Favorable parole rulings drop from ~65% to nearly zero within each session, then return to ~65% after a food break.
Ethical behavior is not a human prerogative
Intelligibility matters for human decision-making too
Extraneous variables can compromise even expert judgment
Key Concept
Fallibility
AI is not more or less fallible than humans. It is differently fallible.
Human Error
Comprehensible within the same cognitive framework: fatigue, distraction, oversight.
AI Error
A fundamentally different category: no evidence was evaluated. Outputs result from pattern completion, not reasoning.
AI-Specific Fallibility
Hallucinations: A New Kind of Error
Humans make typos; AI fabricates plausible but entirely nonexistent entities.
Responsibility distribution:
AI System: Novel failure type (hallucination)
Developer: Confidence calibration and warning design
Deployer: Context of use (casual vs. critical)
User: Verification responsibility (contingent on understanding)
Unequal shares of responsibility; different remedies at each level.
Nuance
Fallibility ≠ Unethical
The capacity for error alone does not render AI unethical.
Self-Driving Car
Ethically problematic not for potential harm (humans share that), but for its lack of accountability.
Drunk Driver
Inherently maleficent and fully accountable. Entirely different ethical profile.
Going Deeper
The Ethical AI Iceberg
Surface-level reasoning yields shallow conclusions. The substantive challenges lie beneath.
Image by Alex Issakova. Tip: visible compliance. Below: systemic, structural challenges.
Provocative View
"There's No Such Thing as Ethical AI"
Monique Tschofen, Professor at Toronto Metropolitan University:
The designers of LLMs are monetizing their platforms by delivering content that goes expressly against the values that contemporary universities purport to hold.
Core argument: costs are systemic and infrastructural, not individual. The economy sustaining AI is itself unethical.
A deliberately provocative position, useful for revealing the breadth of the debate.
Live Poll
After Everything We Discussed...
Has your position changed since the beginning?
More of a Doomer now
0%
More of an Accelerationist now
0%
More confused (as intended!)
0%
More confident navigating these issues
0%
0 votes
Conclusion
Actor-Based Decomposition
The key question is not "is this fair?" but: where did the failure enter, and where can it be corrected?
Data bias → representational corrections • Annotation bias → labor practices • Design bias → technical interventions • Deployment bias → institutional due diligence • Regulatory failure → political action • User credulity → education
6 principles. 4 actors. Different remedies at each level.
Whose Ethos ?
You will choose the tools, the vendors, the claims you accept.
