Text-DaVinci-003/ChatGPT made up research sources?

Question

I'm using text-davinci-003 to find some research backed findings to build up my startup pitch. However, when I ask for the research source (title and url) for AI's answer, the report title and urls given by AI are usually not accessible or searchable online. https://beta.openai.com/playground/p/default-chatOvertime, I started suspecting, are those reports and url fake/generated by AI?Examples of the reports referred by AI: 4. Mental Health Foundation (2015). Why do some people not get the help they need? [https://www.mentalhealth.org.uk/publications/why-do-some-people-not-get-help-they-need](https://www.mentalhealth.org.uk/publications/why-do-some-people-not-get-help-they-need) 5. National Alliance on Mental Illness (2020). Mental health conditions & stigmas. [https://www.nami.org/About-Mental-Illness/Mental-Health-Conditions](https://www.nami.org/About-Mental-Illness/Mental-Health-Conditions) 6. Substance Abuse and Mental Health Services Administration (2019). Mental health services in the United States. [https://www.samhsa.gov/find-help/national-helpline](https://www.samhsa.gov/find-help/national-helpline) - National Institutes of Health (NIH):"The Huppert's Mental Health Spectrum: Population Distributions and Associated Factors". https://pubmed.ncbi.nlm.nih.gov/31407549/.

NtochkaNzvanova · Accepted Answer

Yes. This has been well-documented -- in the days since ChatGPT was released, there have been multiple posts with it making up references, book titles and authors, etc. Galactica did the same, and probably every other LLM.An LLM is not a search engine, a database, or a knowledge base. An LLM generates plausible-sounding text based on the distribution of its inputs. There is no constraint that anything it spits out is factually accurate. Why is this so hard to understand?

32gbsd · Answer

Basically all the output can be generated even the specifics. It is its own source of truth.