The Need for an AI Content Verification Layer

ChatGPT is the fastest growing product in history, exceeding the likes of TikTok and Instagram. It’s captured everyone’s imaginations, and rightfully so, the technology is incredible.

That’s partially the problem however: it builds from imagination. It’s a generative AI and like any generative model, it often presents false information. Sometimes it even doubles down on those falsehoods being correct, here’s an example:

Only numbers 4 and 5 are accurate; the rest don’t exist or simply have wrong authors.

This problem of false information only compounds over time and erodes trust in your generated information.

Google and Microsoft (via Bing) are thinking about these problems very deeply. But what if your knowledge base isn’t publicly available for GPT to be trained on or the Google crawler to scrape?

To solve this, organizations need to build their own verification layer. Let’s walk through using OpenAI & Mixpeek together to train and verify our knowledge base.

We’ll assume our OpenAI model has already been trained on a knowledge base. You can review the docs to learn how: https://platform.openai.com/docs/guides/fine-tuning

Ask Question

We’ll ask a simple question about a restaurant’s menu:

import openai

question = "What were the last 10 soups of the day?"

# ask away
responses = openai.Completion.create(
  engine="my_finetuned_model", 
  prompt=question
)

In this experiment, we’re asking a question on our text model, called my_finetuned_model that’s already been fine-tuned on our organizations’ knowledge base. In this case, menu items.

Intercept Answers with References

We can then go through each response and query our intelligent file store:

import mixpeek

# go through each response and search the file store
for response in responses:
  accuracy = mixpeek.verify(response)

  # does the AI-generated response exist in our file store?
  if accuracy >= 75%:
    print("answer is accurate")

By combining OpenAI’s text generation library with Mixpeek’s intelligent file store, we can intercept the responses and look each one up. If the answer has a high degree of accuracy, we can return it to our clients.

Proposed Architecture

Verification Layer Private Beta

Mixpeek’s private beta is opening up to select organizations, to get on the waitlist enter your information: https://mixpeek.com/start