AI MADE IT POSSIBLE · EVIDENCE, DEEPFAKES & CRITICAL THINKING

The Trump–Natalie Harp Image Is the AI Lesson We’re Missing

The same technology that can manufacture evidence can also help us interrogate it. The difference is whether we use AI to reach a verdict—or to build an investigation.

A blurry golf-course image began circulating online. It appeared to show President Donald Trump kissing a woman many users identified as White House aide Natalie Harp. Then came a sharper version. Then the six fingers. Then the AI detectors. Then the arguments.

And almost immediately, the picture became bigger than the picture.

For people already suspicious of Trump, the image could feel like confirmation.

For people who support him, it could feel like one more fabricated attack.

But before Trump. Before Harp. Before politics. Before the Epstein files. Before anybody's tribe gets a vote, there is a smaller question that has to come first:

That question sounds almost too simple.

It is also the question that demonstrates what I think we keep missing about artificial intelligence.

AI is not just a machine that generates answers, images, music, text, video or code. It is becoming part of the infrastructure through which we create evidence, distort evidence, retrieve evidence, compare evidence and investigate evidence.

That makes it extraordinarily useful.

It also makes careless certainty extraordinarily dangerous.

This Is Exactly Why One Master Prompt Is Not Enough

I recently wrote a three-part series beginning with You Can't Prompt Wisdom.

The central warning is simple: asking AI to “act like a world-class expert” does not transfer years of judgment into you.

The same mistake appears here in a more consequential form.

Is this image real?

You can ask an AI that question.

It may inspect the image. It may notice anatomy, shadows, compression, perspective and texture. It may search reporting. It may give you a confident answer.

That can be useful.

It is not the end of the investigation.

You break the problem down.

This is not one magical prompt.

It is a conversation with evidence.

It is the same process I have been describing throughout the wisdom series: generate, challenge, retrieve, verify, compare, refine.

In Responsible AI Starts With the Controls You Put Around Your Own Work, I put it another way:

If AI gives you a bad fictional character name, change it.

If AI helps you publish a false allegation about a real person—including the President of the United States—the consequence is different.

Political Identity Is Not a Forensic Tool

This is where the story becomes uncomfortable.

People rarely arrive at politically charged evidence as blank slates.

If you strongly oppose Trump, you may know about documented court cases, allegations, controversies and relationships that make a new claim feel plausible.

If you strongly support Trump, you may know about years of hostile coverage, false claims, edited clips and partisan attacks that make another viral allegation feel immediately suspect.

Those contexts affect what people notice.

They still do not authenticate the image.

I do not think it helps to dismiss tens of millions of Americans who support Trump as stupid, immoral or incapable of understanding evidence.

People can look at the same country and prioritize different threats. They can trust different institutions. They can have different experiences with the economy, government, religion, immigration, war, crime, culture or media. They can hear the same argument repeatedly and still assign it less importance than something else.

Understanding that does not require agreement.

It requires enough curiosity to understand what another person actually believes before reducing them to the worst stereotype of their political tribe.

The same standard should travel in every direction.

A Trump supporter does not need to accept every allegation because critics consider it important.

A Trump opponent does not need to stop investigating documented questions because supporters consider the scrutiny partisan.

Evidence still has to do the work.

Smoke Is a Reason to Look for Fire. It Is Not Proof That There Is One.

The wider context around Natalie Harp illustrates why this distinction matters.

Reuters reported on September 9 that Trump gave a total of $45,000 in personal cash holiday gifts to three close West Wing aides in 2025, including Harp. The White House said the gifts were lawful, personal gifts unrelated to government work. Former White House ethics lawyer Richard Painter raised questions about the arrangement.

Reuters also described Harp as a trusted aide with close access to Trump and responsibilities around aspects of his Truth Social operation.

Those are documented facts and attributed concerns.

They do not establish a sexual relationship.

They do not authenticate the viral photograph.

They can be context worth understanding without being converted into proof of a different claim.

Source: Reuters, Sept. 9, 2026 — Trump gave cash gifts to Natalie Harp and other aides.

The Epstein Files Show the Scale of the New Problem

This is where the argument becomes much larger than one viral image.

The U.S. Department of Justice says its Epstein-related release under the Epstein Files Transparency Act now totals nearly 3.5 million pages, plus more than 2,000 videos and 180,000 images.

That amount of material would once have required teams of researchers enormous amounts of time merely to index.

AI changes the economics of that work.

One person can increasingly use AI-assisted tools to search millions of pages, extract names, locate dates, compare testimony, build chronologies, cluster organizations, identify repeated references, surface contradictions and generate lists of documents requiring human review.

That is extraordinary.

But DOJ also gives us a warning that could almost have been written for this article.

The department says responsive material can include items submitted to the FBI by members of the public and therefore may include fake or falsely submitted images, documents or videos.

The DOJ's own Epstein Library also warns that its search functionality has technical limitations and that some materials may not be reliably searchable.

This Is What AI Is Already Doing While We Argue About AGI

We spend an enormous amount of time discussing what AI might eventually become.

AGI. Superintelligence. Recursive self-improvement. Machine consciousness. Civilization-scale risk.

I have already argued elsewhere that those claims should be taken seriously without allowing forecasts to masquerade as established evidence. You can read that argument in If AI Is Going to Kill Us, Show Me the Evidence.

But there is another danger in staring too far over the horizon.

We can miss what this technology is already doing.

AI is digitizing trails that were previously too expensive, fragmented or time-consuming for ordinary people to follow.

Court records.

Government disclosures.

Corporate filings.

Interviews.

Podcasts.

Photos.

Video.

Archived webpages.

Financial disclosures.

Public databases.

Millions of pages of documents.

That information can increasingly be searched together, compared together and questioned together.

AI does not make the resulting theory true.

It makes investigation available at a scale that used to belong primarily to large institutions.

AI Can Help Us Understand People We Disagree With Too

This is not only about exposing wrongdoing.

AI can be used to reduce some of the intellectual laziness that political media encourages.

Take a long-form podcast such as Joe Rogan's.

You can watch a twelve-second clip selected because it outrages your political audience.

Or you can use AI to help work through complete transcripts across dozens of episodes and ask different questions:

• Which concerns recur?
• Which remarks were jokes and which were sustained arguments?
• What does the host repeatedly challenge?
• Where do guests disagree with him?
• Which claims are factual and can be checked?
• Which are value judgments?
• Which subjects consistently generate audience response?
• Does the full conversation support the interpretation created by the viral clip?

You still have to check the transcript.

You still have to listen when context matters.

You still have to verify factual claims.

But AI can help a person move from “What is wrong with these people?” to “What are these people actually saying?”

That is not surrender.

It is understanding.

And societies that have to live together need more of it.

The Creative Question Matters Too

There is another part of the Harp image story that creators should not ignore.

Humans have always created political representations.

Cartoons exaggerate. Satire invents scenes. Paintings symbolize. Comedy places public figures into situations that never happened.

AI did not invent that creative impulse.

What AI changes is fidelity.

A cartoon generally announces itself as interpretation.

A photorealistic synthetic image can travel without its original caption and arrive in somebody else's feed looking like documentation.

The same pixels can therefore change function depending on presentation.

“Here is a satirical image representing how I see this politician.”

is fundamentally different from:

“BREAKING: Here is photographic proof this happened.”

The creator's intent matters. So do labels, provenance and the way the next person reposts it.

That is one reason responsible AI creation cannot end at Can I make this?

Creators increasingly have to ask:

• What is this?
• What am I claiming it represents?
• Could a reasonable person mistake representation for documentation?
• What survives when my caption disappears?
• What responsibility do I still have after I press Generate?

AI Is a Tool. “Tool” Almost Feels Too Small.

I have described AI as a nuclear-level tool—not because every AI interaction is dangerous and not because I believe a chatbot is a nuclear weapon.

I mean it in a different sense.

Nuclear technology is not one thing.

It can produce electricity. It can contribute to medicine and scientific research. It can also produce catastrophic weapons.

The underlying capability matters enormously, but the application, controls, institutions and humans around it determine what that capability becomes in the world.

AI has a similar breadth.

It can create misinformation. It can detect patterns inside misinformation.

It can fabricate a photograph. It can help trace where that photograph traveled.

It can generate propaganda. It can compare propaganda with primary sources.

It can bury us in content. It can help us search through information volumes no individual human could reasonably process unaided.

Trust Can No Longer Be Passive

For most of photographic history, ordinary people were trained to ask:

Do I believe my eyes?

That is no longer enough.

We increasingly have to ask:

• Where did this file come from?
• Who first published it?
• What transformations happened along the way?
• What does the metadata or provenance system show?
• What independent evidence corroborates the event?
• What would prove my current interpretation wrong?

There is something almost paradoxical about that.

AI-generated misinformation may force us to develop better standards of evidence than many of us were using before generative AI arrived.

That opportunity only matters if we accept the discipline.

The Human Bottleneck Is Still the Human

AI can give you ten thousand sources and you can cherry-pick three.

It can reconstruct a timeline and you can ignore the inconvenient date.

It can show you the complete interview and you can repost the twelve-second clip.

It can tell you a photograph has uncertain provenance and you can caption it PROOF.

It can produce the strongest argument against your position and you can tell it to try again until it agrees with you.

No model can force intellectual honesty onto a person determined to avoid it.

The Master Prompt and the Political Tribe Have the Same Temptation

This is where this article comes back to my recent series.

The person searching for the perfect master prompt wants the answer before they have developed the judgment required to evaluate it.

The political tribe often wants the verdict before the investigation has finished.

Both tempt us to skip the uncomfortable middle.

The middle is where you ask another question. Where you discover the first source was wrong. Where the evidence weakens the claim you wanted to believe. Where somebody you disagree with makes a legitimate point. Where the AI confidently gets something wrong. Where an expert tells you your theory does not survive the available evidence. Where reality forces you to revise.

So What Do We Do With the Trump–Harp Image?

We do not use it as proof of an event that has not been authenticated.

We do not need to pretend the broader political context does not exist.

We do not need to decide that every question about Trump is illegitimate because one viral image may be synthetic.

And we do not get to treat an unverified photograph as true because other documented controversies make it feel believable.

We investigate the image on its own evidence.

We investigate other claims on their own evidence.

We preserve distinctions between facts, allegations, interpretations and unanswered questions.

And if better evidence arrives tomorrow, we update.

That is not weakness.

That is what critical thinking looks like when the world can manufacture convincing evidence on demand.

AI Made It Possible. Judgment Still Has to Be Earned.

I do not know whether artificial intelligence will eventually become the civilization-changing autonomous intelligence that dominates so much of the current debate.

I do know that it has already changed what a motivated individual can research, create, compare and investigate.

That is happening now.

We should be paying attention to it now.

Because AI can manufacture smoke.

AI can help us look for fire.

AI can also confuse steam for smoke and present the whole thing in a beautifully formatted report.

The difference is not simply the intelligence of the machine.

It is the process we build around it.

Ask. Investigate. Challenge. Retrieve. Verify.

Talk to people who know more than you.

Understand the people you disagree with before you reduce them to a caricature.

And change your conclusion when the evidence changes.

Maybe that is one of the most useful things AI can help us recover.

Not agreement.

A better argument.

One where the source matters more than the tribe, the evidence matters more than the caption, and the answer is allowed to change when reality gives us something new.

Both futures are already happening.

AI is making it easier to manufacture reality.

And AI is making it easier for ordinary people to interrogate reality.

What comes next depends partly on which habit we choose to build.