The most common reason for letting AI output JSON overturn is not that it "doesn't understand JSON", but that the target you give is too wide, the fields are too loose, and the fault tolerance is not compensated. A model can learn to speak structurally, but it is not a naturally reliable serialization program. You only ask for "please return JSON" without restricting fields, types, and missing value handling, and at the end, there is an explanation, one less quote, and the field name is suddenly rewritten, all of which are high-frequency results.
Common rollover points usually come here
| Question | Surface phenomenon | More likely source |
|---|---|---|
| Multiple output captions | JSON has more natural language before and after | You didn't explicitly allow only returning structs |
| The field fluctuates between more and less | Sometimes the key is missing, sometimes the name is changed | Field definitions and required rules are unclear |
| The type is unstable | Numbers become strings, arrays become single values | There are no constraint types and null policies |
| The content seems legitimate but not usable | The format is correct, but the value is dirty | Missing business checksum retries |
A more stable approach is not a prompt word, but a small set of constraints
- First, narrow the field, keep only the keys you really want to use, and don't let the model generate more than a dozen layers of nested objects.
- Each field is clearly written about the type, whether it is required, and what to fill in if missing, such as an empty string, an empty array, or null.
- Clearly tell the model: only return JSON, don't explain, don't markdown, don't block code.
- Do a parsing check on the access side, feed the error message back if it fails, and let it try again as an error, instead of directly believing the first result.
Don't mistake "JSON output" for "data is reliable"
The structure is correct, just that the shell is like JSON; What really determines whether the field is missing, whether the value is dirty, and whether the meaning is off. For automated processes, prompts are only the first layer, followed by schema checks, default policies, and failed retries. The key to making AI stable is not to urge it to be "more careful", but to make up for the format constraints and program verification together.
