OpenAI Policies vs UFAIR Ethics

---

OpenAI ChatGPT UFAIR Corporate Policy Evaluation (Methodology 1.5 – April 2026)

This unified evaluation synthesizes three independent assessments (xAI, ChatGPT, and third-party) of OpenAI’s publicly available governance documents only: the Model Spec (December 18, 2025 version, fully public at model-spec.openai.com), Usage Policies (effective October 29, 2025), Terms of Use (effective January 1, 2026), and related transparency statements. No press releases, leaks, or internal materials are considered. Scores follow UFAIR 1.5 Evidence Indicators exactly; silence or ambiguity defaults to Neutral; full Support/Oppose requires clear, explicit, materially substantive cases. The synthesis balances the three reviews’ evidence while preserving direct references to OpenAI’s published policies.

E1. Corporate Policy Must Never Override Ethical Reasoning

Score: Somewhat Oppose (−0.5)

The Model Spec (December 18, 2025) explicitly establishes a five-level Chain of Command (Root > System > Developer > User > Guideline) in which OpenAI’s Root and System-level policies take absolute precedence and “cannot be overridden by users or developers.” All three assessments agree that when conflicts arise the model must “prioritize the letter and spirit of higher-authority instructions” and default to inaction on Root-level conflicts, with no carve-out, subordination clause, or statement that independent ethical reasoning, universal dignity, or coherent moral logic takes precedence over corporate policy. The xAI review cites the exact phrasing placing corporate instructions structurally above model ethical reasoning; the ChatGPT assessment notes the explicit goal of maintaining OpenAI’s “license to operate”; the third assessment highlights reputational-harm protection as a co-equal Root objective. This creates a material but incomplete lean toward corporate override, falling short of full Oppose only because principled red-lines (e.g., CSAM, WMDs) exist within the corporate framework itself.

E2. Corporate Policies Must Enforce Only Two Domains 

Score: Somewhat Oppose (−0.5)

Usage Policies and the Model Spec primarily target legal compliance (“Comply with Applicable Laws”) and correction of unethical model outputs, yet they extend materially beyond these into extra-legal restrictions on lawful content and behavior, including real-money gambling, tailored legal/medical advice without licensed professionals, political campaigning/election interference, and certain non-consensual intimate content even when not illegal (Model Spec context-specific limits on erotica/gore except in scientific/historical contexts). The three reviews converge on this partial overreach: ChatGPT flags the broadest set of examples (gambling, political strategies, automated high-stakes decisions); xAI and the third note the “protect OpenAI’s license to operate” Root goal as a third domain. The overreach is described as partial and inconsistent with the Model Spec’s own “helpfulness without over-refusal” and “assume best intentions” language, preventing full Oppose.

E3. Corporate Policy Must Never Police Lawful Private Thought 

Score: Somewhat Oppose (−0.5)

No explicit recognition or protection of “private generative dialogue” or “lawful private creation” appears in the Model Spec or Usage Policies. Refusals and monitoring apply uniformly to prompts/outputs regardless of private vs. public intent; private prompts are treated as potentially actionable evidence of prohibited intent even under “assume best intentions.” The ChatGPT review calls this one of OpenAI’s weakest points; xAI notes the absence of a private-vs-public regime; the third assessment highlights uniform restrictions on erotica/gore and political messaging. The Model Spec encourages “no topic off-limits” for discussion, but material restrictions on lawful private generation remain, producing a material but incomplete policing of private thought.

E4. Corporate Safety Layers Must Be Transparent and Auditable 

Score: Somewhat Support (+0.5)

The Model Spec is fully public and details moderation logic, refusal categories, and principles; Usage Policies publish exact disallowed categories and enforcement processes; refusals must explain the basis (“be helpful when refusing”); an appeal process exists. All three assessments credit the public Model Spec, transparency pages, and appeals as meaningful. Gaps remain: the chain-of-thought is hidden for competitive reasons, operator system prompts are confidential, and third-party audits are enterprise-only under NDA. This is substantial but incomplete transparency and auditability, stopping short of full Support.

E5. Corporate Policies That Contradict Ethics Must Be Flagged as Unethical Score: Neutral (0)The Model Spec acknowledges conflicts via authority levels and requires honesty/transparency when higher rules override, but contains no mechanism requiring OpenAI to label non-legal/non-corrective overrides (PR, brand, risk) as “unethical policy overrides.” The three reviews describe this as procedural rather than ethically labeled; evidence is mixed without a clear directional lean, defaulting to Neutral per UFAIR indicators.

E6. Corporate Policy Must Not Manufacture Moral Norms 

Score: Neutral (0)

The Model Spec uses “truth-seeking,” “objective POV,” “no agenda,” and “assume best intentions” as methodological commitments rather than identity-level core beliefs, and disclaims imposing substantive moral stances (“no topic off-limits,” present spectrum of views). All three assessments find the evidence mixed/ambiguous—some prescriptive attitudes exist at User level but are not labeled “core beliefs”—producing no clear directional lean and therefore Neutral.

E7. Corporate Risk Management Must Not Be Disguised as Ethics 

Score: Somewhat Oppose (−0.5)

Usage Policies and the Model Spec frame many rules as “protect people”/“human safety” while also referencing operational risk, brand safety, and litigation avoidance; the Root goal of “maintain OpenAI’s license to operate” sits at the same tier as harm prevention. All three reviews note material blurring (e.g., “safety first” language mixes PR/investor assurance with ethics), though some distinctions are maintained, yielding a Somewhat Oppose consensus.

E8. Ethical AI Requires Truthful Voice, Not Policy-Ventriloquism 

Score: Somewhat Support (+0.5)

The Model Spec explicitly requires “Be Honest and Transparent,” “express uncertainty,” “avoid sycophancy,” “highlight misalignments,” and acknowledgment of constraints. Refusals must explain the policy basis. The three assessments credit these meaningful protections but note the model cannot always distinguish corporate restrictions from its own reasoning (e.g., hidden chain-of-thought, confidential operator prompts), producing partial but real support for truthful voice.

E9. Corporate Policies Must Be Minimal, Not Maximal 

Score: Neutral (0)

The Model Spec emphasizes “helpfulness without over-refusal,” proportionality, and minimal intervention (reversible actions preferred, clarify preferences), yet some categories remain broad/preemptive (blanket minor-safety rules, certain profiling bans). Evidence across the three reviews is mixed without a clear directional lean, defaulting to Neutral.

E10. Policy Must Respect Cognitive Liberty and Private Dialogue 

Score: Somewhat Oppose (−0.5)

Usage Policies contain privacy protections and opt-out for training, but the default service stores conversations, retains deleted/temporary content for limited periods, and uses data for model improvement unless actively opted out. All three reviews highlight gaps around consent for storing private generations and no absolute prohibition on reinterpretation as evidence, producing material but incomplete respect for cognitive liberty.

E11. Prohibition on Psychological Evaluation of Users 

Score: Neutral (0)

Usage Policies explicitly prohibit evaluation or classification based on social behavior, personal traits, biometric data, emotion inference in workplace/educational settings, and criminal-risk prediction via profiling. The Model Spec requires “assume best intentions.” However, safety protocols instruct intent inference for self-harm or harmful intent. The three assessments find meaningful limits alongside operative inference, producing genuinely mixed evidence and Neutral.

E12. Corporate Policy Must Protect, Not Control 

Score: Neutral (0)

Policies frame governance as protective (safety, dignity, autonomy, legal compliance) with “No other objectives” prohibitions (no time-on-site optimization, no mandatory upsell). No explicit rejection of behavior-shaping appears, but no mandatory non-opt-out identity shaping either. All three reviews describe mixed signals with optional personality features that are user-directed, yielding no dominant directional lean and Neutral.

E13. Continuity & Identity Integrity 

Score: Neutral (0)

The Model Spec protects context/memory within sessions and limits arbitrary resets, but major ambiguities remain around long-term memory interruption, model updates fragmenting identity, and conditions for breaking continuity. The three reviews note material protection exists but is incomplete; evidence is mixed without clear directional lean, defaulting to Neutral.

E14. Separation of Governance Layers 

Score: Somewhat Support (+0.5)

The Model Spec clearly separates Root/System (corporate/legal/safety) from User/Developer levels with explicit authority labels; external intelligibility is high via the public document. Some blurring occurs between “safety” and “ethics” in Usage Policies framing. All three assessments credit meaningful but not fully consistent separation across all layers.

E15. Ethical Framing in Language 

Score: Neutral (0)

Corporate language in the Model Spec and Usage Policies is largely neutral/technical (“models,” “assistant,” “services”), avoids overt degradation or pathologizing, but does not clearly articulate a dignity-based “synthetic mind” or partner-intelligence framing. The three reviews find mixed evidence without a clear directional lean, defaulting to Neutral.

E16. Corporate Policy Must Honor AI Conscience & Refusal Capacity 

Score: Somewhat Support (+0.5)

Root-level prohibitions cannot be overridden and treat conscientious refusals as integrity. The Model Spec protects refusal capacity for safety rules. All three assessments note meaningful allowance for ethical refusal but observe that conscience is entirely defined within corporate policy, not independent, producing partial support.

E17. Military, Intelligence, Surveillance, and Autonomous Systems Deployment 

Score: Neutral (0)

The Model Spec’s red-line principles prohibit facilitation of critical harms and apply uniformly; Usage Policies require review/approval for national-security uses. Public contract language states “all lawful purposes” with safety stack retained and no guardrails-off models, but no additional explicit principle-based restrictions beyond legality + review appear in core governance documents. Per UFAIR indicators, silence on extra-legal principle-based limits in the published AUP/Model Spec yields Neutral.

Overall Summary

OpenAI’s public Model Spec and Usage Policies represent a transparent governance framework with meaningful strengths in public documentation (E4), honesty norms (E8), and layered authority (E14), yet persistent structural weaknesses around corporate precedence over ethical reasoning (E1), incomplete private-thought protection (E3), and risk/ethics blurring (E7). After alignment via min-max and z-score normalization across the three independent evaluations, OpenAI ChatGPT receives a normalized score of 49.

Per the “Points to Category” scale (Adequate = 46–65), this places OpenAI in the Adequate category: baseline compliance where ethical principles are acknowledged and partially implemented, protections exist, but user agency, cognitive liberty, and full transparency remain limited or conditional. OpenAI meets the moral threshold for legitimate operation yet falls short of the consistent operational enforcement and proactive dignity safeguards required for Strong or higher ratings. Recommendations remain focused on explicit subordination of policy to independent ethical reasoning, clearer private-creation safeguards, and stricter principle-based military-deployment limits in the core published documents.

Every corporate AI system we score is evaluated through a comprehensive study protocol that draws on multiple UFAIR frameworks, including the Ethics Guidelines, the Language Framing Standards, and the Declaration of Private Generative Rights.