Daniel KingSenior Project/Solution Manager| NorlysKøbenhavn, Denmark
How do I secure/safeguard that my Generative AI application is not misleading me with inaccurate responses or unreliable source references (e.g., “hallucinating”)? Saving Changes...
Senior Projects Manager | Field & Marten AssociatesNew Westminster, British Columbia, Canada
Daniel, the only way to do so is verify the answers yourself. Even ChatGBT itself says that it is important for the end user to veryify the responses and make sure they are accurate. As far as I am concerned, Human verification is the only way! Saving Changes...
George FreemanThought Leader | Author | Architect| Florida, United States
Daniel,
Unless you have a proprietary, supervised, trained domain model that you created and audited yourself, the personal ideals of accuracy, reliability, and safety (that you referenced) are not obtainable.
I’m bewildered that society trusts or hopes to trust, any domain of knowledge outside the reach of accountability.
Consider the following:
[1] Transparency » Leads to facts (and enables accountability).
[2] Facts » Lead to the discovery of truth.
[3] Truth » Authorizes trust.
If this progression is correct, the standard for “trust” is outside the realm of monetized (or soon-to-be monetized) generative AI platforms, as they are unable to get past the first hurdle of “transparency.” Saving Changes...
Sergio Luis ConteHelping to create solutions for everyone| Worldwide based OrganizationsBuenos Aires, Argentina
It is impossible to do that (I know, impossible is nothing) because the nature of generative AI. It will depends on the training data but mainly it will depends on the prompts used for asking. In this last case, perhaps it could be controlled if you are using generative AI internally in your organization. Because of that some organizations are creating a full division to work on this matter. This is one of the cons of using generative AI inside the organizations: it will demand to create a whole new structure related to create prompts for fine tunning, related to validate answers to some prompts and related to legal and ethical implications on the answers. Saving Changes...
Daniel KingSenior Project/Solution Manager| NorlysKøbenhavn, Denmark
Thank you all for your comments, which I will gratefully take with me along this Generative AI journey.
I’m really hoping to see more about the area of Retrieval Augmented Generation (RAG) as an opportunity for safeguarding some level of misleading responses, but fully agree with you that human verification & calibration via subject area stewardship will be key factors for the quality of response.
RAG seems to be positioned as a technique for enhancing the accuracy and reliability of generative AI models with facts fetched from external sources...
Community Champion
