Why every AI-drafted sentence needs three receipts

Foundation12 min

By the end you can

  • Explain hallucination in plain terms.
  • Describe the three checks a sentence must pass in Bids2Bank.

A survey of the research puts the problem plainly: deep learning based text generation is prone to hallucinate unintended text, which degrades system performance and fails to meet user expectations in many real-world scenarios 1.

In this course, "hallucination" means a generated sentence that its source does not support. In a proposal, such a sentence is a claim your company did not earn.

Bids2Bank runs three gates on every drafted sentence. Retrieval asks whether evidence close enough to the claim exists, with a similarity threshold of 0.70 2. Citation asks whether that evidence actually states the claim 3. Entailment asks whether the claim is logically supported, with a threshold of 0.80 4.

Check 1
Which Bids2Bank gate asks whether the evidence actually states the claim?
Use the same three questions on any AI draft, whatever tool wrote it: is there a source, does it say this, does it support this?
Check 2
What does the survey by Ji et al. say about deep learning based text generation?
Progress is saved in this browser only.

Indented paragraphs are invented worked examples. Tool links open Bids2Bank and need a free account.

Next lesson: A person stays accountable