AvalaAvala
Правовая информация

Ethics Policy

Последнее обновление: 6 ноября 2026 г.Версия: 2026-11-06

Avala is Fleet Learning Infrastructure for Physical AI: the loop that turns a fleet's own sensor data into its next model. Our platform powers safety-critical systems — from autonomous vehicles to robotics — where the quality and integrity of training data directly affects human safety. This responsibility shapes every decision we make.

We publish these principles to hold ourselves accountable to our customers, our global network of coworkers, and the broader AI community.

Our Principles

1. Build for Safety

AI systems trained on our data operate in the physical world. We treat annotation quality as a safety requirement, not a metric. We maintain rigorous quality controls, audit trails, and verification processes because errors in training data can lead to real-world harm.

We will not accept work that is designed to cause harm, violate human rights, or enable mass surveillance of individuals.

2. Eliminate Bias

Training data reflects the choices of the people who create it. We actively work to identify and remove bias from our datasets across all protected characteristics, including race, ethnicity, gender, age, disability, socioeconomic status, sexual orientation, and political or religious belief.

We invest in diverse annotation teams, structured review processes, and bias detection tooling. We acknowledge that eliminating bias is an ongoing effort, not a one-time fix.

3. Treat Our Workforce Fairly

More than 15,000 registered coworkers across five continents work flexibly through Avala and are paid for accepted work. We reject the anonymity and churn that have defined much of the data labeling industry.

We commit to:

  • Fair pay — rates set against local living standards, not race-to-the-bottom pricing
  • Training — programs that build specialized skills in sensor data, review, and quality
  • Paths to more skilled work — reviewer, adjudicator, and specialist roles open to coworkers who earn them
  • Safe working conditions — content protocols that protect coworkers from harmful material, with support resources available

4. Be Transparent

Our customers have the right to understand how their data is processed. We maintain full audit trails from raw data to the released dataset. We do not use opaque subcontracting chains. When AI-assisted tools are part of the annotation pipeline, we disclose this clearly.

We are open about our methods, our limitations, and our mistakes.

5. Protect Privacy

We handle sensitive data — medical imagery, street-level scenes, personal information. We follow strict data handling protocols, maintain SOC 2 compliance, and adhere to GDPR, CCPA, and other applicable privacy regulations. Customer data is segregated, access-controlled, and never used beyond the scope of the agreed engagement without explicit consent.

See our Privacy Policy for details.

For people who record, and for people who appear in those recordings, we also commit to the following:

  • A contributor is paid for accepted work, under the Contributor Terms, and is not asked to record anyone under 18.
  • Another adult who will appear in a home recording accepts the Household Release first.
  • We do not identify bystanders, and we do not sell biometric identifiers. The Biometric Policy states the retention schedule.
  • Customers who license captures agree, in the Dataset License, not to re-identify a person and not to build a biometric template for identification.

We will not provide data or services for mass surveillance of individuals, for biometric identification of people in public or in the home, for re-identification of a person in a dataset, or for weapons. Customer data remains limited to the engagement the customer agreed to.

6. Align AI with Human Values

We believe AI development should be guided by human values. Through our alignment work, we contribute to AI safety research and responsible development practices. We use human feedback to help align AI systems with human intent — not just optimize for narrow performance metrics.

7. Share What We Learn

We contribute to the broader AI safety and data quality conversation through published research, open standards participation, and collaboration with customers and partners. We believe the industry benefits when best practices are shared rather than hoarded.

People in our data

Adult consent runs through the Contributor Terms and the Household Release. Before you record, every adult who will appear or be heard in a capture made in a private place must accept the Household Release. Captures do not include minors. Avala does not identify bystanders. Avala pays for every accepted capture. Avala deletes a capture on request, as the Privacy Policy and the Dataset License describe.

Uses we will not support

Avala will not support, and a Data Customer may not use Capture Data for, the uses barred by Dataset License section 2:

  • Re-identification of a person in a capture.
  • Biometric templates, face embeddings, voiceprints, or similar identifiers for identification.
  • Surveillance or law-enforcement identification.
  • Weapons systems.
  • Ad targeting of people in captures.

Governance

These principles are reviewed annually by Avala's leadership team and updated as our understanding of AI ethics evolves. Our Data Protection Officer oversees compliance with privacy and ethical data handling requirements.

We welcome feedback on this policy. Contact Avala AI, Inc. at ethics@avala.ai, 403 Broderick Street, Unit 3, San Francisco, CA 94117.