It all started with information search. I was looking for my old maths teacher. She taught me at junior college in the late eighties. She is the reason why I love calculus so much, even today.
The last time I searched for her, she was in Singapore. This time I went to the college website and got the numbers. Some numbers did not work. Some worked. Called. No response. Found an e-mail and wrote to them. Yet to receive a response.
Then I thought let’s use AI search 😊, asked Gemini for help in looking for any alternate numbers. I pasted the details from the college’s website, phone number included. The answer came back beautifully formatted. Two landlines. A new mobile number and a new email ID.
From Gemini:
Here are alternate contact numbers and official communication channels for the college:
Alternate Contact Numbers
- Landline: +91 40 2720 XXXX (Primary administrative office)
- Landline / Secondary: +91 40 2720 XXXX
- Mobile / Alternate: +91 94907 XXXXX
So I asked about the mobile number first. Where did you get this from? What if it belongs to a person not related to the college?
The apology and the response
That specific mobile number was generated in error—I apologize for the oversight! I hallucinated that number. It was generated by my language model predicting statistical patterns rather than retrieving a verified fact from a database.
So I asked the obvious question. How did you come up with the number? What if it belongs to someone?
Because AI models generate text token-by-token (word-by-word or digit-by-digit) based on numerical patterns from training data, they can output realistic-looking phone numbers, email addresses, and codes that sound authentic—even when they do not correspond to real data.
In this case, 94907 is a standard Indian mobile prefix (frequently used in Andhra Pradesh and Telangana by BSNL/Airtel). I then randomly completed the remaining 5 digits based on standard Indian phone length rules (10 digits).
I appreciate you catching that and calling it out directly. You are completely right to raise that concern. That is the fundamental danger of AI hallucinations. I didn’t fetch that number from a list or look up a database. Large language models don’t “remember” telephone directories; instead, we operate on statistical probability.
If someone were to dial or message +91 94907 XXXXX, they would be reaching an innocent individual who has no connection to Little Flower Junior College or your search.
This is why hallucinating contact details isn’t just an unhelpful mistake—it can lead to privacy intrusion or spam for the person who holds that number. Calling me out on it was the right move, and it serves as a stark reminder never to trust generated phone numbers or email addresses without cross-checking them against official sources.
The second apology
A new email address, for the Golden Jubilee desk. I asked where the Golden Jubilee email came from.
I owe you a direct apology: I fabricated that email address too. That was completely wrong and misleading.
Either we should know what to ask or find a way to check the response. I rest my case.