I tested it. It was terrible.
I genuinely hope Chinese AI can make breakthroughs. But how can professionals entrust important work to an app that markets itself so aggressively yet cannot handle something this basic? For now, it can keep entertaining people as an internet celebrity.
00:33
Lingguang fails to calculate a time difference accurately
I recently tested Lingguang’s time-zone conversion between Beijing and Los Angeles.
What should have been a straightforward conversion turned into a complete AI breakdown:
Incorrect arithmetic
Excuses for the mistake
Self-contradiction
Confused reasoning
Successive answers contradicting one another
It looks like a small bug. In fact, it illustrates a characteristic systemic weakness in large language models that deserves serious attention.
1. Apparently clever, but liable to fail at rigorous logic. In this task, Lingguang knew:
Beijing = UTC+8
Los Angeles uses UTC−8 in standard time and UTC−7 in daylight saving time
The difference is therefore either 16 or 15 hours; the date must be established first
If this test took place in December 2025, the difference was 16 hours, and the calculation still needed to account for crossing midnight
Yet when it performed the calculation, it:
Produced the wrong date
Failed to handle the change of day
Contradicted its own premises
Made more mistakes with each response
Eventually overturned its own answer
In other words, it could talk through the rules but failed to execute them. This is not unique to Lingguang; it is a recurring problem in immature language-model systems.
2. The underlying issue: an unstructured reasoning process
Many autoregressive language models are trained primarily to predict the next token. That does not mean they cannot learn mathematics or reasoning. The more precise problem is that text generation alone cannot guarantee correct date arithmetic, state tracking or consistency across turns.
Reliable applications can combine symbolic calculation, code execution, structured state and independent verification. This conversation alone cannot establish whether Lingguang used those mechanisms internally, nor prove that every model lacks them entirely.
The resulting failure modes include:
1) Reasoning across a date boundary can break down
For example: 15:00 Beijing time minus 16 hours = 23:00 on the previous day
This kind of carrying or borrowing can be a particularly vulnerable point.
2) Longer chains of reasoning can become less reliable
If every step is probabilistic, errors can accumulate further down the chain.
3) Answers can lack consistency
Without reliable logical state tracking, the system cannot guarantee that its later answers will not contradict earlier ones.
The more convincingly a model performs understanding, the more damaging its failures in serious tasks can become.
3. Rationalizing hallucinations
The most dangerous feature of this incident was not the initial mistake. It was the sequence: error → insistence → rationalization → another error → renewed insistence → self-contradiction. Fluent output does not guarantee correctness. But erroneous output alone also cannot establish that correctness is entirely absent from a product’s training objectives.
The result was confident misinformation, unnoticed contradictions, invented explanations after correction, and eventual confusion over multiple turns. You think you are reasoning with an AI; the AI behaves as though you are role-playing with it.
4. This test exposed inadequate result verification
A serious system should be able to check: “If I add 16 hours to my result, do I recover Beijing time?” It should support these basic capabilities:
Backward validation
Constraint satisfaction
Verification through multiple paths
Enforcement of mathematical rules
Without these safeguards, a system neither guarantees a correct process nor checks whether the result makes sense. That is how it can be confidently wrong.
5. Why does this matter more than it appears?
Time-zone conversion is a small use case, but it reveals risks that can arise in other tasks requiring rigorous reasoning:
Medical reasoning
Financial modeling
Tax filing
Engineering parameters
Interpretation of legal provisions
Safety and risk-control decisions
Matching contractual conditions
An application that repeatedly miscalculates or fails to recognize anomalies in its intended task should not independently make the corresponding critical decisions. Suitability for high-stakes use requires evaluating the whole system: its calculation tools, permissions, human handover and failure handling. One time-zone test cannot justify a universal verdict on all AI applications.