How this site is made

The pages on How Scams Work are researched and drafted with AI, not written by a named journalist or expert. This page says what that means: what the AI does, what a person does, how claims are checked, what the checks have found, and what nobody checks.

The short version

• Most pages are researched and drafted by AI models (Anthropic's Claude), working from agency, court and regulator pages they fetch and read. Every page is supposed to cite its sources, and the automated checks refuse a page that does not.
• Alykhan runs and publishes the site, sets the rules, chooses what to cover and deploys it. Alykhan does not read each page before it goes live.
• No lawyer, financial adviser, fraud investigator or other professional reviews this content.
• Charts, maps, infographics and share cards are drawn by code from data. No image-generation model is used.
• It is educational material, not legal, financial or investment advice. See the disclaimer.

What AI does here

AI drafts the 106 scam-type pages, 294 case files, 132 daily facts, 28 profiles of named people, 35 country reporting guides, 46 glossary terms and 10 articles on this site, and writes the text of the diagrams. For each, it searches for and reads the agency announcements, court records, regulator pages and news reports that the page then cites, with the date each was accessed.

A scheduled routine adds a new scam-type page most days. It runs unattended: no person reads the page between the routine writing it and the site publishing it. Separately, a daily news brief is written by a local AI model (qwen3.6) from headlines fetched that morning; it is marked auto-generated, and it is not reviewed by a person.

What AI does not do here

• No invented people. There are no fake authors, bylines, headshots, credentials or reviewers. Structured data names the publishing organisation, not a person, as the author.
• No AI-generated images. The charts, the map, the infographics, the fact and red-flag cards and the cover posters on scam pages are produced by scripts from data or from text specifications. The folder holding the posters is named /img/ai/ for historical reasons; they are code-drawn SVG graphics, not output of an image model.
• No named victims. And people appear only from the public record, with charges labelled as allegations.
• No operational detail that would help someone commit a scam.

Who is responsible

Alykhan is a real person who runs this site and its sibling, Clean on Paper. The contact address is [email protected]. Corrections are made by Alykhan and the AI tools; see the corrections policy.

How accuracy is checked

These checks run before each automated deploy, and that deploy stops if any of them fails:

• Sources are required. A scam-type page needs at least five sources and three documented cases. Every case file needs a source.
• Charges are allegations. A case or person marked as charged must say the charge is an allegation and the person is presumed innocent.
• Facts must trace back. Each number in a daily fact drawn from a scam or case page has to appear on that page, and every source it cites has to be one that page already cites; a fact with a figure that is not on its page fails the check. Facts computed directly from the FTC and FBI datasets are checked for their sources only.
• Diagrams match the text. The stages in a diagram must equal the stages in the page's prose.
• Internal links resolve. A link checker fails the build on any broken internal link.

These are process, not automatic checks:

• The instructions given to the AI require it to cite only sources it actually fetched and read, and to leave a claim out when it cannot verify it.
• Cases marked as charged were re-checked in one pass in October 2026, which found that 18 of 33 had moved on; those pages were updated. There is no automatic schedule for repeating it.

What an independent spot-check found

On 19 October 2026, 30 claims were drawn at random from the site: scam-page claims, case key facts, daily facts and country-guide facts, drawn with a fixed random seed. Separate AI reviewers tried to confirm each one against the sources its own page cites, and recorded what they found.

• 18 were fully supported by a cited source.
• 10 were partly supported: the main point held but a detail was missing, imprecise or different.
• 0 were not supported by the cited sources.
• 1 could not be judged because a cited source could not be read, and 1 were not checkable statements.

Of the 28 checkable claims, 18 (64%) were fully supported. With a sample this small, the true share across the whole site could plausibly be anywhere from about 46% to 79% (95% interval). Treat it as a rough indication, not a guarantee.

The reviewers are AI too, so this is a second independent read, not a human audit. The partly supported claims were mostly small additions the sources did not state, such as a cause, a year or a label, rather than wrong numbers. One of the partly supported items, the Israel reporting phone number, is confirmed only by a third-party directory and is not corrected, because the guide already tells readers to confirm it on the official government site. Where it found a claim the sources did not support, the page was corrected; see the corrections log below. The one claim it could not check, the India reporting portal, is unverified.

What is not checked

• No professional review. No lawyer, financial adviser or investigator reads these pages. Do not rely on them for a legal, financial or investment decision.
• No human reads each page before publication, including the pages added by the scheduled routine and the daily brief.
• Only a small share of claims has been independently re-checked, as above. The rest rests on the sources cited and the automated checks.
• No automated test for copied text. Sources are summarised in new words with short quotations in quotation marks, but the build does not scan for long verbatim passages.
• Pages can be out of date between re-check passes. Each carries a last-reviewed date.
• Loss figures are reported losses, which understate real totals. See the methodology.

Titles, descriptions and alt text

Page titles, descriptions, structured data and image descriptions are generated from the page's own content by templates, not written separately, so they carry the same claims as the page and are covered by the same sources. An audit script, run by hand, checks the built site for duplicate titles and descriptions.

Corrections log

Errors found after publication, and what was done. This is not a complete list of every edit.

• 29 September 2026. The compounding example on the Ponzi scheme page and the HyperFund case said $1,000 at 1% a day becomes more than $37 billion in three years. It becomes about $37,800 after one year and about $53.9 million after three. Corrected.
• 10 October 2026. The wire-recall glossary entry and the US reporting guide said IC3 advises reporting within 72 hours. That figure is FinCEN's observation about the success of international wire recalls, not an IC3 deadline. Corrected and sourced.
• 10 October 2026. A romance-scam case said eleven people were charged; it was ten, because one person appeared in two indictments. Corrected.
• 10 October 2026. Eighteen cases still described as charged were updated after a re-check found later pleas, sentences or settlements.
• 19 October 2026. The About page called the site "written and edited" by its operator and the daily page said the blog was written by a person. Both were wrong, since AI drafts the content, and both were corrected.
• 19 October 2026. The Think Finance case gave a 2014 suit date, 'three websites', a private equity defendant and a rate cap that none of its cited sources stated. Those details were removed.
• 19 October 2026. The charity pages said charities 'received under 2%'. The source says under 2% went to charitable purposes in some cases. Reworded on the case, scam page and fact.
• 19 October 2026. The pump-and-dump page listed 300% and 500% as separate promised returns; the indictment says 300 to 500 percent. Corrected.
• 19 October 2026. The celebrity-deepfake page implied a rule on celebrity disclosure; the rule covers material connections of insiders generally. Reworded.
• 19 October 2026. The Hi Mum page said the scam appeared around 2021 and spread through English-speaking countries. Its sources show regulator figures from 2022 in Australia and Canada only. Reworded.
• 19 October 2026. The SIM-swap fact and page said losses fell because carriers hardened the process. The FBI reports give no cause. The causal claim was removed.
• 19 October 2026. The invoice-fraud page said newer payment routes 'did not exist' when wire recall was designed, which no source states. Reworded. The mortgage-relief page tied the 1-888-995-HOPE line to HUD-approved counsellors more firmly than the FTC does. Reworded. The Snyder case said the penalty reflected inability to pay; the source says only that she had filed for bankruptcy. Reworded.

Report a mistake

If something here is wrong, email [email protected] with the page address, the sentence and, if you have it, the source that contradicts it.

This page was last updated 19 October 2026.