A video that says "studies show" is only as good as the studies. As more content is drafted by AI, the question of where a claim comes from matters more, not less. This article explains why AI systems produce unreliable references, which techniques make sourced content more trustworthy, and what a person should still check. The examples come from the research step of influencer automation in AI Influencer Studio.

Why does AI invent sources?

A language model writes the most plausible continuation of a text. A citation is just another piece of text with a familiar shape: an author, a title, a publisher, a web address. If the model has not actually looked something up, it can produce a reference that looks right and does not exist, or a real address that does not say what is claimed. This is not deliberate; it is how text generation works when it is not connected to retrieval.

Technique 1: search first, then write

The first defence is to make the model search the web before writing and to give it the results. Sourced content should be built from retrieved documents, not from memory. This alone reduces invented facts considerably, but it does not stop the model from attaching a wrong or invented address to a correct sentence.

Technique 2: verify every source address

The second defence is mechanical. After writing, every source address in the output is compared with the addresses the search actually returned. Anything that did not come back from the search is removed. In influencer automation this check is strict: unverified sources are dropped, and a run with fewer than two verified sources is retried once and then fails with a full refund, rather than going ahead with thin evidence.

Technique 3: link claims to sources

A source list at the end of a script says little about which sentence it supports. Linking each point to the specific sources behind it makes the content auditable: you can open the source for the one claim you doubt. Points that cannot be linked to a verified source are removed. The result is often shorter than an unsourced draft, which is usually an improvement.

Technique 4: keep dates

Facts age. Each source carries its publication date where available and the date it was accessed. For prices, versions, statistics and anything in the news, these two dates tell you how much to trust the claim today.

What sources do not guarantee

A verified source is a real page that was found by search. It is not necessarily correct, current, independent or representative. Sources can be marketing pages, can disagree with each other, or can be summarised without the caveats the original contained. Verification removes one category of error, invented references, and leaves judgement to people.

Sensitive topics need different rules

In health, finance and law, a confident sentence can cause harm even when it is mostly right. For these areas the safer behaviour is to avoid definitive claims and personal advice, to describe what is generally known and what is uncertain, and to add a short disclaimer. That is the rule the research step follows; it is a floor, not a substitute for expert review.

Commercial honesty

Sourced content often names real products. Two habits keep this fair: say clearly when a review is not sponsored, and do not reproduce logos or interfaces that are not yours. If the product is yours, show your own screens rather than a generated imitation.

A reviewer's checklist

  • Open the sources behind the two or three claims that matter most.
  • Check dates on anything numerical or recent.
  • Look for at least one independent source, not only the product's own site.
  • Read the script aloud: does any sentence say more than its source?
  • Keep disclaimers and not-sponsored notes when you publish.
  • Add the source list to your post description so viewers can check too.

The research and sources guide shows where these items appear before you approve a video.

Frequently asked

Does a source list make AI content trustworthy? It makes it checkable. Trust comes from someone checking the claims that matter.

Why not let the AI judge source quality itself? It can help rank sources, but its judgement has the same weaknesses as its writing. Mechanical checks and human review are complementary.

Should I show sources inside the video? Usually the description is the better place: it is readable, clickable and does not crowd the picture. A short on-screen mention can work for a single key claim.