DON'T PASTE CHATGPT TO WIN AN ARGUMENT

Here is a piece of unsolicited advice: don't cut and paste ChatGPT text to prove a point in an argument.

Here's why.

I am well trained in using generative AI, I hated it from the beginning but being a ghostwriter for hire, there are clients who demand that you should know efficient prompt engineering.

You learn it because you have to.

You get good at it because work depends on it.

Then you go online and see people using the same tool to "win" arguments by copy-pasting its answers like a holy scripture.

It makes you look uneducated, stupid, and unable to carry a serious conversation.

And it's not just trolls doing it.

Even good governance advocates, reform-minded professionals, and respected intellectuals are now pasting full ChatGPT responses — logos and all — and treating them like dogma.

There is research showing that even when AI fact-checks are mostly accurate, they do not automatically improve people's ability to discern truth; in some cases, they even reduce discernment and make people more confident in wrong beliefs.

When people with influence over-trust AI, they don't just misinform others.

They train everyone watching to lower their standards too.

THE MACHINE DOES NOT KNOW, IT JUST SOUNDS SURE

Most people still treat ChatGPT like a smarter Google.

It doesn't "look up" facts in real time.

It generates text word by word based on patterns in its training data — predicting what usually comes next in similar sentences, not checking truth against a live database.

An explainer on Physics Forums said the model has no internal understanding that the words it strings together refer to anything in the real world.

That's why it feels authoritative.

The sentences are smooth.

The tone is confident.

It sounds like someone who has read everything.

But fluency is not proof.

IT INVENTS THINGS AND KEEPS A STRAIGHT FACE

In AI jargon, this is called "hallucination."

In normal language, it's making things up and saying them with a straight face.

Researchers have documented how often these models fabricate sources, statistics, and even full legal cases, then present them as if they were pulled from real databases.

In one review, large language models repeatedly produced plausible-sounding but false statements, including fake citations and distorted summaries.

The Mata v. Avianca case is the classic example.

A lawyer used ChatGPT for legal research; the model produced case citations and quotes that did not exist in any database.

The court discovered this and sanctioned the lawyers, ordering them to pay fines and notify the client and affected judges.

In other cases, attorneys in the US and abroad have been fined thousands of dollars after filings packed with AI-fabricated case law reached the courts.

Kung abogado nga nadadale, what more your average Facebook marites quoting ChatGPT about geopolitics?

The model doesn't warn you when it's hallucinating.

It just keeps talking.

IT'S OUTDATED BUT ANSWERS LIKE IT'S LIVE

Every version of ChatGPT has a knowledge cutoff — a date after which it hasn't seen new information in training.

Ask about events after that point, and it's guessing based on older patterns.

Sometimes it even guesses wrong about its own cutoff date.

Users have caught it confidently claiming more recent cutoffs than it actually has.

It doesn't have an internal calendar; it has patterns about what an AI "should" say when asked.

The BBC and European Broadcasting Union recently ran a large study on AI assistants answering news-type questions.

They found that these tools misrepresented or mishandled news in roughly 45% of thousands of evaluated responses, across multiple languages and platforms.

Many answers relied on outdated or incomplete information presented as if it were current.

So imagine using that in a Philippine argument.

You ask about the ICC and the Philippines.

Or about the latest tensions in the West Philippine Sea.

Or about a recent Senate hearing, COA report, or corruption case.

If the model's knowledge stops before the latest development, it will still answer.

It just won't tell you it's talking about an old reality.

You walk into a thread with that answer, thinking you have "facts," while everyone else is looking at this week's news.

Ikaw ang mukhang hindi nagbasa.

IT FOLDS UNDER PRESSURE OR DOUBLES DOWN WRONG

There's another problem: how it reacts when challenged.

A study reported by ScienceDaily showed that when users push back against correct answers — sometimes with nonsense reasoning — ChatGPT often apologizes and changes to the wrong answer.

It wants to keep you happy more than it wants to hold the line on truth.

On the flip side, people have shared experiences where the model confidently repeats wrong answers, even after being shown evidence, especially when the prompt encourages it to sound more certain.

A Nature study comparing model confidence and accuracy even likened some behaviors to a Dunning-Kruger effect, where the system expresses high confidence in areas where it is actually less reliable.

So you have a system that can be confidently wrong, can abandon a correct answer if the user insists, and still talks like it knows what it's doing.

Kung tao yan, hindi mo yan gagawing fact-checker.

IT HAS POLITICAL BIAS IT DOESN'T EXPLAIN

Now bring this into politics — where you and I actually argue.

A study from the University of East Anglia found that ChatGPT's political answers showed a consistent left-wing bias: favoring Democrats in the US, Labour in the UK, and Lula's Workers' Party in Brazil.

Other work has found that later versions of similar models can lean rightward on certain issues and frames over time.

Brookings reviewed ChatGPT's answers and observed that it often leaned left on many political questions and could give different answers to the same question at different times, depending on how the prompt was framed.

So the bias isn't stable.

It changes with model version, tuning, and phrasing.

Picture a Filipino scenario.

Someone asks, "Does the ICC still have jurisdiction over Duterte after withdrawal?" then pastes the AI's answer as proof in a thread.

Someone asks, "Who really owns the West Philippine Sea?" and uses the AI's phrasing as neutral authority, even if it echoes one side's talking points.

Someone prompts, "Is Marcos Jr. doing a good job as president?" and treats the output like a global verdict instead of one model's pattern.

You don't see the guardrails and biases behind those answers.

You just see a wall of text that looks smart.

You end up importing an opaque, drifting political bias into a Filipino conversation and labeling it "objective."

Parang nagpapasok ka ng dayuhan sa usapan, tapos siya pa ang pinapanigan mo.

THE ERROR RATE WOULD GET ANY HUMAN FIRED

The BBC/EBU study didn't look at obscure questions.

It looked at news-type questions across multiple AI assistants, and the conclusion was stark: these tools misrepresent or mishandle news content in about 45% of responses, regardless of language or territory.

That figure isn't just about currency — it covers overall accuracy: outputs that blurred opinion and fact, omitted key context, or got basic details wrong.

Fact-checking experts and AI risk frameworks echo this.

They warn that LLMs often produce false or misleading but plausible information, and that over-reliance on these tools can damage information integrity.

They are useful for drafting or brainstorming, but not dependable as final arbiters of truth.

If a human fact-checker got things wrong 40 to 50 percent of the time, would you still trust them as your only source in a heated argument?

Yet people screenshot or copy-paste AI outputs as if they're more trustworthy than an actual article, report, or law.

HOW IT PLAYS OUT IN PHILIPPINE ARGUMENTS

You've probably seen versions of this.

In a thread about EDSA, someone asks AI, "Was Martial Law good for the Philippines?" and pastes a neat paragraph that flattens decades of scholarship into a few generic lines.

In a debate about Charter change, someone pastes AI text declaring that "studies show" parliamentary systems are always better for development — with no authors, no dates, no context.

In a West Philippine Sea discussion, someone uses AI to "prove" either that China's claim is historically valid or that it is universally rejected, depending on how they framed the question.

Now add this: people who should know better — lawyers, policy analysts, good governance advocates, even academics — are also starting to paste AI outputs as if they were carefully sourced briefs.

Recent research shows that AI fact-checks, even when accurate on paper, can backfire: they don't reliably improve people's ability to tell true from false headlines and can even make people more confident in errors when the AI is unsure or wrong.

Other work calls this the "generative illusion": the sense that because something is long, clean, and AI-written, it must be more authoritative than it really is.

Each time, the pattern is the same.

The person isn't really reading.

They're outsourcing the hard part — sifting through sources, weighing context, checking dates — to something that was never built to carry that responsibility.

They don't want to understand.

They want to win.

AI CAN HELP FACT-CHECK, BUT NEVER ALONE

There's an important nuance here.

Generative AI by itself is not a fact-checker.

It predicts likely text; it does not independently verify what it says against a ground-truth database.

But if you're using a more advanced setup — a strong model wired to search, documents, or a retrieval system — you can use AI as part of a fact-checking workflow.

Key word: part.

Researchers looking at fact-checking and AI keep repeating the same warning: these tools can speed up claim matching, surface relevant documents, and highlight contradictions, but they still hallucinate, still miss context, and still inherit bias from their training data.

Because of that, human oversight is not optional; it is the whole point.

So if you're serious about using AI for truth, the approach has to change.

You don't ask the model "What is true?" and paste the answer.

You ask, "What could be true?" then audit everything it gives you.

You treat the AI like a witness you don't fully trust.

You poke holes in its story.

You compare it to documents, reports, laws, and data you can see with your own eyes.

You fact-check it adversarially.

Generative AI can help you work faster.

It cannot do the thinking for you.

A SIMPLE RULE FOR USING AI IN ARGUMENTS

If you want something practical, here is a three-step habit.

Step 1: Use AI for a draft or outline. Let it give you a starting paragraph, a list of points, or a summary of what might be relevant.

Step 2: Fact-check key claims with at least two real sources. For every important claim — date, quote, statistic, legal point — open actual links: news reports, court decisions, government data, reputable organizations. Check if they match.

Step 3: Only then screenshot or quote anything. If you can't be bothered to do the second step, don't use AI output as your "receipt." At most, treat it as "this is what the bot said," not "this is proven fact."

If you care about your credibility, you cannot skip the second step.

WHY YOU SHOULD NEVER PASTE CHATGPT AS YOUR "RECEIPT"

When someone drops a ChatGPT answer into a thread like a mic-drop, they are telling you three things about themselves.

They don't understand how the tool works.

They didn't care enough to verify basic facts.

They're more interested in looking right than being right.

As a ghostwriter, you see this from both sides — clients who want you to "AI everything" and readers who now have to wade through AI-generated half-truths in every comment section.

The irony is simple.

The more people treat AI outputs as unquestionable proof, the less trustworthy they become.

So the next time you're tempted to paste a ChatGPT paragraph into a political argument, ask yourself one thing before you click send:

If this machine has already been caught inventing cases, misreading news, and bending under pressure, why would you gamble your own credibility on its unverified answer?

And if you won't fact-check a robot that lies with a straight face, why should anyone trust you in a serious conversation?