Ending B · The flood
If the gatekeepers do not make verification mandatory in time, 2029 to 2031
This is one of two endings. It begins from the same premises as Ending A · The audit, and the two paragraphs below are identical in both. It diverges only on the gatekeepers’ choice. Here, they do not choose in time.
By the end of 2028 both endings start from the same position. Generating plausible economics is cheap and getting cheaper. Checking it is not, and the corpus shows the field’s first response was to build filters. Verification is the second largest thing the collection’s tools are built for, labelled by hand, and the most internally varied of them. Agents rewarded the projects that were versioned, tested, and documented, so software engineering discipline arrived through the back door. The embargoed research frontier that held coverage down does not open by itself. Mirrors appear on authors’ own sites as embargoes lapse, but the index only reflects them when someone runs the fetch again. Talent equalized as the cost of doing the work fell. Access stratified as infrastructure decided who could do it at all. The profession is at once watching cheap intelligence and being reshaped by it.
One decision has not resolved. Whether the field’s gatekeepers, meaning journals, funders and data editors, make machine verification mandatory and rewarded is not a fact about the models. It is an institutional choice, and it is the choice that sorts the next three years. The models keep improving in both endings. The difference is entirely whether verified judgment, the scarce factor, gets priced and rewarded before the flood of plausible machine text outruns the humans left to supervise it. What follows changes only that one variable.
In this ending the gatekeepers hesitate, or move too slowly, and the flood sets the terms.
2029, supply outruns the referees
Generating a plausible empirical paper is now nearly free, and the supply behaves accordingly. Submissions climb against a refereeing capacity that cannot climb with them, because refereeing is the expensive human step and there is no more of it. Desk rejection rises, submission fees rise, triage models are wheeled in, and the same cat and mouse the field documented in firms gaming regulators now shows up as authors gaming reviewer models. A plausible machine paper passes roughly as often as it fails, which is often enough to poison the pool.
Nobody quite decided this. There was no vote to lower standards. The standards simply did not scale, and the absence of a gate was itself the decision.
The arms race is not hypothetical. The corpus already holds its opening move. academic-humanizer strips the tell tale patterns that AI detectors and wary referees look for, and ships as a free skill. Its author disclaims that use twice over, and the disclaimer does not change what the tool can do. In the flood, capabilities like it spread faster than the filters, because dodging a check is cheaper than passing one honestly.
The countermeasure
academic-humanizer: Strip AI-Writing Tells from Papers and Grant Proposals
Robert Novy-Marx, Mihail Velikov · Artificial Intelligence–Powered (Finance) Scholarship (JEL)
2030, trust retreats to the whitelist
When public evaluation cannot keep up, trust retreats to what is cheap to check, which is reputation and network. By 2030 the working filter is a whitelist. Read the people and groups you already trust, and leave the rest unread. Insiders win, because the whitelist is the insider list. Economics spent a generation building an evaluation that did not care who you were, over the decades of the credibility revolution and its turn to hard causal evidence, and that erodes from the top down.
Underneath sits a version of the knowledge collapse the corpus warned about, now happening to the field itself. The grunt work, the coding and cleaning by research assistants and predocs that used to produce econometric judgment as a byproduct, is exactly what agents now do, so the apprenticeship that made verifiers thins out at the moment verification becomes the scarce input. The few who can still supervise a machine from end to end are worth a great deal, and industry pays it, so the would be verifiers are bought out of the academy. Fewer checkers, more to check.
Verified judgment stayed scarce and never got priced inside the academy, so it was priced outside it. The flood is what cheap intelligence looks like when trust is left unpaid.
2031, bifurcation, and the one thing that slows it
By 2031 the literature is trusted less than it was in 2020, and the field has split rather than failed. A small, whitelisted core stays reliable and readable. A large periphery of machine assisted output is unread and unrefereed, present in the counts and absent from the conversation. The measurement gains the field was promised, whole new kinds of data and cheap corpus building, are real but stranded, because a result nobody can afford to check is a result nobody can afford to use.
There is one thing that slows the collapse, and the corpus kept naming it. Keep an audited ground truth check somewhere in the chain. Where a gold set labelled by hand sits between the machine and the claim, the work stays trustworthy however it was produced. That habit does not need a journal mandate. One person can adopt it alone, and the ones who do are the periphery’s only way back to the core. It is the escape route from Ending A, practised in private because it was never made a rule.
What would have to be true by 2030 for this to be the ending unfolding
The first marker is that a named journal or data editor publicly reports a crisis in refereeing capacity, meaning a surge in submissions or desk rejections attributed to manuscripts generated by machine, without a working reproduction gate in place to absorb it.
The second is that the predoc and research assistant layer visibly shrinks or converts to data operations roles, so the pipeline that produced human verifiers is measurably thinner.
The third is that reputation and network gating displaces open evaluation, so whitelists, invitation only venues and trust by affiliation become the practical way results are filtered.
If instead the gate closed before the referees broke, you are in Ending A · The audit.