In the summer of 2026, Kenji had become accustomed to asking his AI assistant questions before asking anyone else.
The assistant lived in his smartphone, listened continuously for commands, and had access to a web of public databases, business directories, government filings, archived websites, customer reviews, and real-time search results. It could summarize a hundred pages faster than Kenji could read a paragraph. Most days, it seemed nearly omniscient.
One afternoon, however, a simple phone call exposed a weakness that thousands of engineers had spent years trying to eliminate.
The call arrived from an unfamiliar number.
Kenji ignored it. A few minutes later he asked his AI:
“Who owns this number?”
The assistant responded confidently.
“It belongs to the customer support center of a rental service you use.”
The answer sounded plausible. He was indeed a customer of a vehicle-sharing service whose support team occasionally called users regarding reservations.
Still, something felt strange.
When he attempted to return the call, the line was perpetually busy.
That evening he searched for the number himself.
The search results told a different story.
Several websites identified it as a source of automated sales calls. Some users described prerecorded marketing messages. Others suspected it belonged to a call-center contractor handling campaigns for multiple companies.
Now Kenji became curious.
“Could the caller have been pretending to be the rental service’s support center?” he asked the AI.
The assistant paused briefly while consulting external sources.
“The number may be associated with business outreach activities. Definitive identification is unavailable.”
That answer was far less certain than the first.
The next day he asked again.
This time the AI delivered an entirely different conclusion.
“The number appears to be associated with the customer support department of an online shopping company.”
Kenji stared at the screen.
Rental service yesterday.
Online retailer today.
Sales operation according to public reports.
Which answer was correct?
To settle the matter, he visited the rental service’s official website and checked its support page. The listed telephone number was completely different.
The original answer had been wrong.
What interested Kenji was not the mistake itself.
Humans made mistakes constantly.
What fascinated him was how the mistake had emerged.
As a data scientist working in telecommunications security, he knew that modern AI systems rarely possess a single authoritative database of telephone ownership.
Instead, they assemble a probability estimate from fragmented evidence.
Telephone numbers change hands.
Companies outsource customer support.
Call centers represent multiple brands.
Marketing firms purchase temporary blocks of numbers.
Voice-over-IP providers can issue thousands of virtual numbers in hours.
Caller-ID information can be manipulated through techniques collectively known as caller-ID spoofing.
The AI had not truly “known” who owned the number.
It had constructed the most probable explanation from incomplete information.
And the available information was messy.
⸻
The following week, Kenji attended a conference on digital trust in Tokyo.
One presentation focused on a growing problem.
Since the rise of large language models, users increasingly treated AI answers as verified facts rather than probabilistic predictions.
The speaker displayed a slide.
A confident answer is not evidence.
The audience laughed uneasily.
The presenter then showed several real-world examples.
An AI had identified a law firm as a bank.
Another had attributed a scientific paper to the wrong researcher.
A third had merged details from two companies with similar names and invented a nonexistent customer-support department.
The errors occurred because language models predict plausible continuations of text. They are excellent at synthesizing information but are not inherently connected to a permanent ground-truth registry.
Even systems augmented with web search face another challenge: the internet itself contains contradictions.
Business directories disagree.
Archived pages persist long after information changes.
User reports are unreliable.
Fraudulent websites deliberately imitate legitimate organizations.
The AI may faithfully summarize inaccurate evidence.
⸻
After the conference, Kenji performed a small experiment.
He asked three different AI systems about the same telephone number.
One identified it as a marketing agency.
One claimed it belonged to an e-commerce company.
One admitted uncertainty and listed several possibilities.
The third answer turned out to be the most useful.
Not because it was more intelligent.
Because it was more honest about what it did not know.
⸻
Several months later, governments and telecommunications providers were deploying new standards based on cryptographic verification of caller identity. Similar to email authentication systems that had evolved to combat phishing, these frameworks attempted to prove that a displayed phone number genuinely originated from an authorized network source.
The technology helped.
But it did not eliminate the deeper problem.
Verification systems can confirm whether a signal came from a particular number.
They cannot automatically determine whether the organization behind that number is trustworthy, whether the call is relevant, or whether the information available online is accurate.
Those questions still require judgment.
Kenji eventually forgot who had placed the mysterious call.
Perhaps it had been a sales campaign.
Perhaps it had been a legitimate business.
Perhaps the number had changed ownership several times.
The answer no longer mattered.
What remained memorable was a lesson about knowledge itself.
The AI had not failed because it lacked intelligence.
It had failed because certainty had emerged from ambiguity.
The world contained incomplete records, outdated databases, recycled phone numbers, contradictory websites, and intentional deception. The AI merely reflected that uncertainty while disguising it with fluent language.
From then on, whenever an AI answered a factual question with absolute confidence, Kenji asked himself a different question first:
“How much of this is knowledge—and how much is just a very convincing guess?”
All names of people and organizations appearing in this story are pseudonyms
Comments