Ghanbari supports deeper AI participation where objectives are clear, performance can be measured and errors can be appealed. He rejects the idea that predictive accuracy alone justifies unlimited political authority.
Three levels of machine power
Advisory systems rank options while officials decide. Delegated systems execute within preset limits—adjusting traffic signals, balancing power demand or flagging procurement anomalies. Sovereign systems would choose goals or impose final decisions without meaningful human override. Public debate often collapses these levels, producing either exaggerated fear or careless enthusiasm.
Where bounded delegation is plausible
Urban traffic control is a strong example: the objective can be specified, feedback arrives quickly and interventions are reversible. Grid balancing, tax-fraud triage and preventive maintenance share some of those properties. Even there, guardrails matter. Fraud scores should trigger review, not automatic punishment; optimization of traffic must not systematically disadvantage a neighborhood.
Where delegation becomes dangerous
Criminal sentencing, asylum, child protection and coercive surveillance involve rights, incomplete evidence and severe asymmetric harm. A model’s confidence cannot replace due process. The greater the irreversibility, value conflict and power imbalance, the stronger the requirement for accountable human judgment and appeal.
Performance must be public
Delegated authority should require a domain-specific license: documented objective, benchmark against human performance, distributional error analysis, security testing, an incident register and an expiration date. NIST’s risk-management approach is useful precisely because it treats trustworthiness as a lifecycle problem rather than a marketing label.
Authority should be earned by domain
There is no coherent reason to grant or deny “AI” authority in general. A system may be excellent at scheduling ambulances and unacceptable at determining political speech. Delegation should expand only through evidence, not technological prestige.
The automation asymmetry
Automated decisions scale faster than appeals. A small model error can become a national pattern before institutions notice. Any system that acts at machine speed needs an equally fast audit and remedy layer.
The future question is not whether AI enters government. It already has. The question is which decisions remain advice, which become bounded automation and which must stay under direct, contestable human authority.
QUESTIONS READERS ASK
Further questions
Could AI replace politicians?
AI may replace parts of analysis and administration, but goal-setting, constitutional judgment and legitimate representation are different functions.
Which government tasks are best suited to AI?
Tasks with clear objectives, measurable feedback, reversible actions and strong appeal mechanisms are stronger candidates than rights-sensitive coercive decisions.
EVIDENCE LAYER
Sources and further reading
- Governing with Artificial IntelligenceOECD ↗
- AI Risk Management FrameworkNIST ↗
- European approach to artificial intelligenceEuropean Commission ↗
Sources support factual context. Interpretive conclusions are presented as Meysam Ghanbari’s perspective, not as settled scientific fact.
