-
chatGPT vs deepseek answer much different for same question about same question
I had a question from chatGPT and deepseek that how much STPP to make 300ppm hard water soft?
chatGPT answered 1100ppm and deepseek answered 5ppm.
When i asked chatGPT why your answer is very different than deepseek? It answered both are correct and 1100ppm is theoretical and 5ppm is real world dosage and gave some reasons too that i didn’t understand.
As there is a huge difference between 1100ppm and 5ppm i want to ask real chemists here that how much STPP to make 300ppm hard water soft?
This is the link
https://chatgpt.com/share/67f8a379-029c-800a-a43e-3786b304d06c
-
7 Replies
-
Based on google the suggested amount is:
“When used as a water softener, a rough guide of 3-6 ppm of STPP should be used for each ppm of total hardness (Calcium and Magnesium) as CaCO3:
Therefore if you have 300ppm of hardness the suggested amount is:
from 3 * 300 to 6 * 300 == from 900ppm to 1800ppm.
Both networks are correct. ChatGPT gives you the average final answer, while deepseek gives you the average dose per 1ppm of hardness.
P.S.: Is AI worth spending so much time when you can find the answer in google in 1 sec?
-
Actually i was just testing both ais.
I think deepseek answer is 5ppm STPP to 300ppm hardness. Not 1ppm hardness. This is why it confused me.
-
I ran this on Deepseek and got the following…
“Final Answer:For 300 ppm hard water, you need 0.6–0.9 grams of STPP per liter, depending on efficiency requirements. For 100 liters, you’d need 60–90 grams of STPP. Adjust based on exact water chemistry and system conditions.”
This is much closer to what ChatGPT was saying. I don’t think there can be an “exact” answer because some of the answer relies upon real world conditions. ChatGPT gives you the theoretical maximum.
Interestingly, I asked ChatGPT why there was a discrepancy and it said this…
ME: “did Deepseek make an error? Shouldn’t your responses be the same?
ChatGPT said:
Great question — and you’re absolutely right to ask this. Let me give it to you straight:
Did DeepSeek make an error?
Not exactly — but it was inconsistent across answers.
In DeepSeek’s first answer (3.5–5 mg/L STPP), it clearly underestimated the amount needed to soften 300 ppm hard water. That amount would only remove about 5 ppm of hardness, which is nowhere near 300 ppm.
In DeepSeek’s latest answer (600–900 mg/L STPP), the recommendation is much more realistic and aligns closely with mine from the start (~1.1 g/L for full chelation).
So what happened?
DeepSeek initially applied a rule of thumb commonly used in detergent or light-duty cleaning formulations (where you’re just reducing hardness a bit to prevent scale or improve cleaning), but misapplied it to your question, which was about fully softening 300 ppm hard water.
Then, in the second response, it corrected course by:
Applying proper molecular weight ratios (STPP 368 g/mol vs. CaCO₃ 100 g/mol)
Recognizing that you need a 2:1 or 3:1 ratio to account for real-world inefficiencies
Recommending 0.6–0.9 g/L, which makes sense
Should our responses be the same?
Yes — ideally, all responses should be consistent if the question and assumptions are clear.
In this case, my response and DeepSeek’s final corrected one are aligned. The inconsistency came from their initial answer being too low.
Summary:
DeepSeek’s first answer (~3.5–5 mg/L) = way too low for 300 ppm
DeepSeek’s final answer (~600–900 mg/L) = realistic, consistent with stoichiometry
My original answer (~1.1 g/L) = slightly more conservative, full chelation basis
Bottom line = we agree now
Let me know if you’d like a version of this written for a report or presentation — I can lay it out cleanly.
-
That is good
Thanks a lot
-
-
-
Google is not necessarily more reliable than the AIs and the AI also gives a description of how it calculated the results. So, there are some good reasons to ask the AI a question like this.
-
Both ways are very subjective, but to me verifying AI responses takes much more time than google search on trusted sources. I think to use AI efficiently you need to be an expert in the knowledge domain to eyeball AI errors. Then you can use it as automation. Otherwise much easier would be ask here)
-
I agree. It is very easy to be mislead by AI because it presents answers as if they are just true. If someone does not know enough about a subject they can easily accept misinformation that would be obvious to a subject expert.
-
-
-
Log in to reply.