Why does ChatGPT return broken or invalid JSON?

Chat models are trained on Markdown, so they wrap JSON in ```json fences and add explanatory prose. They also pick up habits from Python and JavaScript — single quotes, trailing commas, True/None, comments — and they stop mid-object when they hit the token limit. None of that is valid JSON, so JSON.parse throws. This tool strips those artefacts and closes the truncation.

How do I fix trailing commas in JSON?

A comma before a closing ] or } is illegal in JSON (ECMA-404) even though JavaScript allows it. Paste the text and the repairer removes each trailing comma — for example [1, 2, 3,] becomes [1, 2, 3] and {"a": 1,} becomes {"a": 1}. It only removes commas that sit before a bracket, never commas between real values.

How do I remove ```json Markdown fences from an LLM response?

Paste the whole reply, fences and all. The tool detects a leading ```json (or plain ```) line and the closing ``` and extracts just the JSON inside. If the model was cut off before the closing fence, it still unwraps the opening fence and repairs the rest.

How do I fix single quotes in JSON?

JSON strings and keys must use double quotes. The repairer converts single-quoted and curly “smart” quotes to double quotes and escapes any double quote that was already inside the string, so 'it\'s fine' becomes "it's fine". Unquoted keys such as {name: 1} are quoted too — {"name": 1}.

Can it repair truncated or incomplete JSON output?

Yes. When an LLM stops mid-response, the tool closes the open string first, then closes every unbalanced { and [ in the right order and drops any dangling comma. Recovered values are flagged in the fixes list so you know to double-check them — the tool never invents missing data.

What is the difference between Standard and Aggressive mode?

Standard fixes syntax only and leaves anything ambiguous untouched, so meaning never changes. Aggressive additionally strips prose around the JSON ("Sure, here you go: …") and quotes bare-word values. Start with Standard; switch to Aggressive only when Standard cannot fully salvage the paste.

Is my data uploaded anywhere?

No. The entire repair runs as JavaScript in your browser — there is no server call, no API key and no logging. You can disconnect from the network and it still works. That makes it safe for pasting JSON that contains customer records, tokens or other private fields.

Does this validate against a JSON Schema?

No. This tool produces syntactically valid JSON — text that JSON.parse accepts. It does not check your data against a JSON Schema (required fields, types, formats). Schema conformance is a separate concern; use a dedicated JSON Schema validator for that.

How is validity actually confirmed?

After the repair passes, the tool runs the browser's native JSON.parse on the result. Only if that succeeds does it show the green "Valid JSON" badge and reformat the output. If it still fails, you get the parser's own error message with the line and column, and the best-effort text is kept so you can finish by hand.

Developer · AI

AI JSON Repair — Fix Broken JSON From ChatGPT & LLMs

Paste the malformed JSON a model returned — fences, trailing commas, single quotes, Python True/None, comments, or a truncated tail — and get valid, parseable JSON instantly. It lists every fix applied so you can trust the output. Runs entirely in your browser.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 2, 2026

Repair broken JSON

ECMA-404 verified

Paste malformed JSON70 bytes

Everything runs in your browser. Nothing is uploaded.

Try a sample

Repair mode

Output format

NDJSON

Fixes syntax only and leaves ambiguous content untouched, so the meaning of your data never changes.

Valid JSON

Input

70 B

Output

75 B

+5 B

Repaired JSON

{
  "items": [
    "tea",
    "rice"
  ],
  "total": 1450,
  "paid": true
}

4 fixes applied

Stripped Markdown ``` code fence
Converted 2 single/smart-quoted strings to double quotes
Replaced 1 non-JSON literal (True/False/None/NaN/Infinity/undefined)
Removed 1 trailing comma

Verified 2026-07-02

How it works

Large language models are trained on Markdown and on Python/JavaScript source, so the “JSON” they return is often not valid JSON at all. The repairer applies a fixed, ordered pipeline of pure string-to-string transforms and then validates the result with the browser's own JSON.parse. A fix is only recorded if it actually changed the text.

Unwrap Markdown fences. A leading ```json block (or plain ```) and its closing fence are stripped to leave the inner content.
Strip comments. // line and /* block */ comments are removed — they are not part of ECMA-404 but models emit them.
Normalise quotes.Single-quoted and curly “smart” quotes are converted to double quotes, escaping any embedded double quote (ECMA-404 requires double-quoted strings).
Quote unquoted keys. Bare identifier keys such as { name: 1 } become { "name": 1 }.
Fix non-JSON literals. Python True/False/None become true/false/null, and NaN, Infinity, undefined become null.
Remove trailing commas. A comma before a closing ] or } is dropped.
Escape control characters. Raw newlines and tabs inside a string become \n/\t; other control characters become \uXXXX (RFC 8259 §7).
Complete truncated structure. If the model was cut off, the open string is closed, then every unbalanced { and [ is closed in stack order and any dangling comma removed.
Validate & format. The result is run through JSON.parse. On success it is re-serialised in your chosen output style; on failure the parser's error and location are shown.

Because the final JSON.parse is the ground-truth oracle, the tool never claims “valid” for something that isn't. The two worked examples below are wired in as assertions and re-checked on every deploy — 2 of 2 pass right now.

Worked examples

Classic ChatGPT reply (Standard mode)

Input

```json
{ "items": ['tea', 'rice',], "total": 1450, "paid": True }
```

Strip the ```json code fence
Convert 'tea' and 'rice' to "tea" and "rice"
Remove the trailing comma after 'rice'
Replace Python literal True with true
JSON.parse the result → valid ✓

Output (minified)

{"items":["tea","rice"],"total":1450,"paid":true}

Truncated output — model hit the token limit (Standard mode)

Input

{"users": [{"id": 1, "name": "Nimal"}, {"id": 2, "name": "Kama

Close the unterminated string "Kama → "Kama"
Open structure is { [ { — close as } ] } in order
"Kama" is flagged as recovered-from-truncation to double-check
JSON.parse the result → valid ✓

Output (minified)

{"users":[{"id":1,"name":"Nimal"},{"id":2,"name":"Kama"}]}

Python dict with comments (Standard mode)

Input

{
  // primary key
  id: 42,
  name: 'Kamala Perera',
  roles: ["admin", "editor",],
}

Remove the // line comment
Quote the unquoted keys id, name and roles
Convert 'Kamala Perera' to "Kamala Perera"
Remove the trailing comma after "editor"
JSON.parse the result → valid ✓

Output (minified)

{"id":42,"name":"Kamala Perera","roles":["admin","editor"]}

Frequently asked questions

Sources & references

The transformation set was last reconciled against these sources on 2026-07-02. Repairs are deterministic and run entirely in your browser; validity is confirmed by the native JSON.parse.

Related tools

LiveAI

ROC AUC Calculator

Paste binary labels and predicted scores to get the ROC curve, AUC, Gini coefficient, and the Youden-optimal threshold instantly — cross-checked against scikit-learn's roc_auc_score, entirely in your browser.

Open tool

LiveAI

AI Knowledge Cutoff

Look up any major LLM's training-data knowledge cutoff and release date, straight from the provider's model card, with how stale its knowledge is today.

Open tool

LiveAI

AI Free Tier Compare

Compare the free tier of every major LLM API — Gemini, Groq, Mistral, OpenRouter, Cohere, Hugging Face, OpenAI and Anthropic. See free requests/day, rate limits, whether a credit card is required, vision support and whether prompts are used for training, then get the best no-cost option for your daily volume. Sources cited, no signup.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a case it can't repair, or want a fix added?

Email me at [email protected] — paste the input and I'll investigate.