How ordinary citizens can use AI to investigate claims and make a difference

Most people believe they cannot affect what happens around them. Politicians introduce and pass legislation. Local councils create ordinances. Officials make decisions. Lies get repeated on social media and in the mainstream press, about organizations, about governments, about churches, and the people who know better say nothing, because what difference would it make?

It makes more difference than you think.

In 2007, two researchers, Nicholas Christakis of Harvard and James Fowler of the University of California, published a study of 12,067 people followed for thirty-two years in the Framingham Heart Study. They found that a person’s habits spread through his social network to three degrees: to his friends, to his friends’ friends, and to the friends of those friends.

They found the same pattern for smoking, for happiness, and for loneliness, and they wrote a book about it, Connected, in 2009. Everyone you know is connected to everyone they know. What you believe and talk about travels farther than you would expect.

If you have twenty friends, each of them has twenty, and each of those has twenty more, you are three steps away from as many as 8,000 people. They found the pattern in politics too. In one experiment, canvassers knocked on doors and urged whoever answered to vote. That person became about ten percent more likely to vote. The other adult in the house, who never heard the appeal, became about six percent more likely.

“Instead of each of us having only one vote,” the authors wrote, “we effectively have several.”

A network, they found, magnifies whatever is put into it. That is why lies move so fast. It is also why the truth can, if enough people pass it on and can prove what they say. Information is power when you can establish your point with facts or poke holes in a story someone wants believed.

For most of my life, having that kind of impact required a staff, money, access to a big library, and a mass media outlet that gave you a voice. Today it requires a phone, a free account, and discipline. This article is about that discipline.

The machine was confident. The machine was wrong.

I have been an investigator for almost sixty years. I started as an intelligence analyst in the Air Force, worked undercover inside extremist groups, ran investigations into organized crime, official corruption, drug trafficking, and money laundering, and directed the Illinois legislature’s multi-year investigation into the sexual exploitation of children.

For my whole career, I’ve used interviews, surveillance, personal observations, library and court record research, and written or made in-person requests for public records from government agencies to gain insight into the targets of my inquiries. I still do. In addition, since they became available, I’ve used internet search engines like Brave, Yahoo, Bing, and Google. More recently, AI has largely replaced search engines, except in certain circumstances, such as searching for Russian or Chinese sources.

I have used artificial intelligence every working day for two years. Although it is the most powerful research tool I have ever had, I do not trust it and always remain suspicious of its findings.

Last year I ran a test to determine AI’s reliability. I asked an AI platform to describe my work for the United States Senate in 1997, when the Permanent Subcommittee on Investigations retained me to investigate foreign money being used to fund President Clinton’s reelection. The primary target was the Lippo Group and its Chinese partners. During the hearing, Senator John Glenn asked me to help locate witnesses who had fled to Asia. On that background, on my testimony, on this exchange with Senator Glenn, on every element of the case, the AI got it right.

Then I asked a second AI to fact-check the first. It returned two confident corrections and cited Senate documents to support both. It said Glenn could not have asked for my help, because Senator Fred Thompson ran that investigation. It said the full Governmental Affairs Committee had conducted the inquiry, so the Subcommittee could not have hired me.

Both corrections were wrong.

The Subcommittee is part of the Governmental Affairs Committee. Glenn did ask. It’s in the record. Nevertheless, the second AI’s answer sounded more convincing than the first.

Here is the problem. AI does three jobs well.

• It collects: give it a question, and it finds documents, filings, studies, and statutes in seconds or minutes.
• It organizes: give it two hundred pages, and it builds the timeline, extracts every name, and puts every dollar figure in an Excel table.
• It challenges: give it your draft and tell it to find every error, every gap, and every claim you cannot support, and it will.

One job it must never do. It must never decide what is true. An AI produces the most likely next sentence. Usually the most likely sentence is also correct. Sometimes it produces a sentence that sounds correct, cites a document that doesn’t exist, and reports it with the same confidence as everything else. The machine often can’t tell the difference. That’s your job.

AI collects. You conclude.

The tax returns that killed my theory

This fall I prepared a lecture for graduate students at the Jane Addams College of Social Work on charities that lose their mission. The opening case was Hull House, which Jane Addams founded in 1889 with her own money and which closed in 2012 as a government contractor drawing more than ninety percent of its revenue from public contracts.

I wanted to compare how Hull House and five peer agencies spent their money across a decade: about thirty scanned tax returns. An AI system extracted revenue, government share, executive pay, and overhead into one table in about an hour. Then I opened the filings myself and double-checked the figures.

I had a theory going in. In the 1990s, a former director of the Illinois Department of Children and Family Services led the Hull House Association. Its budget roughly quadrupled during that decade, while the number of children in state care more than doubled. My theory was that foster care drove the growth, and I already had that in my draft. Then the AI read the returns. Foster care accounted for 12 percent of program spending in the Association’s biggest year. The growth came from government contracts across the board. I found out two weeks before the lecture, not two weeks after.

I used the same method to build my IFI series on the hunger industrial complex. I assembled a database of every state-designated regional food distributor under the federal Emergency Food Assistance Program: 264 organizations, 1,314 tax filings across five years. No one reads 1,314 tax returns. An AI extracted the figures. I verified the ones that carried weight, including the officers who appear on more than one board. That check found the keystone. Kate Maehr sat at all three tiers at once: the board of Feeding America, the leadership of Feeding Illinois, and the chief executive’s office of the Greater Chicago Food Depository. The AI found the connection. The filings proved it.

One word through three institutions

In my series Assault on Our Language, I traced the term “chestfeeding” from its origin through the institutions that adopted it. I followed the broader gender-neutral revisions into Illinois law. Who started it, and how did one activist’s word become mandatory in a professional journal?

The AI read years of journal issues, conference programs, and organizational statements and created a list of every author, editor, board member, and institution that used the term. The record led to Aiden Farrow, a transgender man (a woman) who was the original and individual source of the revised language; to the International Lactation Consultant Association, which made that language its professional standard in 2017; to the Academy of Breastfeeding Medicine, whose 2020 protocol concedes that “significant research gaps exist in this field”; and to the Journal of Human Lactation, which imposed a language mandate on its authors in 2021.

One person, Kathleen Marinelli, held positions in all three institutions in that chain. A connection, through which a word traveled from a fringe figure into the standards of a profession, and from there to the CDC and the hospital intake form.

“Chestfeeding” was one word in a broader revision. Once the professions had adopted the new vocabulary, the law followed. Illinois Public Act 104-0448 struck “father” 55 times and “mother” 34 times from the state’s parentage law, and where the law said “mother,” it now says, “woman or person who gave birth.” Public Act 104-0850 requires construction employers to provide special accommodation for “women and individuals who menstruate.” Read plainly, that covers trans women (men who identify as women), trans men (women who identify as men), and biological women.

I found Marinelli using a link chart. Every person, organization, document, and event goes in with its source attached. By the third version, the chart held 138 entities and 177 links. That chart linked her to the three institutions.

Six theories and a county resolution

A few weeks ago I was assigned to investigate a proposed development on about 165 acres near Sheridan, Illinois, promoted by a company called Vairt. I looked at six elements: the real estate transactions, zoning, financing and securities, infrastructure, governance, and the reliability of the claims made by the sellers and by their critics.

Reviewing these six components produced hundreds of documents in a few weeks. The AI helped me keep them organized and easy to access. The project was also emotionally charged for the people in the area of the proposed development. If I started with a position hostile to the development, the AI would have no trouble building a case against it, maybe using fabricated information. It will help you back up a case that you already believe based on your own bias.

So I used a method we routinely used in intelligence: analysis of competing hypotheses. I wrote six explanations for what the record showed, from the most innocent to the most alarming, and had the AI build a matrix: every piece of evidence down one side, every hypothesis across the top, and a judgment in each cell about whether that evidence supported, contradicted, or said nothing about that hypothesis. The matrix forces you to look for the evidence that would kill your favorite theory.

Some hypotheses died. Some survived. What survived went into my article on the resolution the LaSalle County Board later passed, and the article held the sellers and the critics to the same standard. When someone asked why I hadn’t reached the conclusion they wanted, I could point to the record.

Epstein

For several years I have been investigating Jeffrey Epstein. You have heard that he was a rich man with an appetite for young girls and friends who shared it, and that the only remaining question is which politician from which party more frequently flew on his plane or was a guest at one of his properties. That is the story most of the press tells, and the story each side uses against the other.

The record shows something much larger and far different. What I believe, and cannot yet prove, is that the dirty-old-man story distracts from international corruption on a scale never before seen, involving politicians, intelligence services, bankers, business, media, and entertainment figures on almost every continent. That is a theory. I am doing my best to disprove it. So far, I have not been able to.

Here is some of what the record shows. JPMorgan Chase dropped Epstein as a client in 2013 and waited until 2019, after his arrest, to report his activity to regulators. While he was a client, the bank filed seven suspicious activity reports flagging $4.3 million. After his death, it filed reports on more than 5,000 wire transfers totaling about $1.1 billion.

Federal law requires that reports be filed within thirty days.

The bank’s head of anti-money-laundering compliance tried for years to end the relationship and was overruled by senior leadership. Deutsche Bank retroactively flagged more than $250 million and paid a $150 million penalty to New York regulators. Bank of America waited five to seven years to flag about $170 million in wires from Leon Black to Epstein, then wrote that “the wire transfer activity does not have a verifiable business purpose.”

This is in the report U.S. Senator Ron Wyden released in August from the U.S. Senate Finance Committee.

Epstein’s closest business partner, Jean-Luc Brunel, ran a modeling agency on three continents and died in a Paris jail in 2022 awaiting trial for raping minors. The Department of Justice’s own Office of Professional Responsibility later faulted the 2008 Florida plea deal for “poor judgment” (AP), and a federal judge ruled in 2019 that hiding it from the victims broke the law. Epstein had documented relationships with people in or near the intelligence services of at least three countries.

Now consider the mass of records available. The Department of Justice released more than three million pages this past January, with 180,000 images and 2,000 videos. Add court filings in four countries, twenty years of flight logs, depositions running hundreds of pages, the bank records Congress released, and the tax returns of a dozen foundations. Ten years ago that took a government agency or a newspaper with scores of reporters.

When the government and newspapers decline to investigate, nobody does. But I can do it with the tools named below. So could you.

One system collects: I ask it a question and a list of targets. It returns filings, transcripts, articles, reports, and other documents, each with an identified source. I have a second AI from a different company, using a different model, read what the first one found. It checks the information against the records it locates and tells me where the original AI got it wrong or right. I take this assessment back to the original AI and ask it to respond.

I keep a record of every hypothesis: what supports it, what points in another direction, what rules it out. Every name, company, and transfer goes into a link chart with its source attached, so that when two names turn up at the same address in the same year, I see it.

The tools have translated hours-long documentaries recorded in other languages for me. And every piece of information is given one of three labels—documented, inferred, do not assert.

The TerraMar Project, Ghislaine Maxwell’s ocean charity, shows the method on a small scale. Its tax returns were public for years, but apparently nobody had read them closely. I had an AI extract the financial schedules, and the schedules showed Maxwell financing the organization herself through interest-free loans never disclosed as such. The machine found the line items. I double-checked the returns myself.

Without the tools, the Epstein investigation is impossible for one person. Without following the rules I’ve laid out, any “findings” would be worthless.

The rules

Start with a question, not a prompt. Most people open an AI and type whatever is in their head, and the tool answers with something that may or may not be useful. Before I do anything, I write down the issue, the questions I need answered to resolve it, and the targets—the people, organizations, and documents those questions point at. Every query serves one of those questions. “Who funds this curriculum?” is a question. “Tell me about this curriculum” is a prompt. The first produces a lead. The second produces a report.

Instruct the AI to show its sources. Then check them. A citation from an AI is a lead. AIs have provided me with citations to court cases never filed, studies never published, and page numbers in real documents that say something else. I have also been given exactly the right document on the first try. You cannot tell which until you look. When you do, read what comes before and after. Then ask: does this document prove the fact, or record someone making a claim? A newspaper quoting an activist is a claim. A court filing, a tax return, a budget, a signed contract is a fact.

Label every claim. Documented: I have examined the primary record, and it says this; I can provide it to you. Inferred: the documented facts point to this conclusion. “The two organizations share an address, a treasurer, and a donor. That suggests coordination.” Suggests. Do not assert: I believe it, I may be right, and I cannot support it with anything specific. It stays out of the article, the speech, the email, and the letter. This is where advocates get hurt. They repeat something they heard from someone they trust; it turns out to be wrong, and from that day the other side can undermine your credibility on everything you say. One unsupported claim can cost you ten documented ones.

Keep collection separate from conclusion. I collect everything, including what hurts my position. The evidence that would prove my theory wrong is the evidence I look for most intensely. I develop a full range of theories for what is already known, and as information comes in, I rule those theories out one by one.

If I conclude early, I stop looking for what would prove me wrong, and so does the AI, because it follows my lead.

Use two AIs to attack your draft. When the draft is done, I give it to two AI platforms other than the one I researched with and ask each to find every factual error, every logical gap, and every unsupported claim. Errors get fixed. I close gaps or cut the claim. Unsupported claims go to do not assert. Then I rewrite in my own voice. The argument has to be mine, so that when someone challenges me, I can defend it.

The tools

Every tool I use has a free version.

Perplexity, for collection. Ask a question; it searches and returns an answer with live-link citations.

Gemini Notebook, formerly NotebookLM, from Google, for your own documents. Upload a bill, a curriculum, a set of court filings, a stack of UN reports, and it answers only from what you uploaded, with a citation to the passage. For a volunteer, this is the most useful tool on the list.

Claude and ChatGPT, for analysis and drafting. I use one to analyze and the other to attack what I write. Gemini, for images, quick prompt-building, and a third critique.

One tool to collect, one to analyze, two to challenge. When three systems built by three different companies agree with each other and with the document in front of you, you are on reasonably solid ground. You still confirm against the sources yourself, and when the systems disagree, that is where to look hardest.

Three degrees

Go back to Christakis and Fowler. If what you say reaches as far as your friends’ friends’ friends, then a parent who speaks at a school board meeting with documented testimony reaches more people than they would ever imagine.

The news does not reach most people directly. Researchers who studied elections in the 1940s found that it reaches them through someone they know: a relative, a friend, a coworker who pays attention and tells the rest what it means. Christakis and Fowler found the same thing sixty years later. People are moved less by what they see in the media than by what the people around them say and do. You are that person for somebody, probably for many people.

Suppose a story about an organization you know is being repeated, and you know it is false. Make the AI find the primary source behind the story. Look at it. If the story fails against its own source, you have what you need, and you can pass your findings on to your friends, who pass them on to theirs. This is how false information is corrected.

Most people don’t analyze the information they read, hear, or see, especially if it comes from a source they trust. They take it at face value, even though they know many outlets spin the stories to suit a narrative they push. Misinformation is spreading everywhere, and all sides are responsible. Here are a couple of examples:

Covington Catholic, January 2019. A short clip showed Nick Sandmann, a Kentucky high school student in a MAGA hat, face-to-face with a Native American activist at the Lincoln Memorial. Major outlets reported that the students had surrounded and taunted him. The full video showed the activist had walked up to the students. Sandmann sued for defamation; CNN settled in January 2020, and The Washington Post settled later that year.

Springfield, Ohio, September 2024. A secondhand Facebook post about a neighbor’s cat became a national claim that Haitian immigrants were eating pets, and it was repeated from the presidential debate stage. Springfield police and city officials said they had received no credible reports. The woman who wrote the original post later said she had no firsthand knowledge.

In both of these instances, anyone reporting on them could easily have determined whether they were true or false before passing them on. Instead, both false claims were spread far and wide. In some cases, even lies that have been proven false and widely reported as such persist.

In August 2014, Officer Darren Wilson shot and killed Michael Brown in Ferguson, Missouri. Within hours, the account spread that Brown had his hands raised in surrender and said, “don’t shoot.” “Hands up, don’t shoot” became a national slogan, repeated by members of Congress on the House floor and by professional athletes on the field.

On March 4, 2015, the Department of Justice under Attorney General Eric Holder released its report on the shooting. It found no credible evidence that Brown had his hands up in surrender. The physical and forensic evidence contradicted the witnesses who said so, changed their accounts, or admitted they had not seen the shooting. The Washington Post’s fact-checker gave the claim four Pinocchios, and columnist Jonathan Capehart, who had repeated it, wrote that it was “built on a lie.”

Eleven years later, the phrase is still chanted at protests and posted every August on the anniversary as a statement of what happened.

Defeating these lies and countering them with the truth is every citizen’s responsibility, especially every Christian who claims to represent the truth.

Before you pass it on

Christakis and Fowler found a second thing, and it explains every rule above. Influence fades after three degrees, partly because information decays on the way. They compare it to the children’s game of telephone: each child passes along what he heard, with his own errors added to everyone else’s. By the fourth person, nobody has an accurate account.

Social media is that game played at the speed of light. So the rules that govern an investigation govern the share button. Before you pass on a post, your own or anyone else’s, find the source behind it and examine it. If you cannot find one, do not pass it on.

That a friend posted it, that it has been shared ten thousand times, that it says what you already believe: none of that is evidence. The same research found that a crowd gets things right when each person judges independently and gets things wrong when each follows the one before, which the authors call the blind leading the blind. It also found that people use online political networks more to reinforce what they already think than to test it. The post that confirms your side is the one to check twice.

When you find a false post, correct it, whoever posted it, your own side included. State what is wrong. State what is correct. Link the record that proves it. “This is misleading” corrects nothing, and neither does an insult.

The book describes what happens when ordinary people do this. A television host joked on the air that the world’s elephant population had tripled. Within minutes, the Wikipedia entry said so. Volunteers changed it back, the pranksters tried again, and the people committed to correct information won. An encyclopedia that anyone can vandalize stays accurate because thousands of people take the trouble to fix what is false and explain why.

When the false post is yours, correct it the same way, in the same place, as publicly as you made it.

For a Christian, none of this is optional. “Do not spread false reports. Do not help a guilty person by being a malicious witness” (Exodus 23:1). “Test them all; hold on to what is good” (1 Thessalonians 5:21). The ninth commandment makes no exception for false witness against people we oppose, or for a report we did not write and only forwarded. How we say it stands under the same authority: “speaking the truth in love” (Ephesians 4:15), with no insults, no exaggeration, and no motives assigned that we cannot document.

The same test applies to what we advocate. Before you promote a policy or condemn one, hold it against Scripture. A bill is not right because our side sponsored it or wrong because the other side did. If a policy conflicts with a biblical worldview, explain how. If it conforms, say that too; if you cannot show which, don’t make the claim.

What I am asking you to do

This week, pick one fact you have repeated. Ask an AI to find the primary source. Open it. Read it; if you can document it, great. If you can’t, stop saying it today. Correct it wherever you said it, and be grateful you found out. Then do the same with the next post you are about to share, before you share it.

Next, if you are led to become engaged in our culture, which I hope you will be, pick one issue in your own community: a school board policy, a county contract, a program that keeps growing while the problem it was created to solve does not shrink, a story about a church or a nonprofit that does not match what you know.

Write the questions first. Instruct the AI to collect the information that answers your question. Read what it finds. Label what you learn. Give it to a second AI and tell it to tear the argument apart. Then write it up yourself, in your own voice, and send it to the people who can act on it and to the people who trust you.

You cannot withdraw from the world because you think you do not influence it. You have three degrees of it. You have the most powerful technology ever built sitting in your pocket, free, waiting for a question. Put it to work for the truth.

Remember: AI will make you faster. It will not make you right. Being right is still your job.