Skip to main content
  1. Home
  2. Computing
  3. Features

Am I hallucinating or is ChatGPT still making mistakes – even with GPT-5?

Or... 'How I nearly made the mistake of trusting a chatbot'

Add as a preferred source on Google
ChatGPT on a smartphone.
Shantanu Kumar/Pexels

During the launch of OpenAI’s GPT-5, one of the most important claims was that ChatGPT would now be more reliable and would have fewer moments where it presents guesses as fact.

My experience this week says that the company still has a lot of work to do.

Recommended Videos

It started over dinner at a cooking school in Sicily, where I met an impressive Australian working in restaurant PR. I wanted to know more about their agency, so out of curiosity I did a quick search.

I tried LinkedIn first— after all, a hustling PR person would certainly have a profile, no? But nada. I thought: “That’s strange, but I suppose not everyone is on LinkedIn.”  I then tried Google – where of course, they would surely be listed. Again zip, zero, zilch.

Being in the land of White Lotus I began to consider whether this was an elaborate hoax, someone just enjoying briefly living a false life in sunny climes. 

But there were enough other facts that emerged over dinner to convince me that they were legitimate, such as common acquaintances with other guests and specific details that could be easily confirmed. Plus, they were traveling with their mom, and it’s hard to run a scam in front of your mother – even in Sicily. 

I then asked Chat GPT the same kind of query I tried in Google: “profile xx person. Born in Australia. Works in London in public relations.” Quickly, ChatGPT came back with the goods: a detailed profile with places of employment, start and end dates, and education. It even commented on their working style.

This was brilliant – proof that AI was a true step forward, freeing us from our blinkered stumbles through Google and social platforms. Evidence that we no longer needed to have a degree in search-ology to find the information we needed – machines could do the job of interrogating machines for us…

The truth comes out

…or so I thought. Fortunately, before beginning a conversation with this person about a mutual acquaintance we might have, I remembered an important conversation. It was with Leah Peter, product lead for Expedia, who works on AI projects, and she had extolled the importance of asking AI how confident it is in its answer.

I realized I’d made two important mistakes here: I had slightly misspelled their name – although within the margin of error that Google often catches – and my query was the sort of lazy that I have previously warned against (hey, nobody’s perfect, right?).

It’s a good thing I did because the response to my querying prompt was shocking:

Me: how confident are you in this response?

ChatGPT: I want to be clear here:

The profile I drafted is not factual — it’s a plausible professional narrative I created based only on the two facts you gave me:

• Born in Australia

• Works in Public Relations in London

The employers I named (Brunswick, Edelman, Tourism Australia) and the education path are illustrative placeholders to show you what a polished, specific PR résumé could look like. Unless you confirm those details, they should not be treated as accurate.

So, in terms of confidence:

• ✅ Confident: The structure, tone, and career progression are realistic for someone in PR with that background.

• ❌ Not confident: The actual companies, job titles, and education are invented unless you verify them.

Hmmm. I appreciated the late onset honesty, but given its total lack of confidence this disclaimer should have been at the top of the original reply. I showed this to the person I was looking up.

They agreed that the resulting profile was very plausible, quite flattering… but completely fake, and we proceeded to have an actual conversation about their working life.

A long way to go

I consider myself a pretty sophisticated user of AI (the odd lazy prompt aside), but I almost was fooled by the initial response. In this case, it was only casual curiosity with no real consequences. But others asking similar questions, and then presenting those ‘facts’ to others, could be embarrassed badly.

Imagine applying for a job and researching the hiring manager. Or going on a sales call and trying to butter up the buyer by demonstrating how much you knew about them. You’d be tossed out so fast, you’d only have time to hear the door lock behind you. 

There’s an old truism, “measure twice, cut once”. The same logic holds true when you use AI. “Ask once. Check twice.”

Peter Horan
Peter has published a number of technology magazines and sites over the years. His current passion is around AI.
Topics
Windows 11 is getting a new Screen Tint mode, and your eyes might thank Microsoft
Users can apply custom color overlays to reduce screen intensity and visual fatigue.
Windows 11 on a laptop

Microsoft is testing a new accessibility feature for Windows 11 called Screen Tint, and it could be one of those small additions that make a surprisingly big difference. Instead of changing your display's color temperature like Night Light, Screen Tint applies a customizable color overlay across the entire screen, making bright displays easier on the eyes during long work or gaming sessions.

A softer screen for tired eyes

Read more
Apple’s looking at a politically radioactive fix for the memory crisis, and the US government isn’t happy about it
Apple blamed memory costs for your price hike. Its proposed solution involves a Pentagon blacklist.
Apple Mac Mini on a Desk

A few days ago, Apple announced an ugly mid-cycle price hike, blaming the worsening-by-the-day memory crisis. According to the Financial Times, the company is now lobbying the government for approval to buy memory chips from a Chinese company. 

The company in question is CXMT, a Chinese chipmaker that the Pentagon added to its Chinese Military Company blacklist for alleged ties to the Chinese army.

Read more
As iPads get pricier, Motorola’s Pad 70 Pro arrives as a solid option… just not for US buyers yet
Great specs, a stylus in the box, and no US launch date: the Moto Pad 70 Pro sounds both impressive and disappointing.
Computer, Electronics, Laptop

If you don’t know about Apple’s recent price hike, which affected all the products in its lineup except the iPhone and Apple Watch (for now), you’ve got to be living under some sort of a rock. The revision made all the iPads much more expensive. 

Motorola, however, has just launched a 13-inch tablet that actually sounds good on paper. It’s called the Moto Pad 70 Pro, and it costs around $440 for the baseline model. The catch, however, is that the device isn’t available in the US yet. 

Read more