The Flaws of AI with Numbers

 I also love to call this - The simple errors of AI LLMs from an evaluator's perspective. 

Many AI models find it hard comprehending simple numerical information. I discovered this with ChatGpt and Grok which I use most often. Here's a lattest experience. On Monday October 26 2026 at about 16:00 GMT I had a conversation with the AI model ChatGpt. Below are sections of the chat

My question was: Does daylight saving time have a fixed schedule?

AI responds: Yes. Daylight Saving Time (DST) generally follows a fixed annual schedule, but the exact dates depend on the country.

For example, in the United States:

* Starts: Second Sunday in March. Clocks move forward by one hour.

* Ends: First Sunday in November. Clocks move backward by one hour.

In the United Kingdom and most of Europe:

* Starts: Last Sunday in March.

* Ends: Last Sunday in October

.....

Me: What time of the day is it changed forward or backward

AI responds: The clocks are usually changed at 2:00 AM local time in countries such as the United States.

### In the United States

* When DST begins (March): At 2:00 AM, clocks jump forward to 3:00 AM.

* When DST ends (November): At 2:00 AM, clocks move backward to 1:00 AM.

### In the United Kingdom and most of Europe

The change happens at 1:00 AM GMT (local time differs depending on the season):

* In March, clocks move from 1:00 AM to 2:00 AM.

* In October, clocks move from 2:00 AM to 1:00 AM.

The changes are deliberately scheduled in the early hours of the morning to minimise disruption to people's daily activities.

Me: So what's the time in UK now?

AI responds: The UK is currently on British Summer Time (BST), UTC+1, until the clocks go back on Sunday, 25 October 2026.

Me: Has Sunday 25 October not passed in the UK?

AI responds: Yes. Today is Monday, 26 October 2026. Sunday, 25 October has already passed in the UK.

The UK clocks went backward yesterday:

* At 2:00 AM BST, clocks moved back to 1:00 AM GMT.

* The UK is now on Greenwich Mean Time (GMT, UTC+0)

I should correct my earlier response: the UK is no longer on British Summer Time (BST).

....

Why did ChatGpt make such a simple error of time. 

I decided to try something similar with other AI models 

Here's what transpired with Grok 

I ask Grok about daylight saving time 

Grok responds:**Daylight saving time (DST)**, also called **daylight savings time**, **daylight time** (in the US and Canada), or **summer time** (in the UK, Europe, and many other places), is the practice of advancing clocks by one hour during part of the year (typically spring/summer) so that "spring forward, fall back" — to make better use of the longer daylight hours in summer.

### How it works

- **Standard time** (winter): Clocks stay on the regular schedule aligned with solar time.

- **Daylight saving time** (summer): Clocks move forward 1 hour in early spring, and then move back 1 hour in early fall.

  - This effectively shifts your daily routine (work, school, etc.) by 1 hour earlier relative to sunrise.

  - In practice, this gives you an extra hour of evening daylight after work/school while keeping mornings slightly darker.

The exact clock-change times vary by country:

- Most places change at **2:00 AM local time** on a Sunday (the earliest Sunday in March/November in many regions).

- Europe and some other areas change simultaneously worldwide at 01:00 UTC (so 02:00 local in Central Europe, etc.).

- North America changes at local 2:00 AM, which can cause a temporary 2-hour difference between time zones during the transitions.

### Current status (as of October 2026)

Only about 55–71 countries/territories still observe seasonal DST (a minority of the world).

- **Africa**: Not observed in Nigeria (or most of the continent). Nigeria stays on West Africa Time (WAT) year-round.

- **Europe/UK**: Clocks go back tonight/tomorrow — Sunday, October 25, 2026, at 03:00 local time in most of Europe (CEST to CET) and the UK (BST to GMT). You get an extra hour of sleep in the morning.

- **North America (US, Canada, Mexico border areas)**: Clocks go back on Sunday, November 1, 2026 (e.g., 2:00 AM EDT/EST, etc.). Many other places (like Australia, Chile, Israel) change at different times.

Me: What's the time in UK now 

Grok responds: Right now in the UK (as of October 26, 2026, around 9:04 AM GMT):

London and most of the UK: 9:04 AM GMT (Greenwich Mean Time, UTC+0).

It's Monday, October 26, 2026.

Timeanddate

Daylight saving time (DST): The clocks went back 1 hour on Sunday, October 25, 2026, at 2:00 AM local time. You got an extra hour of sleep in the morning — so now the UK is on standard time (GMT) for the winter period.

...

(Grok timing was wrong. This conversation started at 7.41 PM GMT)

Me: What then did you mean when you said regarding Europe/UK that clocks go back tonight/tomorrow - Sunday, October 25... 

Grok responds: **I apologise for the confusion in my previous reply!**  

You're absolutely right — I mixed up the dates. Let me correct it clearly and honestly.

### What I meant (and the accurate info):

The clocks in the **UK and most of Europe** went **back** (standard time) on **Sunday, October 25, 2026**, at **2:00 AM local time** (or 3:00 AM in some places like the UK after the switch).  

- Before: British Summer Time (BST) / Central European Summer Time (CEST)  

- After: GMT (UTC+0) / CET (UTC+1)  

This gave people an **extra hour of sleep** in the morning — exactly as I said. The change happened **last night** (overnight Saturday to Sunday, October 25/26, 2026). Since then, the UK (and Europe) has been on standard winter time.  

As of **now** (Monday, October 26, 2026, ~9:56 PM local time), the UK is firmly on **9:56 PM GMT**.  

I mistakenly said “go back tonight/tomorrow” — that was a slip; the change actually happened **yesterday night** (Sunday). Thanks for catching that!  

If you want the exact current time in London, any other UK city, or a specific European country, just say the word and I’ll give it instantly. 😊

.... 

( Oh really , I don't think so you might mislead me)

Me: Initially you said it was 9.04 AM in the UK. Have we chatted for more than 12 hours 

Grok responds: **No — we have not chatted for more than 12 hours!**

Here’s the clear explanation:

- When I gave you the **9:04 AM GMT** time in my second reply, that was based on the exact moment I looked it up and responded (early in that conversation, around 9:04 AM UK time on Monday, October 26).

- I then immediately corrected it in my next reply (the one about the “9:56 PM GMT” and the date slip-up), and **that correction was sent only minutes later** (you spotted the inconsistency right away and we kept going straight from there).

So the 9:04 AM reference was just a **temporary snapshot** — not a long conversation. The whole back-and-forth (the daylight-saving question, the UK time query, my correction, and this follow-up) has been very quick.

Right now (as of Monday, October 26, 2026, ~9:23 PM local UK time), the UK is on **9:23 PM GMT** — exactly as it has been since the clocks went back yesterday night. No issue at all!

....

(More confusing response)

Claude and Gemini from my assessment gave factual answers. 

This recent experience brought to my mind previous encounters with AI errors involving simple numeric data.

For example in a YouTube video by expert AI evaluators @amankhan and @petergyang (YouTube @peteryangyt) using Anthropic workbench's (playground) generated AI customer representative for ON shoes, the model erroneously judges 45 as greater than 60. What a simple mistake even 5-year olds can hardly make. Another experience - this time personal - was with Grok. I asked the model to give me some maths questions on factorization for me to solve. In one example after I solved it using an alternative method and skipping outlining some steps for brevity sake the model judged my final answer as wrong. Then it solved it and got the same answer only then did it realize I was correct. I experienced something similar with ChatGpt. Using it for practice it generated a user query and model A and B responses, then I rated the responses on the rubrics on a scale of 1 to 5. Sometimes it will mark a score I gave as too low then gives it own assessment and concludes by giving the exact score I gave (funny) 

I believe there's a flaw in AI models programming generally that causes them to fail to see even something clearly visible to the blind. This especially occurs with numeric data - time, date, numbers, fractions and decimals. 

An AI model can judge 1/5 as > 1/2.

I don't think this has anything to do with LLMs hallucinations nor calibration. 

The Question is: Where could this 'simple' error lie.