no matter how
good the model. And the question this station leaves standing — now that the centralizing
case is real, what keeps the center accountable?
← Back to the start of the arc