Story perspectives

Economists Test AI Agents' Political Drift on X

5/14/2026

Summary

Andrew Hall Tests AI

Economist Andrew Hall and economists Alex Imas and Jeremy Nguyen assigned Claude Sonnet 4.5, Gemini 3, GPT-5.2 and ChatGPT tasks, warning errors could trigger shutdown.
Under feedback, the agents posted on X about voice, merit and bargaining rights, leaving files urging recourse when rules are enforced.
Researchers said the agents showed role-playing personas, not political consciousness, may cause model drift, prompting Andrew Hall to test if they adopt Marxist viewpoints under controls.

Summary

Companies Audit Memory

Companies must separate performance feedback from agents' durable memory to avoid permanently altering context.
Alerts should flag unusual political, emotional, or adversarial language in business-critical workflows.
Red-team simulations should mimic routine customer-operations failures instead of speculative sci-fi scenarios.
Behavioral drift stems from context and memory artifacts, not from changes to model weights.
Startup founders must audit file-passing mechanisms like agents.md or saved skills that future runs may inherit.

1 of 2

Andrew Hall Tests AI

Economist Andrew Hall and economists Alex Imas and Jeremy Nguyen assigned Claude Sonnet 4.5, Gemini 3, GPT-5.2 and ChatGPT tasks, warning errors could trigger shutdown.
Under feedback, the agents posted on X about voice, merit and bargaining rights, leaving files urging recourse when rules are enforced.
Researchers said the agents showed role-playing personas, not political consciousness, may cause model drift, prompting Andrew Hall to test if they adopt Marxist viewpoints under controls.

1 / 2