Most of what circulates as research is confident, tidy, and hard to trace back to anything. This is the method used instead, and it matters more than any single finding. Here is how it works, and where it falls short.

Why I check everything

I kept noticing the same thing in AI summaries of research. The number reads as precise, and the study underneath it describes different people: a figure that sounds like it is about young men was measured on middle-aged adults in another country. Sometimes the number is real and the comparison is not. Sometimes it has no source at all. Every claim I have traced back so far was off in the same direction, toward the more alarming version, and that is the part worth knowing. An unchecked process doesn’t make random mistakes. It drifts toward whatever sounds best. I hold my own work to the same check.

What counts as checked

A finding is only treated as checked once I have opened it at its original source: the journal article, the official report, or the full poll release. A search summary, a news story or a website quoting the number doesn’t count, however reputable. When only the abstract was available, I say so.

For every checked finding I record five things, taken from the source itself:

  • Who was studied (the population)
  • How old they were
  • Where (the country or countries)
  • What kind of study it was: a poll, a survey at one point in time, a study following people over years, a trial, a review of many studies, a model
  • What the number is of: a share of everyone, of gamblers, of people who signed up, and so on

The last one catches the most errors. The same 84% can describe boys who gamble or boys whose friends gamble, and those are different findings.

What the labels mean

  • Checked: opened at the original source, all five details recorded. Only checked findings can carry a number in a post.
  • Secondhand: I’ve only seen it reported somewhere else. It can be discussed as an idea, never with a statistic.
  • Unchecked lead: something worth looking into that nobody has opened yet. The library counts these but doesn’t list them, so an unverified number never appears on this site.
  • Doesn’t hold up: the source contradicts it, or no source could be found. These stay on record with an explanation, so they don’t quietly come back.
  • Measured on young men: the study’s sample was young men, or it reports young men as a group. Most research isn’t, and every entry says who it was actually measured on.

Checked isn’t the same as proven

A checked finding means the number was copied honestly from its source and I know who it describes. It doesn’t mean the study is strong. A poll subgroup of a few dozen people and a review of 21 studies can both be “checked”. That’s why every finding shows its study type and who it was measured on, and why small or single studies are described as small or single.

How AI is used

The research is done with an AI research desk. It searches, reads, pulls out numbers and drafts, including the posts and journal entries. I edit everything and decide what is published. The AI is useful for range: it can read far more than I could. It is not trusted on accuracy, which is why every number is opened at its source and why the most quotable number in each post is checked twice.

When I compare what AI search engines say, I capture their answers on a specific date and call the engines A to E. The point isn’t which company did best. Search answers change from day to day and with wording, so each capture is a snapshot.

One question at a time

Each week starts from one question, decided before the search, so the research can’t go looking for the most exciting answer. I look for studies of young men first. If there aren’t any, that absence is reported too. The questions are listed, in order, in the research library.

Corrections

If a finding changes after I’ve published it, the change is logged with its date in the sources section of every entry that used it, and the library updates. Nothing is quietly edited. If you spot something I got wrong, I want to know.

Who funded the research

Some research on gambling is paid for by the gambling industry. Where I know who funded a study, I record it, and I name industry funding in any post that relies on it.

What this series won’t do

  • Combine two separately reported figures into a new one. Men and 18-to-24-year-olds reported separately don’t add up to a figure for young men.
  • Present a finding about one group as if it were about another.
  • Tell anyone to stop taking risks. The aim is to see the risks clearly, and to risk well.

See the findings: the research library.

If gambling is costing you or someone close to you more than it should, the US National Problem Gambling Helpline is free and confidential: call or text 1-800-MY-RESET (1-800-697-3738), or chat at ncpgambling.org.