AI can make an individual task cheaper while making the system around it harder to sustain. A publisher loses the visits that help finance reporting. A recruiter faces more applications because submitting each one costs less. A future model learns from generated material that has replaced the observations it needs. These are different feedback loops, connected by a common accounting problem: the apparent saving leaves an essential input outside the ledger.

Victor Tangermann's October 4 Futurism article describes AI-related “doom loops” across publishing, hiring and other parts of the economy. Its most useful question is how a tool can undermine the conditions that make its own output valuable.

That question deserves a precise answer. Evidence of a damaged referral channel, a congested hiring process or a degraded training experiment establishes a particular mechanism. An economy-wide collapse would require evidence at a much larger scale.

Publishing can lose the return path

The publishing version begins with original work that costs money to produce. An answer service uses that work to satisfy a reader inside its own interface. If the reader has less reason to visit the originator, the service can gain utility while the publisher loses an opportunity to earn advertising revenue, acquire a subscriber or build a direct relationship.

The news plaintiffs' September 17 summary-judgment brief , in the OpenAI copyright litigation, cites an internal Microsoft document describing an AI content strategy that threatens the economic base of its suppliers. The relevant passage appears on printed page 2. It is a document quoted in the plaintiffs' legal argument, rather than a judicial finding that resolves infringement or measures the whole economy. 404 Media's account helped bring the filing into public discussion.

Independent browsing evidence makes the referral issue concrete. Pew Research Center studied 68,879 Google searches by 900 US adults during March 2025. Traditional-result clicks occurred on 8 percent of visits with an AI summary and 15 percent without one. Links inside the summaries received clicks on 1 percent of visits with a summary.

Those are observed associations. Pew reconstructed search results in April, and the searches receiving summaries differed from other searches. The study does not identify a uniform causal revenue loss for publishers.

The business question extends beyond displaying a citation. A citation identifies an origin. A visit, subscription or licensing payment supplies a different kind of return. An interface can perform the first function while doing little of the second. Preserving the production of fresh reporting therefore requires examining what actually reaches the reporting organization.

Hiring can turn cheap applications into expensive screening

Hiring has a different scarce input: attention to a person's suitability for a particular role. Lowering the effort needed to apply can help an applicant, but it can also increase the volume that a hiring team must evaluate. More filtering then gives applicants another reason to optimize their applications for the filter.

Greenhouse's explanation of its Ezra AI Labs acquisition describes this cycle from the employer-software side: applicants use AI to submit more applications, recruiters use AI to filter them, and trust deteriorates. The company proposes structured voice conversations as a remedy. That is a vendor's diagnosis and product proposition, rather than independent proof that its remedy improves every hiring process.

The distinction changes the success measure. Applications sent per hour measures the applicant's throughput. Applications screened per hour measures the recruiter's throughput. Neither establishes whether qualified people receive a fair assessment or whether a role is filled well.

Consider a hypothetical hiring team that saves five hours on initial screening but spends seven additional hours checking dubious submissions and repairing false rejections. Its screening tool can be faster while the hiring process consumes two more hours. These numbers are illustrative, not findings from Greenhouse. They expose why the accounting boundary matters.

An effective comparison would count accepted matches, review effort, correction work and the experience of candidates who were screened out. Making rejection cheaper is a different achievement from improving selection.

Model training has its own feedback problem

The technical version concerns the data distribution a model learns. If generated outputs progressively replace the original data, uncommon cases can disappear from later training sets. Repetition makes those omissions part of the next model's starting point.

Ilia Shumailov and colleagues' research on recursive training describes this loss of distributional detail as model collapse. The linked preprint was revised in April 2024. It is a study of training conditions, rather than evidence that every deployed chatbot is deteriorating.

A separate 2024 preprint by Matthias Gerstgrasser and colleagues tested an important alternative. In their experiments, retaining original real data while accumulating successive synthetic datasets avoided the progressive collapse seen when data were replaced. The results cover their tested models and settings, rather than a guarantee for arbitrary web-scale training.

The useful lesson is that data management changes the outcome. “Uses synthetic data” is too broad a description to diagnose a training failure. Which original observations remain, how generated samples are selected, and how performance is checked all affect the comparison.

This also separates a statistical problem from an economic one. Keeping an old dataset can help preserve its distribution. It cannot supply tomorrow's investigation, a newly observed event or an experience nobody has documented yet. Training-data preservation and the continued production of new knowledge address related but different needs.

Follow the input that the shortcut consumes

A useful way to assess a proposed AI saving is to follow three things through the entire process: the output being accelerated, the scarce input that makes it useful, and the return that sustains that input.

For publishing, the input is original reporting and the return includes reader relationships and revenue. For hiring, the input is credible evidence of suitability and the return is a useful assessment. For model training, the input is representative information and the return is a learning process that preserves meaningful variation.

The concept of an externality helps describe a cost imposed beyond the party making the immediate decision. A faster process can look efficient because another participant absorbs the review, financing or recovery work.

Newsroom's earlier analysis of cost per completed AI task provides the organizational version of this test. The wider test includes suppliers and recipients as well as the tool's purchaser. Count the work required to obtain an accepted result, then ask whether the people and institutions providing its inputs can continue doing so.

The featured photograph shows The New York Times building in 2019. Defears' original image and the resized, neutrally matted derivatives used here are available under CC BY-SA 4.0 . It identifies a publishing institution at the center of the legal dispute, rather than depicting an AI-related incident.

AI's doom-loop risk becomes clearer when each missing return is named. The question for a new system is whether its faster output also preserves the reporting, attention and original information on which the next useful output will depend.