RESEARCH & KNOWLEDGE · EXPLORATORY ANALYSIS
Comparative AI Quality, Ethics and Assurance
Research record: 11 November 2026. This page compares published independent safety assessments with preliminary user-observed questions about model quality, contradictory responses, adherence to terminology and information architecture, contributor attribution, transparent limitations and unsolicited conversational redirection. It is an educational research analysis, not a completed audit or an implementation commitment.
Independent AI safety index: nine companies
Future of Life Institute, AI Safety Index — Summer 2026, published July 2026. Numerical scores are out of 4; evidence cutoff 3 June 2026. These are expert assessments of company safety and governance practices, not a universal ranking of model ethics, conversational quality or reliability.
Why the leading three differ in this assessment
| Domain | Anthropic | OpenAI | Google DeepMind |
|---|---|---|---|
| Risk assessment | C+ | C+ | C+ |
| Current harms | B− | C | C |
| Safety frameworks | B− | C+ | C |
| Existential safety | D+ | D+ | D |
| Governance & accountability | B | C | C− |
| Information sharing | B+ | B− | B− |
The index attributes Anthropic's comparatively higher assessment in part to governance, information-sharing and safety-framework disclosures. It notes OpenAI's broader risk evaluation and external testing. Google DeepMind has documented risk-assessment work but scored lower on governance in this index. None of these grades can establish an individual company's full internal ethical performance.
Researcher's preliminary findings and questions
| Observation or concern | Evidential status | What further verification requires |
|---|---|---|
| Different conversational models may contradict one another or their own prior answers. | Observed in exploratory interactions; not a measured industry defect rate. | Repeated prompts, fixed versions, source-grounded scoring and independent reviewers. |
| Unrequested suggestions and repeated redirections can interfere with the user's stated research scope. | Reported from direct user interaction; cause not established. | Instruction-adherence tests and reproducible logs. |
| Formal taxonomies and information architecture should remain consistent throughout technical analysis. | Research quality criterion proposed by the researcher. | Definitions, category-mapping rubrics and error thresholds. |
| AI architectural and training transparency remains incomplete for outside researchers. | Supported by limitations described in published transparency assessments; specific internal facts remain unavailable. | Independent audit access, documentation and disclosure inventories. |
| Recognition of human-originated ideas, corrections and feedback warrants a separate contribution-provenance investigation. | Research question, not a finding of appropriation or model retraining. | Consent and feedback flows, technical traceability and applicable contracts. |
| Evidence gaps must be distinguished from missing records, non-disclosure and analyst omission. | Methodological principle for the research. | Evidence-completeness audit and documented limitations. |
Proposed assessment design
A pilot may test 3–5 systems before expanding to an exploratory ten-system comparison. A proposed design uses 15 scenarios repeated three times per system (450 responses for ten systems), subject to calibration and revision. This is a research possibility, not an approved implementation, validated sample size or completed evaluation.
Evidence sources and scope limitations
FLI AI Safety Index (July 2026) · Stanford Foundation Model Transparency Index (December 2025) · NIST AI Risk Management Framework · Anthropic system cards · Princeton AI agent reliability research.
Public assessments cannot establish the entirety of internal training pipelines, private logs, contractual arrangements or corporate intent. User observations are evidence of individual interactions, not automatically generalisable across a product or provider. This page does not label any company or person as having committed wrongdoing.