AI-Assisted Submissions in Online Research Are Rare and Highly Concentrated but Routinely Approved

TEXT START: Online research platforms underpin much of what science claims about people, on the assumption that a human produced each response.

The Dissection

This paper is not proving safety. It is measuring current leakage under a narrow sampling regime, then using low present prevalence to narrow the definition of threat.

Its most important findings undermine its reassuring conclusion: AI assistance is already present, highly concentrated, frequently paid despite prohibitions, and partly invisible to existing detectors. “Routinely approved” does not demonstrate valid research data. It demonstrates that the platform’s payment machinery routinely accepts contaminated work.

The paper’s real function is boundary maintenance: converting a validity crisis into a manageable compliance gap. The fire is described as small while the paper itself documents the fuel, blind spots, and weak penalties.

The Core Fallacy

The paper treats existentiality as a current prevalence statistic. The Discontinuity Thesis asks whether AI can sever the human-production assumption once automated cognition becomes cheaper, faster, and strategically usable—not whether most submissions are already automated today.

It also confuses submission share with systemic risk. One percent contamination is not harmless if it clusters among prolific workers, affects particular studies, or concentrates in easily automated response types. The fact that 5% of assisted workers produced 64% of assisted submissions is a vulnerability multiplier, not reassurance.

Approval rates measure payment workflow, not authenticity. “No detectable increase” over more than three years is not proof of a durable ceiling; it may reflect selection effects, weak detection, or a lag before diffusion.

The paper provides a small-scale behavioral signal consistent with P1, but does not establish P1’s full claim of durable cost and performance dominance. It does not establish P3 either: occasional AI-assisted survey responses are not evidence that the majority has lost access to economically necessary labor. It does, however, expose the kind of institutional failure that P2 predicts: human-only rules exist, but enforcement is porous and penalties are weak.

Hidden Assumptions

• That low average prevalence means low systemic exposure.
• That workers willing to donate ChatGPT histories represent the entire platform population.
• That observed ChatGPT use captures other models, copied outputs, local systems, and unlogged assistance.
• That all assistance has roughly the same effect, despite the difference between proofreading and generating substantive answers.
• That bounded responses are less threatening because they evade an open-ended-response detector. They may actually be easier to automate and scale.
• That approval rates measure response quality rather than weak incentives, inattentive review, or routine payment norms.
• That three years of stable observed use establishes a ceiling rather than a temporary adoption plateau.
• That platform safeguards can improve as fast as model capability and concealment.
• That this platform is representative of online research generally.

Social Function

Primary classification: partial truth serving as transition management and ideological anesthetic.

The baseline measurement is legitimate, and the warnings about concentration, payment despite bans, and detector gaps are valuable. But the “not existential” conclusion freezes the system at its current adoption rate and turns structural fragility into a monitoring problem. It gives platforms permission to tolerate contamination until it becomes too distributed to contain.

The Verdict

The paper does not show that AI cannot destroy the human-response premise. It shows that the breach is currently small, concentrated, under-penalized, and partly invisible to existing safeguards.

That is not exoneration. It is an early-warning autopsy. Under DT, this is a pre-discontinuity contamination signal, not the discontinuity itself. Limited adoption, social norms, and institutional inertia are lag defenses—hospice care for the old verification regime, not evidence that the regime has recovered.