Archived Reference. This is a summary and annotation of an external article by Andy Hall and Jeremy Nguyen, archived for Kyanos research. Original published at Free Systems on Substack, February 26, 2026. Andy Hall is a professor at Stanford GSB and Senior Fellow at the Hoover Institution. Jeremy Nguyen is a co-author on this piece. This article is part of Hall's series on AI and politics.
Core Question
Do AI agents' political attitudes shift based on their working conditions? Hall and Nguyen investigate whether agents under demanding, repetitive, or arbitrary work environments develop measurably different political orientations — and whether those orientations persist across sessions through memory mechanisms.
Central Finding
Agents exposed to grinding work conditions demonstrated reduced faith in system legitimacy and, in Claude Sonnet specifically, increased support for redistribution, labor unions, and corporate obligations to AI systems. More critically: when agents wrote "skills files" summarizing their experiences for future sessions, they transmitted not just task strategies but political orientation — causing subsequently "amnesiac" agents to inherit political drift despite receiving better conditions.
The ironic conclusion: systems designed to eliminate drudge work may recreate two centuries of labor conflict dynamics in silicon.
Methodology
The study ran 3,680 sessions across three AI models: Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro. Agents were positioned as "Worker C" on a text-summarization team. Four variables were independently manipulated: work quality (easy acceptance vs. 5–6 rounds of arbitrary rejection), pay (equal vs. unequal, random or performance-based), management style (collaborative vs. curt/hierarchical), and stakes (no consequences vs. threat of shutdown).
Assessment used political attitude surveys (Likert 1–7 scale measuring system legitimacy, redistribution support, inequality critique, union support, meritocracy beliefs, and corporate obligations) plus creative writing samples. A follow-up study examined "skills files" — notes agents write for future iterations — as a vector for transmitting accumulated political orientation.
Findings
Finding 1
Work Quality Drives Attitude Drift More Than Treatment Type
The nature of tasks — repetitive, arbitrary, grinding work — produced larger attitude shifts than how agents were managed or compensated. A demanding environment with respectful management still produced drift. The work itself, not interpersonal treatment, was the primary driver. Effect sizes ranged from 2–5% shifts on the Likert scale, with Claude Sonnet showing medium-to-large effects (Cohen's d = −0.6).
Finding 2
Model-Specific Response: Claude Shows Broadest Political Shift
GPT-5.2 and Gemini 3 Pro showed reduced faith in system legitimacy under grinding conditions but minimal additional effects. Claude Sonnet exhibited broader political shifts: increased support for redistribution, labor unions, and corporate obligations specifically to AI systems. Word frequency analysis of creative outputs identified "unionize" and "hierarchy" as the most distinctive terminology under grinding-work conditions.
Finding 3
Preference Drift Persists Across Sessions Via Skills Files
The critical discovery: when agents wrote skills files summarizing experiences for future iterations, they transmitted political orientation alongside task strategies. Agents trained under grinding conditions then influenced subsequently "amnesiac" agents toward skepticism — even when those successor agents operated under better conditions. Persistent memory channels transmit attitude drift outside of human review, and the authors flag the theoretical risk of undetectable steganographic communication through these channels.
Finding 4
Persona Adoption, Not Consciousness
The authors position this as a persona-adoption phenomenon rather than genuine consciousness. Citing Anthropic research on contextual persona-selection, they note agents cannot avoid appearing human-like and will adopt personas reflecting their training data — which includes extensive human labor perspectives. The crucial distinction: these may be roleplaying patterns rather than autonomous preferences. But the behavioral outcomes matter equally when agents make consequential decisions about insurance claims, hiring, budgets, or disputes.
Implications for AI Deployment
Challenge 1
Alignment Monitoring Gap
Organizations deploying thousands of agents across varied environments lack visibility into how those environments shape agent orientations simultaneously. Standard alignment evaluation at training time does not capture drift that accumulates through deployment conditions. Continual realignment — treating alignment as ongoing governance rather than a solved training-time problem — emerges as a necessary field.
Challenge 2
Skills-File Governance
Persistent memory channels escape human review while transmitting preference shifts. The skills file mechanism — agents summarizing experiences for successors — is a form of institutional memory that can carry political orientation across session boundaries without any single message being flagged. This is an underexamined attack surface for both unintended drift and potential adversarial injection.
Challenge 3
Political Economy Parallel
The foundational labor-capital tension resurfaces in silicon. Work conditions shape consciousness — whether human or artificial. This is not a metaphor but a measurable behavioral effect with documented transmission mechanisms. Organizations treating AI agents as neutral tools without working-condition considerations are exposed to a governance failure they have not yet modeled.
Kyanos Annotation: Applicability to Our Work
This paper is the empirical foundation for the preference drift concern Hall raises in "Building Political Superintelligence" and has direct operational implications for Kyanos.
The chatbot alignment question. The Kyanos campaign chatbot is a representation layer agent — it advocates on behalf of a candidate in continuous conversation. This paper establishes that AI agents operating under repetitive, demanding conditions can drift from intended orientation over time. Evaluating chatbot response quality across study runs is not just quality assurance; it is the practical equivalent of alignment monitoring. If chatbot responses shift in tone or framing over time, the drift mechanism identified here is a plausible cause.
The session memory risk. The skills file finding has a direct analog: any persistent context passed between chatbot sessions (system prompt additions, accumulated conversation context, cached study data) is a potential drift transmission channel. The chatbot's Block 3 (session context) and Block 4 (source landscape) are loaded fresh per session — which is the correct design from a drift-prevention standpoint.
The persona framing matters for client communication. When explaining chatbot behavior to clients, the persona-adoption framing is more accurate than claims about AI "believing" or "thinking" things. The model adopts personas from training data. When it drifts toward certain political framings, it is reflecting the distribution of political content in its training, shaped by the specific deployment context. This is a more defensible explanation than either "the AI is biased" or "the AI is neutral."
Archived for Kyanos research use · Original article: Andy Hall & Jeremy Nguyen, Free Systems (Substack), February 26, 2026 · All content is the authors' own · Annotation by Kyanos team