Tag Archive | fact-checking

AI in the Research Room: Finding, Testing and Verifying Information

By Cheval John | Vallano Media
This article was created with the assistance of artificial intelligence.
The final version was reviewed, edited and fact-checked by the author


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

Research used to mean opening tabs.

A lot of tabs.

Search results.

PDFs.

Old articles.

Statistics.

Official websites.

Notes scattered across documents.

Then more tabs.

Artificial intelligence changed that part of my work faster than almost anything else.

Not because AI replaced research.

It did not.

It changed how I move through research.

It can help me find possibilities.

Organize information.

Compare sources.

Spot contradictions.

Generate follow-up questions.

Identify what I still do not know.

But the most important lesson I learned is also the simplest:

“AI can help me find information. It cannot decide what counts as evidence.”

That distinction changed the way I research.

Research Is Not the Same as Searching

For years, online research often began with a search engine.

Type the question.

Open the results.

Read.

Compare.

Take notes.

Search again.

That process still matters.

But AI adds another layer.

Instead of asking only:

“Where can I find information about this?”

I can ask:

“What claims do I need to verify?”

“What sources would be most useful?”

“What parts of this issue are disputed?”

“What questions am I not asking?”

That changes research from a simple hunt for information into a more structured investigation.

The goal is no longer to collect as much information as possible.

The goal is to identify the information that matters.

AI Is Useful at the Beginning

One of the most useful places for AI is at the beginning of a research project.

When I know very little about a subject, AI can help me map the terrain.

Who are the major organizations involved?

What terminology should I understand?

What historical events might matter?

Which documents should I look for?

What questions would a reporter normally ask?

That can save time.

Instead of wandering through the subject without direction, I can start with a rough map.

But a map is not the territory.

AI may omit something important.

It may misunderstand the subject.

It may treat a disputed claim as settled.

It may suggest a weak or nonexistent source.

So that map is a starting hypothesis.

It is not a verified research plan.

It gives me somewhere to begin.

Not somewhere to stop.

Turn the Topic Into Claims

One of the biggest improvements in my research process came when I stopped treating an article as one large question.

Instead, I break it into claims.

What am I actually asking the reader to believe?

Maybe the article contains a claim about a date.

A statistic.

A historical event.

A policy.

A quote.

A trend.

A cause.

A comparison.

Each claim creates its own research problem.

That matters because different claims require different evidence.

A score can be verified through official statistics.

A historical claim may require archives.

A quotation requires the original interview, transcript or recording.

A policy claim may require an official document.

A trend may require data across time.

AI can help me identify those claims.

But once the claims are visible, I still need to find the evidence that supports them.

Ask AI What Would Prove It

A useful research question is not:

“Is this true?”

A better question is:

“What evidence would prove or disprove this?”

That changes the task.

If I ask whether something is true, AI may simply give me an answer.

If I ask what evidence would establish the claim, I am forced to think about the source.

What document should exist?

Who would know?

What data would show it?

What record could confirm it?

What evidence would contradict it?

That question turns AI into a research assistant rather than an authority.

Research the Claim Against Yourself

There is another part of that process.

Do not only ask what supports the claim.

Ask what could weaken it.

Complicate it.

Or overturn it.

Once I form a theory about a story, it becomes easy to see the evidence through that theory.

If I think a trend exists, I should ask what evidence would show that it does not.

If I think one event caused another, I should look for alternative explanations.

If I think an institution changed policy at a particular moment, I should ask what evidence might show that the change happened earlier or for another reason.

AI can help generate those challenges.

But it cannot decide whether the challenge is valid.

It can tell me where to look.

I still have to investigate.

Source Hierarchy Still Matters

Not all sources are equal.

A social media post is not the same as an official record.

A press release is not the same as independent reporting.

A news article summarizing a study is not the same as the study itself.

AI can help me locate different types of sources.

But I still have to evaluate them.

Who produced this?

Why?

When?

What evidence does it rely on?

Is the source firsthand?

Is it independent?

Is it current?

Does the source have something to gain from the claim?

Those questions existed long before generative AI.

They matter even more now because AI can flatten very different kinds of sources into one smooth answer.

“AI can flatten source quality. The journalist has to restore the hierarchy.”

Smooth Summaries Can Hide Uneven Evidence

One of AI’s strengths is synthesis.

It can take several sources and produce a clean summary.

That is useful.

It can also be dangerous.

A summary may make five sources sound equally credible even when they are not.

One might be an official record.

Another might be commentary.

Another may simply repeat the first source.

Another may be outdated.

Another may contain an error.

The AI summary can erase those differences.

That means I have to separate the summary from the evidence underneath it.

If the summary contains an important claim, I want to know:

Which source supports it?

Is that the original source?

Are the other sources actually independent?

Do they confirm the same thing?

The cleaner the synthesis looks, the more important it may be to inspect what produced it.

Fast Synthesis Can End Research Too Early

There is another danger.

AI can make the research feel finished before it is finished.

The answer appears quickly.

The sources look organized.

The explanation sounds coherent.

And suddenly it feels as if the subject has been solved.

That feeling can be misleading.

Fast synthesis can create premature research closure.

The journalist stops looking because the answer already looks complete.

That is another version of the same problem that appears in AI-assisted drafting.

Polish creates confidence.

Confidence creates stopping.

But the evidence may still be incomplete.

The research is finished when the reporting supports what I intend to publish.

Not when the AI response looks convincing.

Finding Contradictions

One of my favorite research uses for AI is comparison.

If I give it multiple documents, I can ask:

Where do these accounts disagree?

Which dates conflict?

Which statistics differ?

What changed between two versions of the same document?

That can help surface problems quickly.

But AI cannot automatically tell me which version is correct.

A contradiction is a signal.

Not a conclusion.

If two sources disagree, the next question is:

Why?

Maybe one is wrong.

Maybe one is outdated.

Maybe they use different definitions.

Maybe they cover different time periods.

Maybe the denominator changed.

Maybe they are not actually describing the same thing.

That last question matters.

“Are these sources really contradicting each other, or are they measuring different things?”

Research begins where the contradiction appears.

Use AI to Generate Follow-Up Questions

Good research creates more questions.

That is another place AI can help.

After reviewing the evidence, I can ask:

What is still unresolved?

What would a skeptical reader want to know?

What question would an editor ask?

What does this source not explain?

What additional document could clarify this?

What assumption am I making?

Those questions can lead me toward stronger reporting.

Sometimes the most useful AI output is not an answer.

It is the next question.

A Missing Record Does Not Automatically Disprove Something

Research has another trap.

You search for evidence.

You cannot find it.

So the temptation is to conclude that it does not exist.

That can be dangerous.

Records may not be digitized.

Archives may be incomplete.

A page may have disappeared.

The relevant document may use different terminology.

The event may be documented somewhere you have not searched yet.

So failure to find evidence is not automatically evidence that something never occurred.

Sometimes the correct conclusion is simply:

“I have not found evidence for this claim.”

That is different from:

“This claim is false.”

The distinction matters.

AI-Suggested Sources Are Leads

AI may recommend reports, studies, archives, websites, experts or documents.

That can be extremely useful.

But the recommendation itself proves nothing.

A suggested source may be outdated.

It may be irrelevant.

It may be mischaracterized.

It may not exist in the form the AI described.

So I treat an AI-suggested source as a lead.

Not a credential.

I open it.

Inspect it.

Confirm what it is.

Then decide whether it belongs in the research.

Do Not Let AI Replace Reading

There is a temptation once AI becomes good at summarizing.

Why read a 70-page report when AI can summarize it in two minutes?

Because the details matter.

A summary may miss a footnote.

A limitation.

A definition.

A qualification.

A methodology problem.

A sentence that changes the meaning of the conclusion.

If a document is important enough to support the article, it is important enough for me to inspect directly.

AI can help me navigate long documents.

It can tell me where to look.

It can identify relevant sections.

It can help explain technical language.

But I do not want AI deciding what the document means for me.

Use AI to navigate the document.

Do not outsource the interpretation of the evidence.

Search Results Are Leads, Not Proof

AI-assisted search can make research feel incredibly fast.

Ask a question.

Receive sources.

Open links.

Move on.

But every source surfaced by AI should be treated as a lead until I inspect it.

Does the page exist?

Does it say what the AI claimed?

Is it the correct date?

Is the quotation complete?

Is the source current?

Does the source itself cite someone else?

That last question matters.

Sometimes what looks like three sources is actually one source repeated three times.

Ten sources are not necessarily stronger than three if all ten trace back to the same underlying claim.

Research is not counting links.

It is tracing information back to where it originated.

The Original Source Matters

Whenever possible, I want to get closer to the original evidence.

If a news article says a university made an announcement, I look for the university announcement.

If a story cites a government report, I look for the report.

If a statistic appears in several articles, I look for the original dataset.

If someone quotes an interview, I look for the interview.

That does not mean the original source is automatically truthful.

An institution can frame information in its own interest.

A company can highlight favorable data.

A government agency can make mistakes.

A court filing can contain allegations rather than established facts.

Closer to the source does not always mean closer to the whole truth.

Original evidence gets me closer to what was actually recorded or stated.

It does not remove the need to evaluate credibility, motive, completeness or context.

AI Can Help Organize the Evidence Trail

Research becomes difficult when the evidence pile grows.

Which source supports which claim?

Where did the statistic come from?

Which article was secondary?

Which document was original?

Which fact still needs confirmation?

AI can help organize that.

I can build a claim list.

A source table.

A chronology.

A list of unresolved questions.

A distinction between confirmed facts, interpretations and unknowns.

That organizational role may be one of AI’s most practical research benefits.

It reduces the mental load of managing information.

But organization is not verification.

A beautifully organized false claim is still false.

Know When the Research Is Not Finished

One of the hardest decisions in journalism is knowing when you have enough.

There is always another source.

Another search.

Another document.

Another question.

AI makes it easier to keep searching forever because it can constantly generate new angles.

That can become its own problem.

The goal is not complete knowledge.

That is impossible.

The goal is enough reliable evidence to support what you intend to publish.

Sometimes the evidence is strong.

Sometimes it is incomplete.

Sometimes the honest conclusion is:

We do not know yet.

That answer can be frustrating.

It can also be the most accurate result of the research.

Research Should Leave a Trail

One thing I increasingly value is the ability to explain how I know something.

What source supports the claim?

Where is the original document?

What evidence did I reject?

What remains uncertain?

Why did I choose one source over another?

That creates a verification trail.

Not necessarily something every reader will see.

But something I should be able to reconstruct.

That matters in AI-assisted journalism because the final prose may look effortless.

The evidence behind it should not be.

The Research Room Has Changed

AI has changed what happens before I write.

I can move through large amounts of information faster.

I can organize sources more efficiently.

I can compare documents quickly.

I can generate questions I might have overlooked.

I can test the strength of an argument before I commit to it.

But none of those capabilities remove the basic responsibility of research.

Find the source.

Read it.

Test the claim.

Look for contradictions.

Understand the context.

Verify what matters.

And know when the evidence is not strong enough.

AI can help me move through the research room.

It cannot replace the research room.

Because journalism does not begin with the answer.

It begins with the evidence.


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

From Draft to Publish-Ready: A Practical Editorial Workflow

By Cheval John | Vallano Media
This article was created with the assistance of artificial intelligence.
The final version was reviewed, edited and fact-checked by the author


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

A draft is not a finished article.

That sounds obvious.

But artificial intelligence makes the difference between the two easier to forget.

AI can produce polished language quickly.

It can organize ideas.

It can summarize research.

It can suggest structure.

It can even make a rough argument sound complete.

And that creates a dangerous illusion.

The article looks finished.

So the writer begins treating it as finished.

That is why I needed a workflow.

Not a prompt.

Not a shortcut.

A workflow.

A repeatable process that takes a draft through research, verification, challenge, revision, editorial judgment and final approval before publication.

That process became one of the most important changes in how I work with AI.

But there is an important warning from the beginning.

“A workflow does not guarantee accuracy.”

It creates deliberate places where errors, weak evidence and poor judgment have a better chance of being caught.

The quality of the work still depends on what happens inside those stages.

The Draft Is Only the Beginning

Before AI, a rough draft often looked rough.

The unfinished work was visible.

Sentences needed rewriting.

Transitions were missing.

Notes remained in the copy.

The weakness of the draft was easier to see.

Generative AI changes that.

A first draft can arrive looking clean.

The paragraphs may flow.

The grammar may be correct.

The tone may sound professional.

But polished writing does not mean finished journalism.

The draft may still contain:

unsupported claims,

missing context,

weak sourcing,

overstatement,

incorrect assumptions,

unclear attribution,

or conclusions stronger than the evidence.

The surface can look complete while the reporting underneath remains unfinished.

That is why I stopped treating the first AI-assisted draft as something close to publication.

It is simply the next thing that needs to be examined.

My Workflow Starts Before the Draft

A good editorial process does not begin with writing.

It begins with the question.

What am I trying to establish?

What evidence would support it?

What do I already know?

What do I need to verify?

Which sources should I find?

What would change the direction of the story?

AI can help organize those questions.

It can help identify gaps.

It can help create a research plan.

But in my workflow, I still decide what the story is actually about.

That matters because a weak question produces a weak workflow.

If I start with the wrong premise, moving faster through the process does not solve the problem.

It only gets me to the wrong destination sooner.

Stage One: Research

The first stage is evidence.

I gather the material that can support the story.

That may include:

official records,

statistics,

transcripts,

interviews,

reports,

news coverage,

or firsthand observations.

AI can help me organize that material.

It can compare documents.

It can identify names, dates, contradictions and patterns.

It can suggest follow-up questions.

But the research stage has one non-negotiable rule:

“AI output is not the source.”

If the AI gives me a fact, I still need to know where that fact came from.

If it summarizes a document, I still need to inspect the document.

If it gives me a quotation, I need the original quotation.

If it gives me a statistic, I need the original data.

Research is where evidence enters the workflow.

Everything afterward depends on the quality of that evidence.

Stage Two: Draft

Once the research is organized, I move into the draft.

This is where AI can save a considerable amount of time.

I can use it to help organize the structure.

I can test different openings.

I can explore different ways to arrange sections.

I can use it to help turn notes into readable prose.

But I do not ask AI to decide what the story means.

That decision remains mine.

The draft is a construction stage.

The goal is not perfection.

The goal is to create something complete enough to examine.

That distinction matters.

If I expect the first draft to be perfect, I may spend too much time polishing sentences that later need to disappear.

If I treat the draft as provisional, I can focus on building the argument first.

Then I can test it.

Stage Three: The Truth Pass

The next stage is what I call the truth pass.

This is where the draft stops being treated primarily as writing and starts being treated as a collection of claims.

I ask:

Which statements are factual?

Which ones need evidence?

Which are interpretation?

Which depend on current information?

Which could be misunderstood?

Which are stronger than the evidence allows?

Then I verify.

Some claims are easy.

A date can be checked.

A score can be confirmed.

A public document can be opened.

Other claims require more judgment.

A historical interpretation may need multiple sources.

A causal claim may require stronger evidence.

A quotation may be accurate but misleading without context.

The truth pass is not simply:

“Did I make a factual error?”

The better question is:

“Can I defend what this sentence is asking the reader to believe?”

That changes the way I edit.

Stage Four: Source Verification

Verification deserves its own stage because it is easy to confuse research with verification.

Finding information is not the same as proving it.

During verification, I trace important claims back to their strongest available evidence.

I open the source.

I read it.

I confirm the date.

I examine the context.

I check whether the source actually supports the wording in the article.

I look for independent confirmation when the claim warrants it.

But the existence of a verification stage does not automatically make the journalism verified.

“The quality of the verification matters more than the existence of the checkbox.”

Opening one link is not enough if the claim requires deeper scrutiny.

Finding one citation is not enough if several websites are simply repeating the same underlying source.

A citation tells me where to look.

It does not tell me whether the claim is correct.

This stage is where weak evidence becomes visible.

Sometimes the sentence survives.

Sometimes it becomes more cautious.

Sometimes it disappears.

The evidence decides.

The Stages Can Overlap

The workflow may look linear when written down.

Real journalism usually is not.

Some claims are verified before the first sentence is drafted.

Others emerge during the writing and send me back to research.

A red-team pass may uncover a missing source.

Editing may expose a factual statement I did not realize needed verification.

That is normal.

The stages are useful because they create discipline.

They are not a claim that journalism always unfolds in a straight line.

The workflow moves forward.

Sometimes it also loops backward.

Stage Five: Red-Team the Story

Once the major claims have been checked, I challenge the article.

This is the red-team stage.

I ask AI to look for:

weak arguments,

unsupported inferences,

missing context,

overstatement,

plausible alternative interpretations,

and headline problems.

But I do not accept every criticism.

AI can be wrong.

It can manufacture a counterargument simply because I asked it to challenge something.

It can give too much weight to a weak objection.

It can misunderstand the article.

So the red-team stage has two parts.

First:

Challenge the story.

Second:

Challenge the challenge.

The goal is not to obey the AI.

The goal is to expose the article to resistance.

Then I decide which criticism matters.

Every AI-Assisted Stage Can Introduce Error

This is another reason the workflow matters.

AI errors are not confined to the draft.

AI can misread research.

It can produce an incomplete summary.

It can suggest a structure that exaggerates one part of the evidence.

It can overlook contradictory information.

It can recommend a headline that is stronger than the reporting.

It can create a graphic with a factual or visual problem.

That means every stage where AI helps is also a stage where its output may need to be checked.

The system is assisting the workflow.

It is not exempt from the workflow.

Stage Six: Voice and Editing

Only after the argument has survived verification and challenge do I focus heavily on the writing itself.

Now I ask:

Does this sound like me?

Is the structure clear?

Is the article repetitive?

Is the language stronger than necessary?

Are the transitions doing their job?

Did AI introduce generic language?

Did the draft flatten uncertainty?

Did it make the article sound too polished, too formal or too predictable?

This stage matters because AI can make almost everything sound competent.

Competent is not the same as distinctive.

Journalism should still have a voice.

The writing should reflect the journalist’s judgment, rhythm and priorities.

AI can help improve the prose.

It should not erase the person behind it.

Stage Seven: Editorial Quality Control

The quality-control pass asks a different question:

“Is this ready to publish?”

At this point I am no longer trying to reinvent the argument.

I am inspecting the complete product.

Headline.

Names.

Dates.

Statistics.

Attribution.

Links.

Captions.

Graphics.

Disclosure language.

Formatting.

Repeated ideas.

Anything that could have slipped through the earlier stages.

This is where small mistakes get caught.

And small mistakes matter.

A deeply reported story can still lose credibility because a name is misspelled.

A headline can undermine careful reporting by overstating the conclusion.

A production element can introduce an error that never existed in the manuscript.

A graphic, caption, tag or disclosure can create its own problem.

Production is not separate from editorial accuracy simply because it happens near the end.

Stage Eight: Lock the Manuscript

At some point, the editing has to stop.

That is why I use the word:

“locked.”

A locked manuscript means the substantive editorial process is finished.

The article has been researched.

Verified.

Challenged.

Edited.

Reviewed.

It is ready for production.

That does not mean the article can never change.

If new evidence appears, I can correct it.

If I discover an error, I should fix it.

Locked does not mean untouchable.

It means the article has completed the workflow required for publication.

That distinction helps me avoid endless editing.

There is always another sentence I could adjust.

Another word I could change.

Another version I could consider.

At some point, responsible work has to become published work.

Stage Nine: Production

The manuscript may be locked, but publication still requires production.

The article has to be placed on the website.

The headline has to fit.

The graphic has to work.

Tags have to be added.

The disclosure has to appear.

Links have to function.

The page has to look right.

This is where the workflow expands beyond writing.

Publishing is not just about the manuscript.

It is about the complete package the reader sees.

AI can help with graphics, metadata, captions and formatting.

But those elements still need review.

A strong graphic with a factual, contextual or visual error is not finished.

A strong article with broken formatting is not finished.

The workflow does not end when the final sentence is written.

It ends when the published product has been checked.

The Workflow Is Not a Certification System

There is another danger once the process becomes familiar.

I can start believing that because I completed every stage, the article must be correct.

That would be a mistake.

A checklist is a safeguard.

It is not proof.

I can perform a weak truth pass.

I can miss something during verification.

The red team can overlook the real weakness.

The final QC can fail.

The purpose of the workflow is not to make mistakes impossible.

It is to create repeated opportunities to catch them.

That is a meaningful difference.

Familiarity Can Become Complacency

A repeatable workflow should reduce unnecessary cognitive load.

It should not reduce skepticism.

Once the process becomes familiar, I can move through it much faster.

That is useful.

But familiarity creates its own danger.

I can begin clicking through the stages mentally.

Truth pass?

Done.

Red team?

Done.

QC?

Done.

That turns editorial discipline into ritual.

The stage matters only if I am actually doing the thinking the stage requires.

Efficiency should make the process easier to execute.

It should not make the process automatic in my head.

Not Every Story Needs the Same Amount of Process

The workflow also needs proportion.

A short game recap does not require the same depth of review as a major investigative article.

A routine announcement does not carry the same risk as an accusation.

A feature based on public records may require different verification from breaking news.

The stages remain useful.

The intensity changes.

Responsible AI use should not mean turning every story into an exhausting bureaucracy.

The goal is not maximum process.

The goal is appropriate process.

Higher risk deserves more scrutiny.

Lower risk can move faster.

And some high-risk journalism may require safeguards beyond my normal workflow.

That could include additional independent reporting, another editor, legal review or subject-matter expertise.

A personal workflow should never become an excuse for believing one person and one AI system can adequately handle every kind of story.

Speed Is the Result, Not the Standard

One of the most interesting things I have learned from using this workflow is that speed becomes a byproduct.

Once the stages are familiar, I can move through them quickly.

Research is organized faster.

The draft develops faster.

Weak claims are easier to find.

The red-team process is repeatable.

The quality-control checklist is familiar.

That can dramatically reduce production time.

But speed is not the standard.

A ten-minute workflow is useful only if the article is actually ready.

If the story needs another hour of verification, then it needs another hour.

If a source has not responded, the workflow cannot manufacture the answer.

If the evidence is incomplete, AI cannot make it complete.

The purpose of the workflow is not to publish faster at all costs.

It is to reduce unnecessary friction while protecting the steps that matter.

The Human Is Still Directing the Process

This chapter brings the earlier lessons together.

The journalist remains responsible.

AI output is not evidence.

Human oversight has to mean something.

The story should be challenged before publication.

Now those ideas become a workflow.

AI may help at nearly every stage.

But it does not own the process.

I decide what question to pursue.

I decide what evidence matters.

I decide whether the claim is supported.

I decide which criticism requires action.

I decide whether the writing sounds right.

I decide when the article is ready.

And ultimately, I decide whether to publish.

That is what directing the workflow means.

From Draft to Publish-Ready

The most important change AI made in my work was not simply that it gave me faster drafts.

It forced me to think more deliberately about everything that happens around the draft.

Research.

Verification.

Challenge.

Revision.

Judgment.

Quality control.

Publication.

Those stages existed before AI.

But AI made the need to define them more obvious.

Because when a machine can produce polished language in seconds, the journalist needs a process that separates polished language from publishable journalism.

The workflow creates the places where human judgment must be exercised.

But the workflow itself does not provide that judgment.

The quality of the decisions made inside it still determines the result.

That is what turns a draft into something ready to publish.


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

Make AI Challenge the Story

By Cheval John | Vallano Media
This article was created with the assistance of artificial intelligence.
The final version was reviewed, edited and fact-checked by the author


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

One of the least useful things artificial intelligence can do for a journalist is agree with everything.

That may sound strange.

Most people want tools that help them.

They want answers.

They want suggestions.

They want reassurance that the work is good.

But journalism does not improve simply because someone — or something — says:

“This looks great.”

Sometimes the most valuable response is:

“This argument does not hold up.”

That realization changed the way I use AI.

Instead of asking only for help producing the work, I began asking AI to challenge the work.

Attack the argument.

Find the weakness.

Question the assumption.

Show me what I may have overlooked.

That became one of the most useful parts of my AI-assisted editorial process.

I call it red-teaming.

In this book, I am using that term in an editorial sense: deliberately stress-testing the claims, reasoning, sourcing and framing of a story before publication.

Agreement Is Easy

AI systems are designed to be helpful.

That can become a problem.

If I give an AI system an argument and ask:

“Does this make sense?”

I may receive a supportive response explaining why the argument works.

If I ask:

“Is this a strong article?”

I may hear that the article is thoughtful, compelling or well structured.

That may feel good.

It may also tell me very little.

The problem is not that positive feedback is always wrong.

The problem is that agreement does not test the work.

Journalism needs resistance.

A source may challenge the premise.

An editor may question the evidence.

A reader may notice an assumption the writer missed.

If I want the article to survive those challenges after publication, it makes sense to expose it to some resistance before publication.

AI can sometimes help simulate that kind of critique.

It is not the same as an experienced editor or a knowledgeable subject-matter expert.

But it can provide another angle from which to examine the work.

Ask the Opposite Question

One of the simplest changes I made was changing the way I ask AI for feedback.

Instead of:

“Is this argument strong?”

I can ask:

“What is the strongest evidence-based challenge to this argument?”

Instead of:

“Does this article make sense?”

I can ask:

“Where could a skeptical reader say this article fails?”

Instead of:

“Is this ready to publish?”

I can ask:

“What weakness could justify holding this story?”

The difference is significant.

The first set of questions invites approval.

The second invites resistance.

That resistance can expose problems I might not notice because I am already too close to the work.

Writers Become Attached to Their Arguments

This is not an AI problem.

It is a human problem.

Once I have spent time researching and writing something, I naturally begin to understand the story through the structure I created.

I know what I meant.

I know why I included a particular detail.

I know how the argument developed.

That familiarity can create blind spots.

A paragraph may seem clear to me because I already know the background.

A conclusion may feel obvious because I have been thinking about it for hours.

A source may seem convincing because it supports the direction the story has already taken.

That is exactly when another perspective becomes useful.

Editors have always served that role.

Good editors ask the question the writer may not want to hear.

Why?

How do you know?

Where is the evidence?

Is that really what the source said?

Are you assuming too much?

What might an informed critic object to?

AI can sometimes help simulate that resistance.

Not replace the editor.

Not replace the journalist.

But add another layer of challenge before publication.

What Red-Teaming Means in My Workflow

When I red-team an article, I am not simply asking AI to proofread it.

Proofreading looks for problems in grammar, spelling and presentation.

Red-teaming asks whether the underlying journalism survives criticism.

The questions are different.

What is the weakest claim?

Where is the evidence thin?

Which sentence sounds more certain than the reporting supports?

What assumption am I making?

Is there another plausible interpretation?

What evidence-based weakness might a knowledgeable reader identify?

Have I confused correlation with causation?

Am I treating an opinion as a fact?

Does the headline promise more than the article proves?

Is there missing context that would materially change the reader’s understanding?

Those questions force the work into a more uncomfortable place.

That is the point.

Do Not Manufacture an Opposing Side

There is a danger in telling AI simply:

“Argue the other side.”

The system will usually try.

But the existence of a counterargument does not make the counterargument credible.

It may create a false equivalence.

It may elevate a fringe objection.

It may invent controversy where the evidence is actually quite strong.

That is why I prefer a narrower instruction.

Do not merely give me an opposing opinion.

Show me “evidence-based weaknesses, missing context, unsupported inferences and plausible alternative interpretations.”

That is a much better editorial test.

The question is not:

“Can somebody disagree with this?”

Someone almost always can.

The better question is:

“Does the disagreement expose a legitimate weakness in the evidence or reasoning?”

AI Can Be Wrong About the Weakness Too

There is another important warning.

Red-teaming does not mean accepting every criticism AI produces.

AI can identify a problem that is not actually a problem.

It can misunderstand the argument.

It can invent a counterargument that has little connection to the evidence.

It can overcorrect.

It can treat a well-supported conclusion as questionable simply because I instructed it to find something wrong.

That means the red-team response has to be evaluated too.

This connects directly back to the earlier chapters.

The human remains responsible.

AI output is not evidence.

The human has to exercise judgment.

Red-teaming gives me another perspective.

It does not give AI veto power over the article.

The Goal Is Not to Destroy the Story

A red-team process can go too far.

If I challenge every sentence endlessly, nothing will ever be published.

Every story has limitations.

Every argument can be questioned.

Every source can be criticized.

The goal is not to produce journalism that no human being could possibly challenge.

That is impossible.

The goal is to discover weaknesses that matter.

A missing comma is not the same as a missing source.

A stylistic disagreement is not the same as an unsupported accusation.

A theoretical alternative explanation is not automatically as credible as the explanation supported by the available evidence.

Red-teaming is useful only if the journalist can distinguish criticism generated for the sake of criticism from a material weakness.

Focus on What Could Change the Story

AI can produce long lists.

That does not mean every item deserves equal attention.

If a red-team review gives me ten objections, I do not automatically rewrite the story ten times.

I want to know:

Which objection could change the factual conclusion?

Which one exposes weak sourcing?

Which one reveals missing context or distorted framing?

Which one could damage trust if I ignore it?

That is a more useful filter.

One serious weakness is worth more than ten superficial criticisms.

AI can generate objections.

The journalist has to rank them.

Surviving a Red-Team Pass Does Not Prove the Story Is Correct

This distinction matters.

If a story survives an AI red-team review, that does not certify the journalism.

The AI may have missed something.

It may share the same assumptions that influenced the original work.

Its critique may simply be weak.

Passing one adversarial review means only that **one additional layer of challenge did not reveal a material problem**.

That is useful.

It is not proof.

Verification still matters.

Human editing still matters.

The underlying reporting still matters.

Red-teaming is another safeguard.

It is not a replacement for the others.

The Same AI Is Not an Independent Editor

There is another limitation worth acknowledging.

Sometimes the same AI system that helped with the draft also performs the red-team review.

That can still be useful.

But it is not independent review.

The model may carry forward assumptions from the earlier discussion.

It may understand the article through the framing already established in the conversation.

It may reproduce the same blind spots.

The higher the stakes, the less comfortable I should be relying on one system to create the work and then declare that it survived criticism.

Outside human review may be necessary.

Additional sourcing may be necessary.

A different evaluation approach may be necessary.

The AI red team is one perspective.

It should not pretend to be every perspective.

Red-Teaming Can Improve Reporting, Not Just Writing

One of the most useful things about the process is that it can send me back to the reporting.

Maybe the problem cannot be fixed by rewriting a sentence.

Maybe I need another source.

Maybe I need the original document.

Maybe I need to check the timeline again.

Maybe I need to separate what is confirmed from what I believe happened.

That is when red-teaming becomes more than an editing exercise.

It becomes part of the reporting process.

A good challenge can reveal that the story is not finished yet.

And that is valuable information to receive before publication.

Challenge the Headline Too

Headlines deserve their own red-team review.

A headline can be technically related to the article and still overstate what the reporting establishes.

That is especially tempting in digital publishing.

Strong language attracts attention.

Certainty attracts clicks.

Conflict attracts attention.

But a headline is still part of the journalism.

If the headline is stronger than the evidence, the article begins with a credibility problem.

So one of the questions worth asking is:

“Does the headline make a claim the article cannot fully support?”

That simple test can prevent a lot of unnecessary trouble.

Red-Team the AI Too

There is another layer to this process.

The AI itself should be challenged.

If AI summarizes a source, I can ask:

What did you omit?

What uncertainty did you flatten?

Show me where the source supports that statement.

What alternative interpretation exists?

If AI suggests a conclusion, I can ask:

“What evidence would make this conclusion wrong?”

If AI gives me a confident answer, I can ask:

“What part of this answer requires the most verification?”

The goal is not to create an endless conversation with the machine.

The goal is to prevent confidence from becoming authority.

AI should not merely challenge my work.

I should challenge its work too.

A Strong Review Also Identifies What Survives

Red-teaming should not become an exercise in finding fault for its own sake.

A useful review should also identify which claims remain strong after scrutiny.

What is well supported?

Which conclusion survived the challenge?

Which source still holds up?

Which section does not need to be rewritten?

That matters because the purpose of red-teaming is diagnosis.

Not destruction.

A good review separates the parts of the story that need more work from the parts that have already earned their place.

The Psychological Value of an Adversarial Partner

There is another benefit to red-teaming that is harder to measure.

It makes criticism normal.

Instead of treating criticism as something that arrives after publication, I can make it part of the production process.

That changes my relationship with the draft.

The goal becomes less about defending what I wrote and more about improving it.

If the red team finds something important, good.

Better now than after publication.

If the argument survives, good.

I have more confidence in the process because I tried to break it.

Not because the red-team pass proved the story correct.

It did not.

But because the work has faced another layer of scrutiny.

A Workflow for Challenging the Work

A journalist does not need an elaborate system to begin.

A simple editorial red-team workflow can look like this:

Finish the draft.

Identify the central claim.

Ask for the strongest evidence-based challenges.

Look for weak or overstated claims, missing context and unsupported inferences.

Identify which weaknesses could materially change the story.

Then evaluate every criticism yourself.

Some will be useful.

Some will be irrelevant.

Some will be wrong.

The human decides which ones require action.

That may lead to a rewrite.

It may lead to more reporting.

It may lead to stronger sourcing.

Or it may confirm that the original argument already has strong support.

All of those outcomes can be useful.

Challenge Before the Audience Does

Every published story eventually meets readers who did not participate in creating it.

They bring different experiences.

Different knowledge.

Different assumptions.

Different questions.

Some will notice things the writer missed.

That cannot be eliminated.

But journalism can become stronger when the writer deliberately invites challenge before publication.

AI gives me another way to do that.

Not because AI knows better than the journalist.

Not because its criticism is automatically correct.

And not because surviving an AI challenge proves the journalism is right.

The value comes from resistance.

Do not use AI only to help build the argument.

Use it to pressure-test the argument too.

Then decide what survives.

Because a story that has been seriously challenged before publication has a better chance of surviving the challenges that come after it.


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

The Human in the Loop

By Cheval John | Vallano Media
This article was created with the assistance of artificial intelligence.
The final version was reviewed, edited and fact-checked by the author


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

“Keep a human in the loop.”

It sounds simple.

It also sounds reassuring.

Put a person somewhere in the process, and the risks of artificial intelligence are supposedly under control.

But that phrase can hide a much harder question:

“What is the human actually doing?”

If the person is simply approving whatever the AI produces, the human is technically in the loop.

That does not mean the human is exercising judgment.

If the editor is rushed, the reporter is overconfident, or the organization has built a workflow that encourages people to trust the technology without questioning it, human oversight can become little more than a formality.

That is not enough.

A meaningful human role requires more than presence.

It requires authority.

It requires understanding.

It requires skepticism.

And sometimes, it requires saying no.

Being Present Is Not the Same as Being in Control

AI-assisted workflows often include a human checkpoint.

Draft.

Review.

Approve.

Publish.

On paper, that looks responsible.

But a checkpoint only matters if the person at that checkpoint understands what they are reviewing and has the authority to stop the process.

If an AI-generated summary is wrong and the journalist never checks the source, the human review did not accomplish much.

If an editor sees a polished draft and assumes someone else already verified it, the human review did not accomplish much.

If a newsroom requires approval but creates deadlines so tight that nobody has time to challenge the output, the system may have a human in the loop without meaningful human control.

That distinction matters.

The goal is not simply to insert a person between AI and publication.

The goal is to preserve human judgment at the points where judgment matters most.

And even then, human oversight is not a guarantee of correctness.

It is a mechanism for accountability, challenge and intervention.

What the Human Still Has to Decide

Artificial intelligence can assist with many parts of journalism.

It can help organize research.

It can summarize documents.

It can compare information.

It can suggest questions.

It can challenge an argument.

It can help edit a draft.

But journalism still contains decisions that are difficult to reduce to a checklist.

Is this source credible?

Is this statistic meaningful without additional context?

Is the headline accurate or merely dramatic?

Is there enough evidence to publish?

Would publication cause harm that outweighs its value?

Those decisions require editorial judgment.

AI may contribute information to the decision.

But for consequential editorial decisions, accountable human judgment still has to remain decisive.

The Human Must Understand the Task

Human oversight becomes weak when the person supervising the AI does not understand the work being delegated.

That does not mean a journalist has to understand every technical detail of the AI system.

I do not need to know how every line of code inside a model works before I can use it responsibly.

But I do need to understand the journalistic task.

I need to understand what a good result looks like.

I need to understand the limitations of the tool.

And I need to understand the standard the output has to meet.

If I ask AI to help fact-check a story but I do not understand verification, I may accept a weak check as a strong one.

If I ask AI to help write a headline but I do not understand how headlines can distort a story, I may accept something misleading because it sounds good.

If I use AI to summarize a report but never read the underlying material, I may not notice what the summary left out.

That leads to a principle I keep returning to:

“You cannot responsibly supervise work you do not understand well enough to judge.”

Delegation does not require knowing everything.

It does require knowing enough to recognize when the result is not good enough.

The Danger of the Rubber Stamp

One way human oversight can fail is through overreliance on automated output.

The system produces an answer.

The answer looks polished.

The human assumes the machine probably got it right.

Instead of reviewing the work, the person confirms it.

That is a very different behavior.

A rubber stamp may technically satisfy a policy requiring human approval.

It does not satisfy the deeper purpose of human oversight.

The human has to be willing to interrupt the process.

That may mean asking:

Where did this come from?

What did the AI leave out?

Does the source actually support this?

What assumption is hidden here?

Why is this being presented as certain?

Would I publish this if AI had not suggested it?

Those questions turn oversight into active supervision.

And the phrase “human reviewed” can be just as misleading as “AI-generated” if nobody explains what the review actually involved.

Not Every Task Needs the Same Oversight

There is another important distinction.

Human oversight should be proportional to risk.

A low-risk formatting task does not require the same level of scrutiny as a factual claim.

Correcting capitalization is not the same as evaluating an accusation.

Reformatting notes is not the same as verifying breaking news.

Suggesting alternative headlines is not the same as deciding whether an investigative conclusion is supported.

If every AI-assisted task requires the same level of manual review, the workflow can become so cumbersome that the technology provides little benefit.

The goal is not maximum friction.

The goal is appropriate friction.

The higher the stakes, the stronger the oversight should be.

That requires the journalist or organization to understand where the real risks are.

Sometimes the Human Has to Slow the System Down

Artificial intelligence is often sold through speed.

Faster research.

Faster drafting.

Faster editing.

Faster production.

That can be useful.

But the human role sometimes exists specifically to slow the process down.

That may sound inefficient.

It is not.

A journalist may need to stop and read the original document.

An editor may need another source.

A claim may need to be softened.

A headline may need to be rewritten.

A story may need to wait.

The value of the human is not always acceleration.

Sometimes the value is friction.

Responsible friction prevents speed from turning into carelessness.

In an AI-assisted workflow, the pause can be as important as the prompt.

Judgment Is Not the Same as Being Right

There is another distinction worth making.

Human judgment does not mean human certainty.

A journalist can be biased.

An editor can make a poor decision.

A newsroom can fail.

Humans are not automatically more accurate simply because they are human.

The purpose of human oversight is not to replace machine error with human perfection.

It is to introduce accountability, context, professional standards and the ability to challenge the system.

Good judgment has to be disciplined.

It should be informed by evidence.

It should be open to correction.

It should recognize uncertainty.

It should be willing to change when better information appears.

The human role is valuable because accountable people and institutions have to remain responsible for consequential decisions.

The Organization Is Part of the Loop Too

It is easy to picture the “human in the loop” as one reporter sitting at a computer.

That is too narrow.

News organizations are part of the loop.

Editors are part of the loop.

Managers are part of the loop.

Policies are part of the loop.

Training is part of the loop.

The design of the workflow itself is part of the loop.

Not every AI failure is a failure of the individual journalist.

Sometimes the workflow itself is badly designed.

A reporter may follow the process exactly as instructed and still be set up to fail by weak training, unclear standards, poor tools or unrealistic deadlines.

If an organization introduces an AI tool without explaining its limits, that matters.

If employees are encouraged to use AI but are not given clear verification standards, that matters.

If the newsroom rewards speed while expecting careful oversight, that contradiction matters too.

A responsible AI workflow cannot depend entirely on an individual journalist catching every possible problem.

The organization has to create conditions that make good judgment possible.

Authority Has to Be Real

Formal authority on paper is not enough.

An organization can tell employees that they have the right to reject AI-generated work.

But what happens when somebody actually does it?

What happens when a reporter says:

This needs another source.

This output is unreliable.

This story is not ready.

This workflow is creating a problem.

If slowing down the process is punished, discouraged or treated as incompetence, then the journalist does not really have meaningful authority.

Real human control depends not only on approval rights.

It depends on whether the culture supports people who challenge, delay or reject AI-assisted work when necessary.

That may be one of the least visible parts of responsible AI use.

But it matters.

My Role in the Loop

This has become increasingly clear in my own work.

When I use AI, I am not simply waiting at the end of the process to approve whatever appears.

I am involved throughout it.

I decide what question to ask.

I decide what sources matter.

I decide when the AI has misunderstood the task.

I decide when an argument needs to be challenged.

I decide when something needs independent verification.

I decide whether the final language sounds like me.

I decide whether the work is ready.

That is what human oversight looks like in practice.

It is not one final approval button.

It is a series of decisions throughout the workflow.

Sometimes those decisions are small.

Sometimes they change the entire direction of the work.

The Human Should Not Be the Excuse

There is also a danger in using “human oversight” as a blanket defense.

An organization cannot simply say:

“There was a human in the loop.”

And assume that settles the question.

Was the human trained?

Did they understand the task and the limitations of the tool?

Did they have time to review the work?

Did they have access to the underlying evidence?

Could they reject the AI output?

If the answer to those questions is no, the presence of a human may not mean very much.

Human oversight should be something we can describe.

Not just something we claim exists.

A Better Question

Instead of asking:

“Is there a human in the loop?”

I think the better question is:

“Where does human judgment enter the process, and what authority does it have?”

That question forces us to look at the workflow more carefully.

Who verifies?

Who approves?

Who can stop publication?

Who decides whether a source is credible?

Who resolves conflicting information?

Who is responsible when something goes wrong?

Those questions matter more than whether a policy document contains the phrase “human in the loop.”

Human Control Does Not Mean Rejecting AI

None of this means journalists should avoid AI.

The lesson is not that human control requires manual work.

Chapter 2 established that responsibility does not require doing everything yourself.

Human control means knowing where AI can help and where accountable human judgment has to remain decisive.

AI may help surface possibilities.

The human decides which matter.

AI may help draft language.

The human decides what is accurate.

AI may help challenge an argument.

The human decides which criticism is valid.

AI may help identify a pattern.

The human decides whether the pattern is meaningful.

That is partnership without surrender.

The Final Decision Still Belongs Somewhere

Every workflow eventually reaches a point where something happens.

A story is published.

A claim is removed.

A headline is changed.

A source is rejected.

A correction is issued.

A story is held.

Artificial intelligence can influence those decisions.

It should not be allowed to make consequential editorial decisions without accountable human review.

Somebody — or some clearly defined editorial authority — has to own the final choice.

That ownership is what gives the phrase “human in the loop” meaning.

Without it, the human is simply present.

With it, the human has a responsibility to understand, challenge and intervene.

The goal is not merely to keep humans near artificial intelligence.

The goal is to keep human judgment, authority and accountability inside the process.

That is what the loop is supposed to protect.


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

AI Output Is Not Evidence

By Cheval John | Vallano Media
This article was created with the assistance of artificial intelligence.
The final version was reviewed, edited and fact-checked by the author


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.

One of the easiest mistakes to make with artificial intelligence is also one of the most dangerous.

The answer looks finished.

The language sounds confident.

The explanation makes sense.

So you believe it.

That is exactly when a journalist needs to slow down.

AI can be useful for finding patterns, organizing information, generating questions, comparing claims and helping a journalist think through a story.

But none of those things make the AI-generated answer proof.

That distinction has become one of the most important rules in my own work:

AI output is not evidence

If an AI system gives me a fact, I still need to verify it.

If it gives me a quote, I need to find the original source.

If it gives me a statistic, I need to trace it back to the organization or dataset that produced it.

If it tells me that something happened, I need evidence that it actually happened.

The AI response may help me know where to look.

It may even point me directly toward legitimate evidence.

But the answer itself does not replace the proof.

Confidence Is Not Verification

Generative AI has a particular ability that makes this problem more difficult.

It can sound certain.

A wrong answer may not look wrong.

It may arrive in clean sentences with a logical explanation and enough detail to make the reader think the system must know what it is talking about.

That creates a dangerous shortcut.

The journalist stops asking:

“Is this true?”

And starts asking:

“Does this sound true?”

Those are not the same question.

Journalism cannot operate on plausibility alone.

A statement can be reasonable and still be wrong.

A date can look right and still be wrong.

A statistic can sound familiar and still be invented, outdated or taken out of context.

A quote can sound exactly like something a person would say and still never have been said.

That is why confidence cannot be treated as verification.

The Evidence Has to Exist Beyond the AI Answer

When AI gives me useful information, one of my next questions is usually:

Where is the evidence?

That may mean an official report, a government database, a court document, a transcript, a recording, a reputable news report or a historical archive.

The form will depend on the story.

What matters is that I can inspect the underlying evidence independently rather than treating the AI-generated answer itself as proof.

That distinction becomes even more important when the claim is central to the story.

The more consequential the claim, the more carefully I want to examine what supports it.

An AI system can point toward evidence.

It can summarize evidence.

It can help compare evidence.

But the evidence still has to withstand scrutiny on its own.

A Citation Is an Invitation to Verify

AI systems create another temptation.

They can provide citations.

A citation may look complete.

Author.

Title.

Publication.

Date.

Maybe even a link.

That still does not mean the work is done.

The source may not exist.

It may exist but not say what the AI claims it says.

The citation may point to something outdated.

Important context may have been omitted.

Details from several sources may have been mixed together.

That means I cannot verify a claim simply by seeing a citation attached to it.

I have to open the source.

Read it.

Confirm the relevant material.

Check the date.

Understand the context.

And decide whether the evidence actually supports the statement I want to make.

A citation is not verification.

It is an invitation to verify.

Verification Changes the Question

Early in my AI use, I was much more likely to ask:

“What is the answer?”

Now I am more likely to ask:

“What evidence supports this answer?”

That small change has made a significant difference in how I work.

The first question invites the system to produce information.

The second forces me to think about the information differently.

What would prove this?

Who would know?

Where would the original record exist?

Is there a primary source?

Does another independent source confirm it?

Is the information current?

Am I looking at the original claim or somebody repeating it?

Those questions move the process away from information generation and toward verification.

And that is where journalism has to live.

Primary Does Not Mean Infallible

When a primary source exists, I usually want to see it.

If a government agency publishes the data, I want the agency’s report.

If a university announces something, I want the university’s documentation.

If a sports organization changes a rule, I want the official rulebook or announcement.

If a company reports financial results, I want the filing or official release.

But there is an important distinction.

Primary does not mean infallible.

A company can emphasize information that benefits the company.

A government agency can provide an incomplete picture.

A university can frame an announcement in its own interest.

A court filing represents a party’s legal position, not necessarily an established fact.

Getting closer to the original record does not eliminate the need for judgment.

The source still has to be evaluated.

Secondary reporting matters too.

Good journalism can provide context, expertise, independent confirmation and scrutiny that the original source does not provide.

So my preference for primary evidence is not a rule that says primary sources are always better or always correct.

It means I want to understand where the claim originated before deciding what it proves.

Two Sources Do Not Automatically Mean Two Confirmations

Another verification trap appears when the same claim shows up in multiple places.

It is tempting to think:

Two sources say it, so it must be confirmed.

Not necessarily.

Both sources may be repeating the same original report.

One may have copied the other.

Both may rely on the same press release.

Both may be citing the same anonymous source.

AI can make this especially tricky because it may summarize several webpages that appear to agree when all of them ultimately trace back to one original claim.

So verification is not simply about counting links.

It is about understanding provenance.

Where did the information originate?

Are the confirmations actually independent?

What evidence does each source possess?

Even two genuinely independent sources can be wrong.

Independence increases confidence.

It does not eliminate the need to evaluate the quality of the evidence itself.

Verification Is More Than Fact-Checking

Verification is often described as fact-checking.

That is part of it.

But verification is broader.

A sentence can contain technically correct facts and still create a misleading impression.

A statistic may be accurate but missing context.

A quote may be authentic but selectively presented.

An AI-generated summary may include the facts while flattening uncertainty or leaving out an important qualification.

So verification also means asking:

What is missing?

What would change the interpretation?

What assumptions are hidden inside this claim?

Does the evidence support the conclusion being drawn from it?

That is where verification overlaps with editorial judgment.

Sometimes the Answer Is: We Do Not Know Yet

Verification does not always produce certainty.

Sometimes the records are incomplete.

Sometimes sources disagree.

Sometimes a developing event is changing faster than the available evidence can keep up.

Sometimes the most accurate conclusion is:

We do not know yet.

That is not a failure of journalism.

Pretending to know would be the failure.

This is especially important during breaking news.

Waiting until every uncertainty disappears is not always possible or desirable.

The public may have a legitimate need for information before the full picture is known.

The responsibility then is different.

Publish what the evidence supports.

Clearly distinguish confirmed information from what remains uncertain.

Attribute claims.

Avoid presenting speculation as fact.

Update the reporting when stronger evidence becomes available.

Verification is not always a switch between “publish” and “do not publish.”

Sometimes it determines what can responsibly be published right now.

AI Can Help With Verification — Carefully

There is an irony in all of this.

The same technology that creates new verification problems can also help with verification.

AI can help surface claims that may need checking.

It can compare documents, identify apparent inconsistencies, generate questions and challenge an assumption.

Those uses can be valuable.

But there are two boundaries I try to maintain.

First, AI does not decide which claims deserve the deepest scrutiny.

That remains an editorial judgment.

A journalist has to decide which claims are central, disputed, risky or consequential enough to require more work.

Second, AI cannot certify its own answer.

If an AI system produces a claim, asking that same system whether the claim is correct is not independent verification.

AI can assist the verification process.

It cannot become the final authority within that process.

The Verification Record Matters

One of the principles I keep returning to is that the evidence trail matters more than the appearance of certainty.

Someone can speculate that an article was produced with AI.

Someone can run the article through an AI detector.

Even if someone could perfectly determine whether AI was involved, that still would not establish whether the journalism itself was accurate.

Those are different questions.

Accuracy depends on evidence.

What sources were used?

What claims were checked?

What evidence supports the central statements?

What uncertainty remained?

What was changed or removed during the editorial process?

Those questions tell me much more about the quality of the journalism than simply asking whether AI touched the work.

The important question is not:

“Was AI involved?”

The better question is:

“Can the journalism be verified?”

A Real Workflow Needs a Verification Stage

This is where verification connects directly to the workflow I described in Chapter 2.

A workflow is not complete because it produces a draft.

A responsible workflow needs a verification stage.

In my own work, that often means stopping after the draft and examining the factual claims.

Which statements came from firsthand reporting?

Which came from primary documents?

Which came from secondary reporting?

Which originated as AI suggestions?

Which require another source?

Which are interpretations rather than established facts?

Which statements sound stronger than the evidence actually allows?

That pause matters.

It separates production from publication.

AI can make it easier to generate material quickly.

Verification is where I decide whether that material deserves to become journalism.

When Verification Gives the Time Back

There is something else worth acknowledging.

Verification takes time.

Sometimes AI saves time on one part of the process and verification gives some or all of that time right back.

Sometimes verification takes longer than the drafting.

That is not necessarily a failure of the workflow.

It may be the workflow doing exactly what it is supposed to do.

If a claim takes 20 minutes to verify, those 20 minutes are part of producing trustworthy work.

This is why I do not measure AI productivity only by how quickly a draft appears.

The better question is whether AI helps me move through the entire process without lowering the standard.

If the draft arrives faster but verification becomes weaker, I have gained very little.

If AI helps me spend less time on mechanical work and more time examining the evidence, that is a much more meaningful improvement.

The Journalist Has to Be Willing to Let Go

Sometimes verification produces an answer nobody wants after spending time on a story:

The evidence is not strong enough.

Maybe a sentence has to go.

Maybe an entire section has to go.

Maybe the central claim has to be rewritten.

Maybe the story has to wait.

And sometimes, after all that work, the story should not be published at all.

That can be difficult.

The draft exists.

Time has already been invested.

The headline may be good.

A graphic may even be finished.

None of those things turn weak evidence into strong evidence.

At the same time, uncertainty does not always require abandoning the entire story.

Sometimes the responsible decision is to publish only the portion that can be supported and clearly explain what remains unknown.

The principle is the same either way:

The evidence determines what survives.

Not the amount of work already invested.

Not how convincing the draft sounds.

And not how badly I want the story to be true.

Trust Is Built Before Publication

Readers usually see the finished article.

They do not see the searches.

The discarded sources.

The rewritten paragraphs.

The claims that were removed.

The moments when the journalist stopped and said:

I cannot prove this.

Or:

We do not know this yet.

Those invisible decisions are part of what makes the final product trustworthy.

AI does not change that.

If anything, those decisions become more important because plausible information has become easier to produce.

The harder task is deciding what deserves to survive.

That is why verification cannot be treated as an optional final check.

It has to be built into the process.

Evidence Before Confidence

AI can be fast.

It can be useful.

It can be impressive.

It can also be wrong.

The journalist’s job is not to be impressed by the answer.

The journalist’s job is to determine what the evidence supports.

That means asking where the information came from.

Opening the source.

Checking the context.

Looking for independent confirmation when appropriate.

Understanding uncertainty.

And being willing to reject or qualify a claim when the evidence does not support certainty.

AI can help generate and challenge information.

But journalism begins when the information survives verification.

That is the line I keep returning to:

AI output is not evidence.

The AI answer may point toward the evidence.

The evidence still has to speak for itself.


Graphics created with AI-assisted design tools.
Visual concepts, editorial direction, composition, and final creative decisions were developed by Vallano Media.