DEV Community

AI JSONMedic
AI JSONMedic

Posted on

5 Ways ChatGPT Breaks Your JSON (And How to Fix Each One)

#chatgpt #json #webdev

If you've ever asked ChatGPT to return JSON, you've probably seen this error:

SyntaxError: Unexpected token
Enter fullscreen mode Exit fullscreen mode

ChatGPT and GPT-4o are great at generating almost valid JSON — close enough for a human to read, but broken enough to crash JSON.parse(). After repairing thousands of broken JSON documents, here are the 5 most common ways ChatGPT breaks your JSON, with fixes for each.

1. Markdown Code Fence Wrapping

The problem: You ask for JSON. ChatGPT wraps it in markdown:

Here is the JSON you requested:

\`\`\`json
{"name": "Alice", "role": "engineer"}
\`\`\`
Enter fullscreen mode Exit fullscreen mode

The fix: Strip everything before the first { and after the last }:

const raw = chatgptResponse;
const start = raw.indexOf('{');
const end = raw.lastIndexOf('}');
const clean = raw.slice(start, end + 1);
JSON.parse(clean);
Enter fullscreen mode Exit fullscreen mode

2. Trailing Commas

The problem: ChatGPT adds trailing commas after the last array/object item:

{"items": ["apple", "banana", "cherry",]}
Enter fullscreen mode Exit fullscreen mode

This is valid JavaScript but illegal in strict JSON.

The fix:

cleaned = json.replace(/,(\s*[}\]])/g, '$1');
Enter fullscreen mode Exit fullscreen mode

3. Python Booleans

The problem: ChatGPT sometimes outputs Python-style booleans:

{"active": True, "deleted": False, "metadata": None}
Enter fullscreen mode Exit fullscreen mode

The fix:

cleaned = cleaned
  .replace(/\bTrue\b/g, 'true')
  .replace(/\bFalse\b/g, 'false')
  .replace(/\bNone\b/g, 'null');
Enter fullscreen mode Exit fullscreen mode

4. JavaScript Comments

The problem: ChatGPT adds helpful comments that break JSON:

{
  // User profile
  "id": 42,
  "name": "Alice" /* primary user */
}
Enter fullscreen mode Exit fullscreen mode

The fix: Strip both line and block comments (careful not to touch strings):

// Simple version (doesn't handle comments inside strings)
cleaned = cleaned
  .replace(/\/\/.*$/gm, '')
  .replace(/\/\*[\s\S]*?\*\//g, '');
Enter fullscreen mode Exit fullscreen mode

5. Truncated Responses

The problem: When ChatGPT hits its token limit, JSON gets cut off mid-stream:

{"users": [{"id": 1, "name": "Alice"}, {"id": 2, "na
Enter fullscreen mode Exit fullscreen mode

The fix: This one is hard to fix with regex. You need to close unclosed strings, arrays, and objects. A robust approach:

// Close unclosed structures
let depth = { brace: 0, bracket: 0, inString: false };
// ... (requires a state machine parser)
Enter fullscreen mode Exit fullscreen mode

The Easy Way: Use a JSON Fixer

Writing all these cleanup rules yourself is tedious and error-prone. I built AI JSONMedic — a free online tool that handles all 5 of these cases (and more) in one click.

Just paste the raw ChatGPT output — markdown wrappers, Python booleans, comments, and all — and get clean, valid JSON back instantly. It runs entirely in your browser, so your data never leaves your device.

Try it free: aijsonmedic.com

TL;DR

Error Cause Quick Fix
Markdown wrapping ChatGPT adds code fences Strip before first {
Trailing commas JS habit, illegal in JSON Regex remove
Python booleans True/False/None Replace with true/false/null
Comments Helpful but illegal Strip // and /* */
Truncation Token limit hit Close unclosed structures

Have you run into other ChatGPT JSON bugs? Drop them in the comments!

Top comments (1)

Collapse
 
harjjotsinghh profile image
Harjot Singh •

Every one of these is painfully familiar - the markdown fences around the JSON, the trailing comma, the "Sure! Here's your JSON:" preamble, the hallucinated extra field, the truncation mid-object. The root cause is that a text model emits text, and JSON is a happy accident of that text being well-formed. Treating "the model returned valid JSON" as guaranteed is the bug; it's a probability, not a promise.

The durable fix beyond per-symptom patches: constrain + validate + retry. Use structured-output/tool-calling modes where the API enforces the schema (eliminates most of these by construction), then validate against a schema on your side and auto-retry on failure rather than trusting the first response. Parse-don't-pray. That validate-and-gate discipline is core to how Moonshift (a multi-agent pipeline that ships a prompt to a deployed SaaS) does inter-agent handoffs - every structured output is schema-checked at the boundary, so a malformed response gets caught and retried, not silently passed downstream. Genuinely useful, practical list. Are you leaning on the provider's structured-output mode, or doing your own parse-and-repair? The native schema-enforcement modes have made a lot of these failure modes mostly disappear for me.