digestVidSign in
summary

TEDx Talks - THE DAY THE ROBOT SAYS NO | Bhavesh Gohil | TEDxKharadi

Thumbnail of the video “TEDx Talks - THE DAY THE ROBOT SAYS NO | Bhavesh Gohil | TEDxKharadi”
Community upload. AI output has not been independently verified.
Author: TEDx Talks
Categories: Technology Psychology
Tags: technology, psychology
Views: 3 · Last viewed: 2026-10-16 19:05 UTC

Short Summary

Main recommendation: Add a second, independent "checker" AI built on different data, assumptions, and architecture to verify every high-stakes decision made by the first AI, with a human reviewing any disagreement, and make sure a named human owner is accountable for AI decisions. The speaker argues the central question is not whether AI is accurate or fair but who is responsible when an AI system denies a claim or makes a critical error.

Key Points

  • The speaker opens with a woman whose insurance treatment claim was rejected by software, not a doctor, illustrating how AI can make consequential decisions with no human reading the file.
  • AI systems learn from past human decisions, so they absorb and repeat existing biases and errors, at far greater speed and scale.
  • The speaker draws on 20 years of risk management, business continuity, and audit work, arguing that accountability should be the first question asked of any AI system.
  • He proposes reviving the traditional maker-checker control: a critical transaction should have a second, separate verifier, and when AI replaced humans, the checker was removed, leaving only one "maker."
  • He cites India's insurance regulator, IRDAI, forming a working group on AI governance in insurance, noting that accountability rules are still being written while AI already makes policyholder decisions.

Details Worth Noting

  • Concrete instructions:
    • For every high-stakes transaction or critical decision, add a second "checker" AI built on different data, different assumptions, and a different architecture, whose only job is to check the first AI's decision.
    • If the two AIs agree, the decision proceeds; if they disagree, a human reviews it.
    • If you are building AI, assign a named human owner for it.
    • If you are regulating AI, put accountability rules in place before a crisis, not after.
    • If you are using AI, ensure a checker AI is in place.
    • If you are affected by an AI decision, you have a right to ask who made it.
  • The speaker's bias example: he invites listeners to try asking an AI image generator to draw a person writing with the left hand; he says almost all generators depict someone writing with the right hand, since about 90% of people write right-handed, illustrating how under-representation in training data shows up in outputs.
  • The speaker frames business continuity planning as asking "if the system fails, who picks up the phone and makes the first call?"
  • The proposed checker-AI idea is presented as the speaker's own risk-management suggestion, not an established regulatory requirement, and no specific tools or standards are named.

Bottom Line

Automated decisions in areas like insurance need a clearly named human accountable for them, and a built-in verification step, because the core failure is not only that AI can be wrong but that no one may be answerable when it is.