What AI-Generated Content Changes

The change is economic before it is technical

Fabricated content is not new. What is new is that producing a large volume of fluent, plausible, well-formatted content now costs close to nothing and requires no particular skill.

That matters because several widely used verification habits were never really about truth. They were about effort. A well-written article, a professionally composed photograph, a coherent multi-page report, a video with matching audio: each of these once implied that somebody had invested time. Investment implied a stake, and a stake implied at least some accountability.

That inference is now unreliable. Not because everything fluent is false, but because fluency has stopped carrying information about effort.

Habits that stopped working

Judging by writing quality. Spelling errors and awkward phrasing were once reasonable warning signs. Generated text is grammatical by default. The absence of errors now tells you nothing.

Judging images by realism. Fingers, teeth, and text in backgrounds were useful artefacts for a period. They are much less useful now, and any specific artefact list ages quickly. Building a verification practice on a list of visual tells means rebuilding it every few months.

Treating volume as evidence of independent confirmation. Twenty sites reporting the same claim once suggested twenty newsrooms. It can now mean one source and nineteen automated rewrites. Corroboration requires independence, and independence has to be checked rather than counted.

Treating a citation as evidence a source exists. Generated text can produce citations formatted exactly like real ones, naming plausible authors, journals, and years, for papers that do not exist. A reference list is now a claim to be checked, not a credential.

Treating a voice or a face as identification. Synthetic audio in particular is cheap, short, and effective in contexts where people expect low fidelity, such as a phone call or a voice note.

Habits that still work, and matter more

Provenance. Where did this come from, and can I get to the origin? A claim traced to a named organization that stands behind it is in a different category from a claim that appears fully formed with no trail. This is the habit that gains the most weight, because it is the one generative tools do not touch.

Independence of corroboration. Two sources are only two sources if they did not get it from each other. Checking whether apparent corroboration traces back to a single origin is now the central verification move.

Checking the specific reference. If a claim cites a study, find the study. This single step catches fabricated citations, real citations that say something different, and real citations applied outside their scope.

Expected-evidence reasoning. If this were true, what else would exist? A public statement would have a recording. A regulatory action would have a filing. A weather event would have satellite data. Generative systems produce the artefact in front of you. They do not produce the surrounding record.

Consequence-scaled scepticism. Raise your standard in proportion to what acting on the claim would cost.

Why detection tools are not the answer

There is persistent demand for a tool that examines a file and reports whether a machine produced it. Such tools exist, and they are not reliable enough to carry the weight people want to put on them.

Two failure modes matter. A false positive accuses a real person of fabricating, which is a serious harm, and falls disproportionately on people whose writing is already atypical for the detector’s training distribution. A false negative provides false reassurance. Both errors are worse than no tool if the output is treated as authoritative.

Detection also sits on the wrong side of an adversarial dynamic. Anything that reliably identifies generated output becomes a target to optimize against.

Provenance is the more durable direction, because it attaches a verifiable record at the point of capture rather than trying to infer origin afterwards. The C2PA specification defines how such content credentials are constructed, signed, and validated [1]. Its practical limitation today is coverage: credentials are absent from most content in circulation, and are commonly stripped when a file is screenshotted, re-encoded, or uploaded. Present credentials are informative. Absent credentials are not evidence of anything.

The chatbot as an intermediary

A separate change is that generative systems are increasingly the interface through which people encounter information, not just a way of producing it. The Reuters Institute’s 2026 survey across 48 markets found 10 percent of respondents using AI chatbots for news, up from 7 percent the previous year, rising to 16 percent among respondents under 35 [2].

This introduces a specific problem that is easy to miss. A chatbot answer is a synthesis without a visible selection process. When a search engine returns ten results, you can see that a selection was made and inspect what was excluded. A single fluent paragraph presents no such surface. There is no ranked list to disagree with.

The practical response is to treat any such answer as a lead rather than a finding. Ask for the sources, open them, and check that they say what the summary reports. If the sources cannot be produced or do not resolve, you have a summary of unknown provenance, which is exactly the category this article is about.

What to do differently, concretely

1. Stop using production quality as a signal, in either direction.
2. Ask where a claim came from before asking whether it is plausible.
3. When you see multiple sources, check whether they are independent.
4. Open every cited study or document that a decision would rest on.
5. Ask what else would exist if the claim were true, and look for it.
6. Treat detection tool output as weak evidence at best, and never as grounds for an accusation.
7. Treat a chatbot summary as a starting point with sources to verify, not an answer.

Where this leaves us

The uncomfortable part is that these habits are more work than the ones they replace. Judging by appearance is fast. Tracing provenance is not.

The compensating observation is that you do not need to do this for everything. Most of what passes in front of you does not require a verdict. Reserve the effort for content you are about to act on, share, or repeat, which is a small fraction of the total and the only fraction where the cost of being wrong is paid by someone other than you.

References

Numbered citations in the text above correspond to the entries below.

1. Coalition for Content Provenance and Authenticity. (2025). C2PA technical specification, version 2.2. C2PA. https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html
2. Reuters Institute for the Study of Journalism. (2026). Digital News Report 2026: Overview and key findings. University of Oxford. https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/dnr-executive-summary

Next step

Use the AI content verification checklist

Topics covered: AI LiteracyInformation Manipulation