Cleaning JSON from ChatGPT and LLMs

Quick answer: When ChatGPT returns JSON, it often wraps it in a ```json markdown fence or adds a sentence like "Sure! Here's the JSON:" — both of which make JSON.parse() fail. The fix is to extract just the JSON before parsing. Paste the raw reply into the AI JSON Cleaner to strip fences, prose, and trailing commas in one step.

The model is trained to be helpful and to format code nicely, which works against you when you need raw data:

• Markdown code fences — the reply starts with ```json and ends with ```. Those backticks are not JSON.
• Conversational preamble or trailing notes — "Here's your data:" before, "Let me know if you need changes!" after.
• Trailing commas, because models mimic human-written JavaScript.
• Smart quotes (" ") instead of straight " when the text passes through a chat UI.

Any one of these throws Unexpected token the moment you call JSON.parse().

A pragmatic approach is to pull out everything between the first { (or [) and the last matching } (or ]):

function extractJson(reply) {
  const fenced = reply.match(/```(?:json)?\s*([\s\S]*?)```/i);
  const body = fenced ? fenced[1] : reply;
  const start = body.search(/[{[]/);
  const end = Math.max(body.lastIndexOf("}"), body.lastIndexOf("]"));
  return body.slice(start, end + 1);
}

JSON.parse(extractJson(reply));

This handles the two most common cases — a fenced block, or JSON buried in prose. For messier output (smart quotes, trailing commas, truncation), the AI JSON Cleaner applies a fuller repair pass and shows you the cleaned result.

Cleaning after the fact is a patch. The durable fix is to make the model emit raw JSON:

• Ask for it explicitly: "Respond with only valid JSON. No markdown, no code fences, no commentary."
• Use the provider's structured-output / JSON mode so the response is constrained to JSON at generation time. See Getting Reliable JSON from LLMs.
• Validate against a schema so you catch a malformed or off-shape response immediately — see Validating LLM JSON Against a Schema.

For the broader set of breakages models introduce — and how to repair them — see Fixing Broken LLM JSON.

If you're piping large JSON into or out of a model, remember that JSON's quotes and braces are token-expensive. Before sending a big payload, estimate it with the JSON Token Estimator — see Counting JSON Tokens for LLMs.

Why does ChatGPT wrap JSON in ```json? It's trained to format code in markdown for readability. The fences look right in a chat window but aren't part of the JSON, so they break JSON.parse().

How do I remove the code fences from ChatGPT output? Extract the text between the fences (or between the first {/[ and the last }/]), or paste the reply into the AI JSON Cleaner which strips them automatically.

How do I stop ChatGPT from adding explanations? Instruct it to return only valid JSON with no markdown or commentary, and use the API's JSON / structured-output mode, which constrains the response to JSON.

Why are the quotes in the JSON "curly"? Chat interfaces sometimes convert straight quotes to typographic ones. Replace "/" with ", or let a cleaner normalize them, before parsing.