Responsible AI

Responsible AI.

Areto Labs builds AI to make online spaces safer and more inclusive. Areto protects fans, followers, brands, athletes and other public-facing people from online abuse and malicious activity, while upholding fairness, transparency and accountability.

AI can reduce harm when it’s used responsibly. Built carelessly, it can reinforce the biases it should be catching. This page sets out the commitments behind how Areto is built and used, and the definitions it uses to tell criticism from hate.

01 · Our commitments

Eight commitments, and what we do to keep them

01

Fairness and inclusiveness

We strive to ensure our AI systems don’t perpetuate or amplify biases, treating all individuals and communities fairly, ensuring representation in our training data, while fostering an inclusive online environment.

  • We rigorously test our models using diverse datasets and annotators to identify and mitigate bias.
  • We continuously monitor AI outputs to prevent disproportionate moderation of specific groups.
  • We prioritize language and cultural inclusivity, adapting our models to recognize context in different communities.
  • We provide transparency about potential biases and limitations in our AI’s decision-making.
  • We have developed AI models trained specifically to detect overt and covert racism and sexism, addressing the unique challenges faced by equity-seeking groups.
02

Transparency and explainability

Understanding how our AI systems work is essential for trust. We strive to make our moderation decisions transparent and explainable.

  • We document how our AI models function, including how they classify harmful content.
  • We provide users with clear explanations of moderation definitions, thresholds and actions, enabling them to see what content has been flagged or removed and why.
  • We engage with our customers and communities to refine and improve transparency practices.
03

Reliability and safety

Our AI systems must operate reliably and safely, ensuring consistent and effective moderation without unintended consequences.

  • We rigorously test our AI models to ensure they perform effectively across diverse scenarios.
  • We implement safeguards to prevent AI from being manipulated or misused.
  • We continuously refine our moderation algorithms to address emerging threats, including new forms of online harassment.
  • We prioritize ethical AI safeguards to prevent over-moderation or censorship while still ensuring digital safety.
  • We enable human oversight, ensuring all moderation decisions are reviewable when needed.
04

Accountability and oversight

We take full responsibility for how our AI systems operate and their impact on digital communities.

  • We maintain clear internal processes for reviewing AI performance and addressing potential issues.
  • We provide users with mechanisms to reverse certain decisions immediately, and report concerns.
  • We continuously audit and improve our AI systems based on feedback and real-world performance.
05

Privacy and data security

Protecting user privacy is at the core of our AI approach. We ensure personal data is handled securely and ethically.

  • We comply with global data privacy regulations (e.g., GDPR, CCPA).
  • We anonymize training and testing data.
  • We use encryption and secure protocols to prevent unauthorized access to sensitive data.
  • We provide users with control over their data and offer transparency about how it is used.
06

Reducing harmful human training and moderation practices

We recognize that traditional model training often relies on human reviewers in harmful, exploitative conditions, particularly in low-income regions. We also acknowledge that human-centered content moderation negatively impacts moderators' mental health and well-being, and limits job opportunities in related fields. Areto is committed to reducing this reliance by using ethical, AI-powered automation.

  • We develop AI tools that minimize human exposure to harmful content, reducing the mental health risks associated with content moderation.
  • We train our models in a way that eliminates the need for underpaid, underprotected human reviewers, ensuring ethical and sustainable practices.
  • We ensure our AI complements and enhances human expertise, integrating ethical oversight and clear escalation pathways to maintain accountability.
07

Human augmentation and control

AI should enhance human decision-making, not replace it. We ensure human oversight remains integral to moderation.

  • We design AI to support content moderators, reducing their exposure to harmful content rather than replacing their expertise.
  • We ensure humans retain ultimate control over critical moderation decisions, particularly in sensitive cases.
  • We avoid full automation in areas where human judgment is essential.
08

Continuous improvement and collaboration

Responsible AI is an evolving field. We commit to ongoing improvements and collaboration with industry leaders, researchers, and stakeholders.

  • We actively participate in discussions on ethical AI and content moderation best practices.
  • We engage with external experts to review and enhance our approach to responsible AI.
  • We iterate on our AI models based on new research, user feedback, and emerging risks in digital safety.
  • We share insights and best practices to contribute to the broader responsible AI community.

02 · Definitions

Hate speech and abuse

These are the categories Areto detects and hides. Each one has its own link, so you can point to a single definition.

Always On

Ableism

Discrimination or prejudice against individuals with physical, intellectual, or psychiatric disabilities.

Homophobia & Transphobia

Discrimination or prejudice against lesbian, gay or bisexual people as well as transgender, gender nonconforming or non-binary people.

Racism

Discrimination or prejudice against an individual or group based on the idea that one particular race is inherently superior to others.

Xenophobia

Having or showing a dislike of or prejudice against people of different nationalities.

Religious Discrimination (Antisemitism & Islamophobia)

Discrimination or prejudice against Islam & Judaism or the people who practice Islam & Judaism.

Sexism

Discrimination or prejudice against an individual or group based on the idea that one sex or gender is better than the others.

Physical Threat

Intention or suggestion to inflict pain, injury or violence against an individual or group.

Adjustable

Personal Attack

An inflammatory comment directed toward an individual or group that is likely to upset, alienate, polarise, or silence users.

Violent Language

Language that contains violent or vulgar themes directed toward a group or individual.

Offensive Language

Rude, obscene, or other hurtful language that can cause offence to an individual.

Body Shaming

Mocking, humiliating, or stigmatizing a person on the basis of supposed imperfections in body shape, size, or appearance.

Gender & Racial Microaggressions

A statement, action, or incident regarded as an instance of indirect, subtle, or unintentional discrimination against members of a specific gender, race or ethnicity. For example: covert and unconscious biases; suggestions of inferiority; sexual objectification & sexism; restrictive gender roles.

Plus: upload your own list of keywords or phrases to further customize Areto's automated moderation.

03 · Definitions

Digital threats and piracy

Areto detects, classifies and deletes the following types of digital harms and piracy.

Finance and Crypto Scams

Posts promoting fraudulent investment schemes, “get rich quick” offers, and fake crypto giveaways.

Porn Bots

Accounts and comments soliciting adult content or directing users to porn sites and similar pages.

Illegal Streaming

Links and promotions for unauthorized live streams of sports events, movies or TV shows.

Romance Scams

Fake friend requests and messages aimed at luring users into fraudulent relationships.

Ticket Selling Scams

Offers for fake or overpriced event tickets.

Lottery and Contest Scams

Fake giveaways designed to collect user data or encourage fraudulent engagement.

Sports Betting Solicitation

Unsolicited promotions of gambling platforms.

Celebrity Impersonators

Accounts pretending to be famous figures to deceive users.

Fake Online Stores & Unauthorized Merchandise

Promotion of fake products or stores, including counterfeit or unauthorized team merchandise.

Other Phishing Scams

Accounts pretending to be reputable with the intention of stealing login credentials, bank account numbers, and other personal info.

Other Risky Solicitations

Requests to follow unrelated pages, groups, or engage with content that may lead to audience scams, identity theft or harmful interactions.

04 · How it works

How moderation settings work

01

Always On

The Always On categories cover hate speech and physical threats.

02

Adjustable

Adjustable categories follow the threshold you set for your community, so you decide what counts as abuse on your channels.

03

Your own words

Add your own words, phrases and hashtags to catch, and turn moderation on or off whenever you like.

04

Hidden, not lost

Areto hides abusive comments, so your audience can’t see them, but you can review them and restore anything hidden. Spam and scams are deleted permanently.

05

See what happened and why

Your dashboard shows what Areto has hidden and deleted, and why.

For more on how your data is handled, read our Privacy Statement and Security & trust page.