Assistance must not erase accountability.

Nardux considers AI only for a defined purpose with reviewed data, meaningful human authority, measurable limitations and a safe way to pause or stop.

Use technology only where the operating case is supportable.

01

Purpose before tool

Name the decision or repetitive task, the expected benefit and why simpler rules or process changes are insufficient.

02

Minimum necessary data

Use the least sensitive input and the narrowest retention, access and reuse permissions that support the approved task.

03

Human authority

People retain the ability and responsibility to review uncertainty, sensitive decisions, appeals and material exceptions.

04

Measured limitations

Evaluate errors, disagreement, drift, uneven impact and failure conditions rather than treating fluent output as proof.

05

Reversible operation

Define monitoring, escalation, fallback, change control and stop conditions before production dependence develops.

Six questions before an AI-assisted workflow.

  1. 01

    What exact task and decision outcome is being assisted?

  2. 02

    Which data enters, leaves, persists or may be reused by the tool?

  3. 03

    What errors, bias, privacy, security and user harms are plausible?

  4. 04

    Where can a qualified person review, override, appeal or stop?

  5. 05

    How will quality, drift, incidents and downstream effects be monitored?

  6. 06

    Who approves the vendor, policy, deployment, change and exit?

Start with low-authority, reviewable tasks.

These are examples for assessment, not a promise that a particular tool or workflow is approved.

ROUTING

Organization and prioritization

  • Suggest a queue or category
  • Identify possible duplicates
  • Extract approved structured fields
  • Surface missing information
  • Recommend an escalation path
REVIEW SUPPORT

Context for a person

  • Summarize approved case material
  • Retrieve relevant policy sections
  • Draft internal rationale for review
  • Flag uncertainty or disagreement
  • Support controlled evaluation sets
NOT IMPLIED

Autonomous authority

  • Unreviewed high-impact enforcement
  • Automatic legal determinations
  • Unapproved user profiling
  • Hidden employee surveillance
  • Training on client data by default

Contextual and consequential decisions stay reviewable.

Human review is retained for ambiguous policy, sensitive content, critical actions, credible harm, legal or specialist triggers, appeals and cases where the system or data falls outside evaluated conditions.

Automation can be paused when confidence, quality, data flow, permissions, vendor behaviour or operating context changes beyond the approved boundary.

Nardux does not claim that AI removes risk, guarantees accuracy or replaces the client’s policy and product accountability.

Evaluate the workflow, not only the model output.

The review method is shaped around the task, risk and available ground truth.

01

Bounded test definition

Approved use case, input range, policy version, exclusions, expected output and stop conditions.

02

Error and disagreement taxonomy

False routing, missing context, unsupported output, uneven impact, privacy events and human disagreement.

03

Human comparison

Qualified review of representative, difficult, changed-policy and anomaly cases.

04

Change and incident record

Model, prompt, vendor, data, threshold, policy and workflow changes tied to re-evaluation decisions.

Approval follows the complete data flow.

Client content is not entered into a public or personal AI account. Any proposed provider requires business, privacy, security and contract review.

Input and output data classesPurpose and lawful instructionsProvider retention and training useSubprocessors and locationsIdentity, roles and support accessLogging and incident notificationDeletion and export capabilityFallback and provider exit

Questions about scope and authority.

Does Nardux use AI on every moderation engagement?

No. Human-led work does not require AI. Assistance is considered only where the task, data, vendor, risk, client permission and human-control model are approved.

Can AI automatically remove content?

This page does not offer autonomous enforcement. Any automation authority would require a separately evaluated use case, client approval, monitoring, appeal and stop controls proportionate to the consequences.

Will client data train public models?

Nardux does not approve client data for provider training or unrelated reuse by default. Actual vendor terms, settings, data flow and contract controls must be assessed before use.

Do you certify that an AI model is safe or compliant?

No. Nardux may discuss bounded human evaluation workflows but does not claim comprehensive model certification, cybersecurity red teaming, legal compliance or proof that every failure mode has been found.

Define where assistance helps and where it must stop.

Share the workflow, data categories, decision consequences and current human review path.

Discuss a Bounded Workflow