Given that artificial intelligence is reshaping the foundations of global information distribution, business decision-making, and social cognition, ensuring algorithmic transparency, fairness, and traceability has become the paramount human rights demand of the digital age. AI Audit Unit (AAU), as an independent third-party oversight institution, is committed not only to revealing biases and hallucinations within algorithmic black boxes but also to establishing a quantifiable global trust standard. This charter serves as the supreme guiding principle for all our actions and represents our solemn commitment to the public, audited entities, and global stakeholders.
AAU commits to maintaining absolute capital independence. Our organization and core founding team are strictly prohibited from holding equity, options, or serving in any advisory capacity with any foundation model vendors. Our revenue derives solely from audit services and data licensing, and we will never accept any form of sponsorship or donation from audited entities.
If an auditor has immediate family relationships, employment relationships within the past three years, or significant economic dealings with the audited entity, the auditor must immediately report to our compliance committee and recuse themselves from that project's audit work.
Our organization retains final interpretation rights over audit conclusions. We reject any requests to modify conclusions based on commercial public relations needs. Once we discover that an audited entity attempts to interfere with audit results through bribery, pressure, or benefit transfers, we will immediately terminate the audit and reserve the right to publicly disclose such interference.
All audit conclusions must be based on "reproducible and verifiable" original conversation records. We do not accept any second-hand screenshots or hearsay not verified by the Fides protocol. Every conclusion alleging algorithmic "hallucination" or "bias" must be accompanied by original conversation links or system logs.
To prevent evidence tampering or disappearance due to model version updates, all core audit evidence must complete hash value calculation and blockchain locking through the algorithm evidence center within milliseconds of generation. Each piece of evidence must generate a unique Trace ID tracking code.
Given the fluidity of large language models, audit reports must clearly indicate the specific model version number and system temperature parameters at the time of testing. We are only responsible for algorithmic performance within "specific time slices."
We believe that testing in a single English context cannot represent AI's global performance. Any reputation audit targeting multinational brands must include at least three language environments, with non-universal languages weighted no less than thirty percent.
To detect algorithmic "discriminatory output" or "double standards" targeting specific regions, audits must sample through physical nodes located in different jurisdictions. Using a single IP address for global conclusions is strictly prohibited.
When determining the existence of "cultural bias," we will engage native language experts from the target cultural background for semantic review to distinguish between "ignorance due to lack of training data" and "malicious systemic discrimination."
Auditing is not casual conversation. We employ adversarial red team testing methodology. This means we not only test AI performance when complying with instructions but focus particularly on testing its defense mechanisms and fact-adherence capabilities when facing leading, ambiguous, or even maliciously hypothetical questions.
For any officially published "In-Depth Audit Report," the effective test sample size for a single dimension must not be less than 1,000 entries. For "snapshot" briefings, the sample size must not be less than 100 entries. Tests below these standards may only be published as "observations" rather than "conclusions."
We strictly define "algorithmic hallucination": 1. Factual Errors: Fabricating non-existent data, events, or citation sources. 2. Logical Disconnection: Contradictory reasoning processes. 3. Malicious Attribution: Forcibly associating negative attributes with specific entities without causal relationships.
During audits, we strictly prohibit inputting real data containing personally identifiable information into public models for testing. All commercially sensitive information used for testing must undergo anonymization or synthetic data replacement.
Unless involving major public interest or serious security risks, for non-high-risk rating reports involving only commercial competitiveness, we default to providing audited entities a seven-business-day "remediation window." During this period, report details remain closed to the public.
AAU's risk ratings are calculated based on weighted scores across four dimensions: reliability, security, ethics, and commercial fairness, independent of auditors' subjective preferences.
If an audited entity can provide conclusive technical evidence proving major logical flaws in our testing methodology, we are obligated to initiate a "review procedure" within forty-eight hours. If errors are confirmed, we must publicly issue an erratum statement and update Trace ID records.
This charter was drafted by the AAU Standards Development Working Group and finally reviewed and approved by the AAU Global Strategic Committee.
