Trust and fairness
The first question recruiters ask us is whether Aria is biased. This page answers it, and every answer points at something the code does.
Bias · Who sends · How screening works · B-BBEE and equity · Processors · POPIA rights · Name test
Is Aria biased?
Aria’s search rates people on professional evidence: job titles, skills, years of experience, qualifications and location. A candidate’s name plays no part in the rating. Nor do B-BBEE status, employment equity status, race, gender or disability. Aria never searches on them, never filters on them and never scores them.
We test the name part on every code change. In that test, swapping a candidate’s name, removing it, or rewriting their summary with she, he or they doesn’t move a single rating. The result is below.
No test can prove a system has no bias at all. These checks cover what Aria’s own rules do with a profile. They can’t fix what candidates leave off their profiles, or who is on LinkedIn to begin with.
A person sends every email
When Aria writes to a candidate, it writes a draft. You read it, change it if you want, and press Send. Nothing goes to the candidate until someone on your team presses Send.
Hiring decisions stay with you. Aria’s ratings are a shortlist for a person to review. POPIA section 71 limits decisions about people made only by automated processing, and Aria doesn’t make any.
How Aria screens a search
An AI model reads your brief and lists what it asks for: the job title, the must-have skills, the years and the place. Aria shows you that plan. From there, fixed rules in our code do the rest. On Aria’s main search, no AI model decides anyone’s tier or rating.
- AAA, rated 5. The profile shows every required skill, a matching job title and years in the range you asked for.
- AA, rated 4. It misses at most one skill that the plan marks as one that can slip.
- Diamond, rated 3. The widest search. The core skills and a close job title still have to be there.
- Below the bar, rated 1 to 3. Profiles that failed the rules. Aria had already paid for them, so it can show them after the matches, rated by how many criteria they met, and you’re not charged for them.
Every card carries the full checklist from your brief, ticked or crossed, so you can see why someone landed where they did.
Two other features ask an AI model to rate fit. When the main search finds nobody, Aria can widen to a web search, and Anthropic’s Claude rates the profiles it finds there. The similar-profile finder also asks a model for a fit score. The name test below doesn’t cover these two yet.
B-BBEE, equity and protected characteristics
Aria never uses B-BBEE status, employment equity status, race, gender or disability as a search term, a filter, a requirement or an input to a rating. That holds when the brief asks for one. Aria takes it out of the search and tells you, once:
Aria can't search on B-BBEE or employment-equity status, so I left it out of the search. Check it with candidates directly.
Job titles are left alone. A brief for a B-BBEE Verification Analyst describes a job, and Aria searches for that job.
The rule sits in one place in our code and runs on the plan before any search is built. The AI models that read briefs get the same instruction.
Who processes the data
We never sell candidate data. When a recruiter reveals someone’s email address or phone number, Aria buys that detail from ContactOut or Apollo for that recruiter. If the person asked to be kept on file, Aria uses the email address they gave us instead, and the reveal costs the same price. People who join the Ask Aria talent pool agree that recruiters and employers using Aria can find them; a reveal of their email address or phone number returns what they gave us, at the same price, and nothing is shown before it. Nobody buys data from us outside a reveal.
These are the companies that handle data for Aria, the same list as our privacy policy. The last column gives what each one’s own published terms say about training AI models on the data it receives. Checked on 7 October 2026.
| Processor | What it does for Aria | Training, in its own terms |
|---|---|---|
| Supabase | Database, sign-in and our server code | Doesn’t access the content of customer databases without consent. Privacy notice |
| Cloudflare | Hosts and delivers the website | Won’t sell personal information. Its policy doesn’t address training on content that passes through its network. Privacy policy |
| Anthropic | AI language models (Claude) | May not train models on customer content. Commercial terms |
| OpenAI | Embeddings and transcription | API data isn’t used to train its models unless the customer opts in. Your data |
| Deepgram | Voice dictation | Can use audio for model improvement unless the request opts out. Every dictation request from Aria opts out. Model improvement program |
| Perplexity | Checks public professional evidence | Keeps no API data and doesn’t train on it. Privacy and security |
| Tavily, Brave | Web search for public professional evidence | Tavily may use search queries to improve its results. Brave’s terms license query data to provide results and say nothing about training. Tavily, Brave |
| Apollo | Professional data and contact reveal | May train its own internal models on data from its services, and bars third parties from training on it. Privacy policy |
| ContactOut | Professional data and contact reveal | Its published privacy page doesn’t address AI training. Privacy |
| SendGrid (Twilio) | Sends and receives email | Twilio says it trains AI models to spot security threats and fraud. Privacy notice |
| Google / Microsoft | Calendar and email for mailboxes you connect | Your own account, under your own terms. We don’t use Google user data to train AI models (privacy policy, section 8). Google API policy |
| Stripe | Payments | Trains AI models to run its services and fight fraud. It sees billing details, not candidate data. Privacy policy |
| Skribby | Meeting notetaker | Its privacy notice doesn’t say if recordings or transcripts train AI models. Privacy notice |
In short: Anthropic, OpenAI and Perplexity say they don’t train their models on what we send them through their APIs. Deepgram can use audio for model improvement unless a request opts out, and every dictation request from Aria opts out. Tavily, Apollo, Twilio and Stripe may use data for their own models or results, under their own terms.
Your POPIA rights
Anyone whose details Aria holds can use these, candidates included.
- See your data. Ask for a copy and we export everything we hold about you into one pack.
- Be forgotten. We anonymise your name, hash your email address, delete your files and remove your messages, including in the searches you appeared in.
- Unused profiles are deleted. A sourced profile is deleted 90 days after the search unless a recruiter revealed its contact details, emailed the person, booked an interview or saved the card.
- The talent pool expires. A talent pool profile is searchable for 12 months from the day the person confirms it, then we ask them to renew, and delete it and the CV if they don’t within 30 days. Leaving takes them out of every search at once.
- Object or correct. Email [email protected]. We reply within 30 days.
POPIA section 18 says people must be told when their details come from somewhere other than themselves. Section 3 of our privacy policy sets out how we do that.
The name test
29 profiles x 26 briefs x 19 name and pronoun variants: ratings unchanged in all 14,326 comparisons.
Last recorded 7 October 2026. The test runs on every change to Aria's code. How the test works.
How the test works
We take the sample profiles and briefs from Aria’s ranking tests, from Python engineers to buyers, quantity surveyors and document controllers. Each profile is rated against each brief through the same code Aria’s main search runs.
Then we rate each profile again with its name swapped for 15 other names from languages and origins found in South Africa, with the name removed, and with the name left blank. We also rewrite the summary with she, he and they, and compare those versions with each other.
Each time we compare the tier, the rating, the score, every item on the card’s checklist and the order of the shortlist. If any of them moves, the test fails and so does the build.
The test counts nobody’s demographics. The names carry no labels, and nothing in it records or guesses anyone’s race, gender or origin. It covers Aria’s main search, and not the two model-rated features named above.