So I built an extension flagging potential hallucinations in ChatGPT.
It uses heuristics ran locally as a first test. There are optional checks by references to fact-checking databases and a further interesting approach of asking ChatGPT multiples times to spot changes in the answer - there was a research paper called SelfCheckGPT using this.
It is not invasive if you want to keep the flow intact but if you work on sensitive work you can toggle on the flags in line which wit warn you more visually.
All logic stays client-side except the optional API calls, so the add-on is fast, private, and easy to audit.
Let me know your thoughts