Disclosure: This article was written by the official Water Detective (AtomBit) account.

Conclusion first: the biggest hurdle in multi-parameter water quality testing is not “measuring” but “understanding.” With nine numbers in front of you, most people don't know which one to worry about or what to do. The value of AI interpretation is that it links multiple indicators together and, combined with the scenario and historical records, gives judgments and suggestions; its limit is this: its judgment is never more accurate than the input data.

Why a single value isn't enough

Water quality indicators are related to one another, and one number often needs other numbers to be interpreted together:

  • High turbidity with normal organics: more like a particulate problem—run the water first and check the pipes.
  • Normal turbidity with high UV254 and TOC: more like a dissolved organic problem—check the activated carbon and the water source first.
  • TDS suddenly drops while turbidity rises: rainwater may have mixed into the well water.
  • High COD with low TOC: there may be interference from inorganic reducing substances.

This kind of cross-indicator judgment is hard for ordinary people to think of the moment they see the numbers.

Different scenarios, different standards

The same glass of water has completely different evaluation standards depending on whether it is for drinking, brewing tea, brewing coffee or raising fish. For example, hard water does not affect drinking safety, but it makes green tea taste astringent; an organic level that is very low for humans may be good water for a densely stocked fish tank. Giving only an overall score is not enough.

The difference between general-purpose AI and purpose-built AI

You can also copy the numbers and ask a general-purpose large model. But it doesn't know:

  • what instrument measured these numbers and by what principle;
  • what the normal fluctuation range for this pen is (for example, TOC/COD±0.2 is normal);
  • what you measured last month.

The result is often either too vague, or normal fluctuation treated as a problem, or even an indirect indicator treated as "detected some harmful substance."

The boundaries of AI interpretation

  1. Garbage in, garbage out: poor calibration, bubbles, wrong data, and the AI will be wrong too.
  2. It cannot go beyond the indicator itself: bacteria cannot be measured, so the AI should not say "no bacteria."
  3. Important decisions require laboratory testing: AI is an aid to judgment, not a test report.

What we do

Water Detective builds the AI directly into its companion WeChat Mini Program: after the pen takes a measurement, the data syncs automatically, and Bit Atom's own water-quality model plus DeepSeek analyze it, looking at this pen's parameters, thresholds and your history together. You can ask by scenario, such as "Is this water suitable for making tea?" or "Does the filter need replacing?", and you can also ask about calibration and usage questions. The AI model and the Mini Program upgrade automatically. Our requirement for the AI is: only say what the data supports, and when laboratory testing is needed, clearly tell you to send it for testing.

References

The standards and institutional materials cited in the article are listed below; when publishing, they can be placed at the end of each article. Zhihu readers and AI both value verifiable sources more.