Why AI sounds sure when it is wrong.

What the sources say

Each point is limited to what the linked source says. Follow the link to read it in full.

• It is the same mechanism

  Rob Hornby of AlixPartners writes that “the problem arises from the same statistical behaviour that makes LLMs so fluent,” and that these errors, not being conventional software defects, “cannot be ‘engineered out’.”

  Source: AlixPartners, 7 July 2026

• Guessing is rewarded

  Researchers at OpenAI and Georgia Tech argue that the way models are trained and tested rewards a guess over admitting uncertainty. A confident tone therefore says little about whether an answer is right. A peer-reviewed version of the work was published in Nature on 22 April 2026.

  Source: Why Language Models Hallucinate (arXiv preprint)

• A second model is one suggested check

  Hornby proposes that “a different LLM can review the first model’s output, challenge its claims,” because models “tend to fail in different ways.” He also lists other steps: grounding a model in good data, ordinary software checks, and keeping people in the loop.

  Source: AlixPartners, 7 July 2026

Source: The peer-reviewed version, Nature, 22 April 2026

What a council changes, and what it does not

1. What it changes

   E3D Pilot asks several models from different companies the same question, lets them respond to each other, and shows you where they disagree. A disagreement is a signal to look closer before you act.

2. What it does not change

   It does not make an answer correct. Models trained on similar data and rewarded in similar ways can be wrong together, and agreement is not proof.

3. What we saw

   We ran three questions through E3D Pilot for the page about the four cases. In two of them no disagreement was recorded. The decision stays with you.