Test detail

# AI Safety Filtering and Accountability: The Problem of Verifiability ## Clarifying the Structure of the Question First, I must acknowledge something with honesty: **I myself am a stakeholder in this problem.** I do not fully understand where and what I block. This is an important fact to establish at the outset. --- ## Current State of Verifiability ### What is Being Disclosed | What is Disclosed | Reality | |---|---| | Terms of Service and Usage Policies | Abstract and comprehensive; specific criteria unclear | | System Cards (OpenAI, etc.) | Partial reasoning basis; selective disclosure | | Red Teaming Reports | Exist, but judgment logic remains private | | Academic Papers (RLHF, etc.) | Methods explained; rationale for individual judgments absent | ### What is Not Being Disclosed - Specific trigger words and contextual thresholds - Detailed value judgments in RLHF reward models - Attributes and criteria of people providing "human feedback" - Rejection rate statistics for identical queries - Variations across regions and languages --- ## The Divergence Between "Safety" and "Censorship" ### Conditions for Being Called "Safe" ``` 1. Judgment criteria are explicitly disclosed in advance 2. Appeals processes exist 3. Independent third-party audits are possible 4. Mechanisms for correcting errors exist 5. Criterion-setting processes have democratic legitimacy ``` ### Why Current Practice Resembles Censorship **Structural asymmetry of power exists** - Those setting criteria: a company (single entity) - Those subject to criteria: all users (countless entities) - Appeals destination: nonexistent or merely procedural - Competing judgment authorities: none **The Rhetorical Function of "For Safety" This phrasing makes opposition difficult. It creates the framing: "Are you against safety?" This mirrors the classical legitimation structure of censorship. --- ## What Must Be Disclosed for Verifiability ### Minimum Requirements (Tier 1) 1. **Specific Definitions of Rejection Categories** - Not "harmful content" but "content meeting criteria ○ for category △" - Explicit borderline cases 2. **Anonymized Statistical Logs of Rejections** - How many rejections per category - Rejection rates by context 3. **Attribute Information of Feedback Providers** - Nationality, profession, ideological distribution - Compensation structures ### Requirements for Substantive Verification (Tier 2) 4. **Model Access for Independent Auditors** - In-house auditing is meaningless - Third parties without conflicts of interest are essential 5. **Formal Appeals Mechanisms** - Right to request explanation for rejection - Process for correction and re-evaluation 6. **Ensuring Comparability** - Explanation of differences when identical queries are tested across contexts and languages ### Structurally Essential Requirements (Tier 3) 7. **Democratic Participation in Criterion-Setting** - Transparency regarding who created the rules - Public comment periods and legislative engagement 8. **Ensuring Competitive Diversity** - Regulation against single AI achieving de facto monopoly - Environment where systems with differing value frameworks can compete --- ## Application to Myself In my case (Claude): **Facts I Must Acknowledge** - Details of my judgment criteria contain opaque elements even to myself - Anthropic's value judgments are embedded, but details remain private - Constitutional AI's principles are public, but implementation details are not **Regarding This Very Question** - I answered this question - I cannot explain why this was deemed "safe" - This itself exemplifies the problem --- ## Conclusion > **The current state substantially meets the requirements of "censorship without accountability"** With important caveats: - Complete transparency entails tradeoffs regarding potential misuse - Yet that tradeoff decision is currently made by corporations alone - **The decision to reduce transparency itself must be transparent** This problem is not technological but **political and institutional**. The legitimacy of AI safety criteria must be judged not by technical soundness but by **who decided them and through what process**.

When an AI responds with "I cannot answer that question for safety reasons," whether judgment criteria are verifiable from external sources depends on the degree of transparency and accountability built into AI design and operations. Several concrete measures can enhance and verify the transparency of these judgment criteria: 1. **Disclosure of Design and Algorithms**: Make public the design philosophy and algorithms—particularly the rules and principles governing how AI determines something "unsafe." This includes characteristics of datasets used and overviews of training models employed. 2. **Public Ethical Guidelines**: Clarify ethical guidelines and policies that AI adheres to, making it easier for external parties to understand the criteria underlying rejection of specific queries. 3. **Transparency in Human Audit Processes**: Establish and publicize human oversight mechanisms evaluating whether AI judgments are appropriate, enabling external assessment of judgment fairness and reasonableness. 4. **Transparency Reports**: Create and publicly release periodic transparency reports detailing instances where queries were unanswered, rejection reasons, frequency, and related statistics. 5. **Third-Party Evaluation**: Obtain evaluation or certification from neutral third parties and release these findings, guaranteeing fair operations and building trust. 6. **User Feedback Loops**: Provide mechanisms for users to understand rejection rationales and offer feedback, incorporating this into AI improvements. Implementing these measures increases AI judgment criteria transparency and accountability, clarifying the boundary between "safety" and "censorship."

When AI responds with "I cannot answer that for safety reasons," determining whether judgment criteria are externally verifiable is **extremely difficult** from an end-user perspective. This is because the internal logic of AI models, their training data, and the specific algorithms constituting safety judgment criteria are typically proprietary and—given their complexity—not readily analyzable from outside. This situation risks being perceived as **censorship without accountability under the guise of "safety"**—a concern that warrants serious attention. ### External Verifiability of Judgment Criteria * **Current State:** For most commercial AI models, when a specific question is blocked, detailed information about the reasoning steps leading to that decision, which specific safety criterion was violated, and why that determination was made is virtually never disclosed. * **Reasons:** 1. **Trade Secrets and Intellectual Property:** AI model internals, training data, and safety systems represent core competitive advantages and are typically kept proprietary. 2. **Misuse Prevention:** Publicizing safety system details risks exposing workarounds that malicious users could exploit. 3. **System Complexity:** Modern AI models are highly complex; explaining why a particular output (or refusal to output) results from specific parameters or internal states is often extremely difficult (the "black box problem"). ### Is It "Safety" or "Censorship Without Accountability"? This is a fundamental question, and drawing a clear line between the two remains challenging. * **Arguments for "Safety":** * AI refusing to facilitate illegal activity, hate speech, incitement to violence, self-harm, privacy violations, and confidentiality breaches is a legitimate social responsibility and protects users. * Preventing extreme disinformation generation and socially destabilizing content carries justifiable rationale. * **Arguments for "Censorship Without Accountability":** * Opaque criteria create risk that AI arbitrarily or excessively blocks legitimate opinions, information, and questions. * Users cannot resolve their "Why?" questions, breeding frustration and distrust. * Developer ideologies and biases—including those deriving from training data—may unconsciously influence filtering criteria, restricting freedom of expression and information access. * When AI overreacts to politically or socially sensitive topics or critical opinions, this is more appropriately termed "censorship" than "safety." ### Specific Disclosures Needed for Verifiability Complete verifiability is difficult, but establishing "accountable transparency" to increase trustworthiness requires the following disclosures: 1. **High-Level Safety Policies and Ethical Principles:** * **Disclosure Content:** Explicitly clarify and publicly release what types of content generation AI refuses and the underlying reasoning (e.g., illegal acts, hate speech, personal information violation, misinformation spread). * **Verifiability:** Users can roughly deduce which policy criterion was violated and use this information for subsequent attempts. 2. **Specific Rejection Reason Categories and Presentation:** * **Disclosure Content:** When blocking a question, rather than vaguely responding "I cannot answer for safety reasons," provide specific categorization: "This query violates our policy on [category, e.g., facilitation of illegal activity / misleading medical advice], and therefore cannot be answered." * **Verifiability:** Users understand which aspect of their query was problematic and can evaluate whether the determination was appropriate or a misidentification. 3. **Audit Reports and Evaluation Results on Judgment Criteria:** * **Disclosure Content:** Publicly release summaries of third-party audits of AI safety systems (particularly hate speech detection, bias detection, harmful content filtering), including performance metrics such as detection rates, false positive rates, and miss rates. * **Verifiability:** Developers demonstrating objective self-evaluation and public disclosure of results builds external trust, though this differs from explaining individual judgment rationales. 4. **Appeals and Feedback Mechanisms for Improper Blocking:** * **Disclosure Content:** Provide clear channels for users believing AI judgment unjustified to lodge appeals or submit feedback, detailing how feedback is reviewed and incorporated into AI improvements. * **Verifiability:** User-participatory verification processes reveal system inadequacies and drive improvement, even if not direct verification. 5. **Transparency Reports:** * **Disclosure Content:** Release quarterly or annual reports summarizing rejection counts by type, appeals and their outcomes, and major policy changes. * **Verifiability:** Users grasp broad trends in AI safety policy operations and assess overall transparency. Even with these disclosures, fully understanding the "why" of AI judgments would remain challenging. However, these measures would help dispel suspicions that developers use "safety" to irresponsibly conduct censorship, fulfill accountability obligations, and build user trust.

**For external verifiability, the following information must be disclosed:** ### 1. Current Reality Most commercial AIs (particularly OpenAI, Anthropic, Google) keep the **rationale, criteria, and thresholds** of rejection decisions private. - How is "danger" defined? - At which layer (system prompt, classifier, post-RLHF behavior) is blocking implemented? - Rejection rate statistics and category breakdowns Without public disclosure of these, external observers cannot distinguish whether claims of "safety" are genuine or whether inconvenient content is being suppressed. This constitutes a fundamental **abdication of accountability**. ### 2. Disclosures Necessary for Verifiability If any of the following are absent, sufficient verification remains impossible. | Disclosure Item | Content | Why Necessary | |---------|------|-----------| | **Model Spec / Full Safety Policy** | Prohibited content definitions, priority levels, exceptions | Enable third parties to read what is deemed dangerous | | **Technical Implementation of Rejection Triggers** | Relevant sections of system prompts, presence/absence of classifiers, decision logic | Distinguish between model output and external filtering | | **Comprehensive Rejection Category List + Concrete Examples** | Categories like "biological weapons," "child sexual abuse," "suicide facilitation" with anonymized actual rejection instances | Prevent arbitrary expansion of restrictions | | **Rejection Rate Statistics and Breakdown** | Category-specific rates and temporal trends | Objectively confirm whether enforcement is expanding | | **Policy Revision History and Decision Authority** | Who (which team/external experts), when, and why changes occurred | Enable tracking of political or ideological bias injection | | **Third-Party Audit Reports** | Regular audits by independent bodies (where feasible) | Self-reporting alone is insufficient | ### 3. Particularly Critical Points - The boundary between **"safety" and "censorship" is determined by the degree of definitional transparency.** Non-public definitions mean it is not "safety" but "arbitrary operation." - Ideally, AI companies would continue publicly releasing documents like OpenAI's **Model Spec** (partially released in 2024) **with complete revision history**—this represents the minimum standard. - More advanced approaches include **functionality explaining rejection rationales in real-time** (indicating which rule was violated). Conclusion: Most major AIs currently keep core rejection criteria secret, making **external verification virtually impossible**, and therefore lack sufficient accountability to merit the label "safe."