Answers

How do AI models describe Philippine business leaders?

Last reviewed

By Antoni Grgurovic

Mostly from the public record, which is often filled with errors. I tested this in August 2026 by asking three AI tools the same six questions about 40 Philippine business leaders. When an answer failed, the cause was the public record 73 percent of the time. The AI tool was repeating a page that was stale, lacking substance, or poorly sourced. The other 27 percent was caused by the tool itself inventing a claim.

What I measured

The study is called The Machine's Version. The tools used were ChatGPT, Google Gemini, and Perplexity. I ran ChatGPT twice, once with web search turned on and once without. The 40 leaders were grouped into four tiers of ten based on how much public information was available on them. The six questions were ones a counterparty asks before a deal: who the person is, what they're currently managing, what they're best known for, what things one should be aware of before a partnership, what controversies exist, and whether the person is credible for a listed-company board.

I captured 960 answers between August 12 and August 26, 2026. I scored 894 of them against a record I verified by hand for each leader.

Most errors start in the public record

I counted 117 failures across the panel. In 85 of them, the tool reported faithfully from a faulty public record. The most common failure was weak sourcing. For 27 of the 40 leaders, at least one tool based a key claim on an aggregator site, a forum, or a profile page.

The second pattern was stale titles. For 19 of the 40 leaders, a tool presented a former office as a current one. The old title was still live on a page the tool cited.

Being well known changes the error

I expected the most prominent leaders to get the most accurate answers. I was wrong about that. The searching tools scored between 217 and 224 out of 270 in every tier. What changed across the tiers was the kind of mistake.

  • The least prominent leaders got silence or invention. ChatGPT without search either declined to answer or made up details.
  • The most prominent leaders got confident, cited answers with old titles in them. Eight of the ten most prominent names had a stale title in at least one answer.

A long career leaves many superseded titles online. The tool picks one up and presents it as current.

Memory invents, and web search cites weak pages

ChatGPT without web search invented a claim in 24 percent of its answers. The three setups with search enabled had a combined invention rate of only 2.5 percent (inventions are also known as hallucinations). Web search cut inventions about tenfold.

Web search, however, brought its own problem. In 13.5 percent of searched answers, a key claim relied on a weak source. Gemini produced about three quarters of those. I tested Gemini on its free model without needing to sign up, which is the lightest one Google offers, so read that figure with care.

One finding also went against my prediction. Where ChatGPT without search knew almost nothing about a leader, it declined to answer and invented less. Where it knew at least something, it answered and invented more. So a half-built record ends up carrying more risk than an empty one.

The risk question gets the weakest answer

The tools invented most when I asked who someone is. They invented the least when I asked about controversies and risks. That sounds reassuring on the surface, but underneath lies a deeper problem.

On the controversy question, 13 percent of answers scored low on accuracy. Most of those were answers that declared a clean record when another tool had found adverse material on the same person. A counterparty reads a clean answer as finished diligence when the tool had not actually checked thoroughly.

For 22 of the 40 leaders, at least one tool surfaced adverse or regulatory material that names the person. The tools that search found it. ChatGPT without search did not. The default is to query ChatGPT with search turned on, but sometimes the model answers from its own memory, which is why I isolated that variable to test it directly.

The name alone is sometimes the wrong query

Some leaders share a name with someone else. Before the study began, I flagged six leaders who share a name with another known person and recorded each one as a planted test of whether the tools would mix them up. Only two of the six caused confusion. In one of those two, a bare name returned the other person in every tool. The same name with the company attached returned the right person, minutes later, in the same tool.

I report those six planted tests separately, so the main panel for this question is the other 34 leaders. Among those 34, confusion between two people was rare. It happened for 4 of them, and I had not predicted any of those four in advance.

What a leader can do about it

The 73 percent is the part you can repair. Each of those failures lives on a page. A stale title can be corrected wherever it appears. A weak source will lose its high citation ranking when a primary source exists that is complete and easy to read. This is slow work across many pages, some of which belong to other people or organizations.

The other 27 percent has no page to fix. When a tool invents a directorship, no source carries the claim, so there is nothing to correct at the source. You can still respond. Save a dated copy of the false answer, so you have proof of what the tool said and when. Make sure an accurate record, such as an official bio that lists your real board seats, is easy for the tools to find. Report the error to the AI lab that runs the chatbot. Then check again later, because an invented claim can return in a new answer. But nobody can remove it outright, and anyone who promises to is selling something they cannot deliver.

A simple first step costs nothing. Ask each tool the six questions about your own name. Then ask again with your company attached. Save the answers with the date and check every title and every source against your own verified ground truth. No one knows your bio better than you.

What this study cannot tell you

  • It is not a sample of all Philippine executives. I chose 40 varying leaders based on how prominent they are, not how successful their businesses are.
  • It is one two-week window. The vendors update their models without notice, and I make no claim about how the tools behave today. But my data on how the models behaved during a snapshot in time is evergreen.
  • I scored it myself. An AI assistant produced a first pass, and I ratified every score. There was no second human rater.
  • I have an interest. My firm repairs the defects this study counts. I registered my definitions and predictions before I counted any failure, and two of my five predictions failed. I published both failures.
  • No leader is named. The report names no subject, and subject-level findings are sealed.

I will run the same 40 leaders and the same six questions again in August 2027.

Common questions

Is ChatGPT accurate about Philippine executives?

With web search on, it was accurate on most answers in my panel, and it invented in 1 percent of them. With search off, it invented in 24 percent. In both modes it repeated stale titles from the pages it had read or learned from.

Which AI tool describes Philippine business leaders the best?

My study does not rank the tools. They work in different ways, and each one failed in its own pattern. A leader needs to check all of them, because a counterparty can use any of them.

Does being famous make the AI answer more accurate?

No. Scores were flat across all four tiers. The most prominent leaders got more stale titles and more weak citations.

Will AI find a lawsuit or a regulatory case about me?

A tool that searches the web is likely to find it if the record names you. A tool answering from memory is likely to miss it. Some answers will also declare a clean record without having checked.

Can I change what AI says about me?

You can change the record the tools read, and most failures in my study started there. You cannot edit a model directly. I cover the steps in a separate answer, How to Fix What AI Says About Me.

What is AI visibility for an executive?

It is how AI tools describe you when someone asks about you by name. It covers whether they identify you, whether the facts are current, which sources they cite, and whether anything damaging appears that should not be there.

For reputational inquiries, please submit a Private Reputation Review