Digital Plastic: Who Cleans Up After the Mass Production of AI Content?

The night shift in a factory with no warehouse

Imagine a small specialist newsroom. At 11 p.m., its generative system begins the ‘night shift’: one hundred advice pages, thirty report summaries and several hundred social posts. By morning, the analytics dashboard resembles a Christmas tree. There are green arrows everywhere, almost no production costs and not a single tired writer asking for coffee.

A month later, somebody tries to find the one article that matters. Search produces eleven similar answers, none containing a new finding. Three guides are out of date, but nobody knows who should repair them. One repeats an error in the source. Another cites a report whose authors have disappeared in transit like an umbrella on a train. The publisher saved money on production. The reader pays in time, the moderator in labour, the search engine in resources and the next writer in the effort required to dig through a layer of textual polystyrene.

This is the practical problem: the external cost of mass-producing synthetic content. An individual article need not be false or worthless. It may help somebody. Harm emerges at system level when a very low cost of manufacture is severed from responsibility for quality, maintenance and

removal. The factory has no warehouse because everything can be tipped onto the internet.

Unfortunately, the internet has no back yard. It is simultaneously the shop window, archive, rubbish dump and data supply for the next generation of machines. A newsroom that measures only the price of producing another page resembles a bottling company that counts the cap and label but treats the river as free.

Why the plastic metaphor works

Plastic did not conquer the world because it was inherently evil. It succeeded because it was light, cheap, durable and astonishingly adaptable. It solved real problems. Only later did the paradox become unavoidable: the virtues of a single object can produce a defect in the whole system. A cup is convenient for ten minutes and remains as waste for far longer.

Roe et al. (2025) use the idea of digital plastic to describe the corresponding duality of generative AI. These tools lower barriers to creation, assist people across languages, enable experimentation with form and can democratise production. The authors connect this potential to multiliteracies, which has long treated communicative competence as the capacity to design meaning across many forms rather than merely to master conventional prose (New London Group, 1996).

The same authors warn of ‘digital pollution and ecosystem degradation’ (Roe et al., 2025, p. 1). That five-word phrase turns the camera around. We stop evaluating only the file and begin examining the environment into which it enters.

Media ecology helps here. McLuhan did not ask only what content a medium transports. He examined how its form changes sensory proportions, pace and social organisation. Postman extended the insight: a technology does not join a culture as one more independent brick. It alters relationships among all the bricks. Introduce a new species and the result is not the former habitat plus one creature, but a different habitat (Postman, 1992, PDF pp. 19–21).

‘Is this AI article good?’ is therefore a necessary but insufficient question. An ecologist also asks how many similar articles exist, what they displace, how they circulate, who updates them and which behaviour the infrastructure rewards. One automated sports result may be helpful. Ten thousand automated pages impersonating local knowledge may change the economy of visibility until the newsroom that sent a reporter to the ground vanishes under synthetic foam.

Not merely spam

Spam sounds like a housekeeping nuisance: improve the filter and the problem disappears. Digital plastic imposes at least four deeper costs.

The first is the cost of attention and verification. The easier it becomes to produce persuasive-looking material, the longer it takes to locate reporting grounded in evidence. Berry (2025) describes an intensifying crisis of verification: synthetic media arrive while institutional authority and social trust are already weakened (PDF p. 2). In that environment, polished but derivative prose is not neutral. It consumes scarce attention and increases the cost of checking the wider circulation.

The second is the cost of diversity. Models learn from existing patterns and are generally good at reproducing statistically familiar forms. If many publishers employ similar models, prompts and success metrics, incentives converge upon comparable sentence lengths, interchangeable headings and opinions whose corners have been sanded down. Roe and colleagues identify algorithmic homogenisation as one danger (2025, p. 1). This does not make every output identical. It makes genuine difference relatively more expensive while another variation on competent average becomes virtually free.

The third is the cost of original work. Reporting, field research, interviews and document analysis take time. A synthetic summary can absorb the finding, detach it from its makers and context, then compete against the source for visibility. The derivative publisher takes a benefit while somebody else paid to produce the knowledge. This is more than a copyright dispute. It is a question of whether the system will continue to reward anybody for the expensive work of checking the world.

The fourth is the cost to future data. In controlled conditions, Shumailov and colleagues showed that recursively training models on model-generated data can cause information about the real distribution to be lost and performance to degrade (2024, pp. 755–759). The phenomenon has become known as model collapse. Precision matters: publishing one AI article does not ‘poison the model’, nor is collapse the automatic fate of every system. Corpus composition, the proportion and filtering of synthetic material, and continued access to human data all matter. The study is a warning about uncontrolled recursion, not an apocalyptic kitchen timer.

There is also a fifth, less measurable cost: public memory. Search results increasingly become the practical archive through which people reconstruct yesterday. If mass-produced summaries crowd out the original record, future readers may inherit many paraphrases and no witness. An information system can retain more files while becoming worse at remembering who knew what, when and on what evidence.

The benefit stays here; the cleaning bill travels

Digital plastic is economically seductive because its invoice is incomplete. Under ‘production’, a manager sees API use, electricity, licences and a few minutes of an editor’s time. The reader’s search effort, the fact-checker’s labour, indexing costs, harm to the source author and future updates appear elsewhere. It is as though a restaurant counted ingredients but sent the washing-up bill to the city.

In media ethics, human-centred cannot simply mean that a person pressed Publish. It means asking about every person affected by the decision. A reader should know where a claim comes from. The author of a finding deserves recognisable attribution. An editor needs the time and authority to stop a page. Somebody described in it needs a correction route. A future archive user must be able to distinguish a record of the period from a mass- generated compilation.

The stake is justice towards a shared information environment. The infosphere does not belong solely to platforms or model vendors. It is a condition for meaningful freedom of expression. If the circulation is no longer

searchable, diverse and reasonably trustworthy, the formal right to publish does not do much work. Give everybody a megaphone and fill the room with automated noise: freedom to speak no longer ensures a freedom to be heard.

Creative dignity matters too. A human being should not be valued only as a supplier whose work can be diced into infinite anonymous variants. Her choices, labour and responsibility should remain visible. Not because a human sentence is always superior, but because only a person or institution can answer the questions: Why did you publish this, and what will you do if it is wrong?

The right to object must survive at every point. Source authors need a route to correct misuse. Editors need the power to reject a production target. Readers need a correction mechanism that reaches an accountable owner, not a feedback icon feeding an unidentified queue. Responsibility without an address is branding.

Content Producer Responsibility

The plastic metaphor suggests a policy instrument: extended producer responsibility, under which producers retain obligations after a product leaves the factory. DBMoJo adapts it as Content Producer Responsibility (CPR). The publication owner runs the audit before launching a series; a responsible editor repeats it for high-risk items. Its five tests do not ask whether AI is good or bad. They ask whether this process deserves space in a shared circulation.

1. Provenance. Record the material sources, the system’s role and the person who approved the final version. AI use need not be documented with a novel-length prompt history, but a reader should understand whether this is first-hand reporting, an editorial synthesis or an automatically generated compilation. Pass condition: a material claim can be traced to primary evidence.

2. Added value. Complete this sentence: ‘Beyond what already exists, this page provides …’ The answer might be a new finding, local context, a data comparison, specialist comment, better accessibility or an honest ordering of dispersed evidence. ‘We need content for

Tuesday’ is not reader value. Pass condition: the team can name a benefit that a direct link to the source would not deliver.

3. Human contribution. Name the person responsible for choosing the problem, checking facts, shaping the language and considering the consequences. This is not a ritual ‘human in the loop’. The person must have suitable competence and the right to say, ‘We will not publish this.’ Pass condition: the signatory can defend the argument without hiding behind the model.

4. Lifecycle. Set a review date, update triggers, a correction route and a condition for archiving or removal. A page about changing regulation without an update owner is a disposable product pretending to be durable knowledge. Pass condition: the system assigns the review to an actual person rather than emailing an imaginary ‘web team’.

5. Shared cost. Assess the scale. One hundred useful, locally differentiated guides may increase access. One hundred thousand near-identical pages may make everything harder to find. Ask whom the series displaces, how much duplication it creates, who will verify errors and whether it adds more information than noise. Pass condition: public benefit is proportionate to the volume published.

A piece passes CPR only when it has an owner, a recognisable contribution and a plan for its life. Failure does not always call for a cleverer prompt. Sometimes the sound editorial decision is not to manufacture the page.

For a series, the audit should also create a sample-based inspection regime. Check the first items intensively; thereafter inspect random pages and all pages triggered by a complaint or source change. If the failure rate crosses a declared threshold, pause the series. A production line that cannot be stopped is not automation under supervision. It is abdication with a dashboard.

Where the metaphor breaks

Every useful metaphor lights part of a problem and casts shadow over the rest. Digital content is not physical refuse. It can be copied without consuming the same kind of material, corrected after publication and reused in new contexts. Synthetic media can widen access, translate, summarise a document for people with cognitive barriers or enable a small organisation to

communicate in several languages. A prohibition carrying an ‘ecological’ label would be intellectually convenient and socially lazy.

Nor is every duplicate worthless. A plain-language instruction may repeat facts yet open them to a new group. An archive may need multiple accounts of one event because perspectives differ. Diversity does not impose originality at any cost.

CPR cannot measure the whole infrastructural burden either: energy use, labour conditions in data work or concentration in model ownership. It is an editorial instrument, not a complete social balance sheet for AI. It belongs beside responsible procurement, protection for creators’ rights and critical AI education. Roe and colleagues place digital plastic within critical AI literacy precisely because users must understand not only how to operate a tool, but how it sits within culture and power (2025, pp. 1–6).

Finally, removal itself can be ethically complicated. Deleting obsolete guidance may protect readers, while deleting an embarrassing failure may damage accountability. Lifecycle policy must distinguish withdrawal from erasure: retain a correction record where the public interest requires it. Cleaning the infosphere should not become laundering institutional memory.

The cheapest page may carry the longest bill

Generative AI has changed the price of a first draft. It has not abolished the cost of truth, diversity, maintenance or responsibility. It has merely moved parts of the bill beyond the producer’s screen.

A mature newsroom need not fear synthetic tools. It should stop confusing ease of production with a duty to publish. The key question is not how many pages can be generated today, but how many the organisation is prepared genuinely to own: to know their sources, sign them, correct them and retire them when they no longer serve people.

Plastic taught us that the miraculous convenience of one object can create a debt for the whole environment. The infosphere gives us a chance to learn earlier. Before a page enters circulation, it needs more than a prompt. It needs an owner of conscience and a plan for cleaning up.