FideAI

Why now

The case for independent scrutiny.

AI capabilities are changing. The evidence we use to trust them must keep up. We follow developments that matter for trustworthy systems and informed human judgment.

Selected developments ↓Public commentary ↓How we select sources ↓

Collection reviewed

Access

Can outside reviewers inspect the evidence behind the claims?

Methods

Do the tests measure what the decision requires?

Public judgment

Can people understand the findings and their limits?

Field developments

What changed. Why it matters.

Selected announcements, essays, research, and frameworks. Source dates stay visible; older work is included where it helps explain the present. Each entry separates the source’s position from our interpretation.

OpenAI

Proposed practices

Towards safety cases for frontier AI training ↗

Safety claims need an inspectable chain of evidence.

OpenAI outlines proposed practices for frontier reinforcement-learning training, including evaluation backtesting, immutable transcripts, dissent, and sufficient auditor access to examine safety claims.

Scope and status. The document describes evolving recommendations and an aspirational safety-case framework. Its scope is frontier training, rather than the full set of deployment risks.

Fide’s interpretation

A polished report is only the starting point. Reviewers need to trace its claims to records, test whether safeguards work, and examine what the argument leaves unresolved.

OpenAI · Lama Ahmad

Assessment principles

Priorities and principles for effective third party assessments ↗

Outside assessment is becoming a defined technical practice.

OpenAI identifies safety cases, safeguards, evaluation validity, and incident investigations as priorities. Its principles address agreed scope, proportionate access, technical expertise, conflicts, confidentiality, and editorial independence.

Scope and status. These are published priorities and principles, not an appointment of Fide or a guarantee that every assessor will receive the same access.

Fide’s interpretation

The useful unit of work is a consequential question matched to evidence and expertise. Multiple specialist assessors can contribute without any one organization claiming to cover every frontier risk.

Dario Amodei

Commitment and proposal

We Must Pace the Frontier ↗

Independent scrutiny needs meaningful access.

Amodei announces Anthropic’s commitment to embedded third-party evaluators with ongoing, employee-like access. He describes rights to publish findings alongside protections for sensitive information, and proposes broader coordination.

Scope and status. The announced access commitment is distinct from the essay’s wider coordination proposals. Actual arrangements remain subject to contracts and information protections.

Fide’s interpretation

Independence depends on the conditions of the work: what evaluators can inspect, what they can conclude, and whether they can disclose material limitations and unfavorable findings.

METR

Research · Foundational context

Measuring AI Ability to Complete Long Tasks ↗

Capability scores need a connection to real tasks.

METR studies agent performance through the duration of tasks measured against human completion time, and publishes its infrastructure, data, and analysis. The study discusses task selection and the limits of extrapolating its results.

Scope and status. This is a dated study with a particular task distribution and methodology. Its results are not a current capability estimate or a general deployment guarantee.

Fide’s interpretation

Useful measurement should explain what a score means for a real decision, how the test setting differs from use, and which conclusions the evidence cannot support.

NIST

Voluntary framework · Foundational context

AI Risk Management Framework: Generative AI Profile ↗

Trustworthiness belongs across the AI lifecycle.

NIST’s Generative AI Profile accompanies its AI Risk Management Framework, helping organizations incorporate trustworthiness into the design, development, use, and evaluation of AI systems.

Scope and status. The framework is voluntary. Referencing it does not establish accreditation, certification, or a finding that a system is trustworthy.

Fide’s interpretation

Evaluation should account for people, institutions, and the setting of use. Technical performance alone cannot answer every question about appropriate reliance on AI.

Why now

Independent evaluation is becoming a shared priority.

The statements collected here describe commitments and proposals for outside evaluation. Access terms differ by organization; the posts do not establish an appointment for Fide. Useful scrutiny requires rigorous methods, inspectable evidence, and relevant expertise.

The call for independent evaluation

View on X ↗
01 / 14

Post screenshots saved September 25, 2026.
Public statements, not endorsements of Fide AI.

Editorial approach

Follow the evidence, beyond the headlines.

We select coverage that changes how AI can be tested, scrutinized, or responsibly used. Primary documents anchor our summaries. Public commentary supplies perspective, including concerns about incentives, expertise, and the limits of current methods.

This is a curated collection, not a live news feed. A proposal, announced commitment, research result, and implemented practice are different kinds of evidence. Dates describe the source or our review; they do not establish that every proposal has been implemented.

Outside coverage and public statements are attributed to their authors. They do not imply an endorsement of Fide, a partnership, or an evaluator appointment.

Explore Fide’s research →Use public tools →Commission an evaluation →