Skip to content

Trust center

How the AI interview and scoring work

Last updated: November 2026

This page explains, in detail, what the AI does in an Aethel interview, how your answers are scored, what is never scored, and where it falls short.

The short version

  • An AI asks the questions. It is labeled "AI interviewer" on screen.
  • Only the text of your answers is scored. Not your voice, face, video, or typing.
  • Each field has four areas. An area gets a score only if a quote from your own answers supports it.
  • People at the company make the hiring decision. Aethel never rejects anyone automatically.
  • Your results show how long the live interview took and how many times you left the room and came back. A company that gets the interview through its link sees the same facts. They are not scored.
  • You can ask for an accommodation or another way to be assessed, and you can delete your interview at any time.
  • No independent bias audit has been done yet.

The interviewer

The interviewer is an AI. It appears as a video avatar from our video provider, currently Beyond Presence, and speaks its questions out loud. The interview screen labels it "AI interviewer", and it says it is an AI at the start and whenever you ask. The questions are written by an OpenAI model. If OpenAI is slow to start the interviewer’s next reply, we also send the same request to the same OpenAI model run by Microsoft Azure, through LiveKit, and use whichever answers first. This applies only to interviews you agreed to from October 20, 2026, when the consent screen started naming Microsoft Azure. Microsoft Azure gets only text, never your audio or video. If the interviewer's video cannot start or stops, the interview goes on by voice and text. The backup interview, used only when the live interview room cannot start, has no video: the questions are shown on screen. A company can also choose a voice interview for its interview link: the same AI interviewer by voice, with the questions on screen and no video of it, and the link's page and the consent screen say so before you start.

To write each question, the AI is given:

  • Details from your profile: your field, specialty, years of experience, skills, the summary you wrote, your most recent job, and the project or result you chose from your resume. The backup interview also gets your other jobs and the education you listed.
  • Your recent answers in this interview.
  • If you came from a company's interview link: the role details the company wrote, and any questions it asked us to include.

It is given your first name so it can greet you. It is not given your last name, email address, photo, or anything about you from outside Aethel.

In the live interview room, your microphone turns on by itself when it is your turn to answer, unless you chose Type only, and you can mute it at any time. While it is on, your microphone audio travels through LiveKit to Deepgram, which turns it into text. LiveKit's turn detection also uses the audio to tell when you have finished speaking, so the interviewer does not cut you off. Neither is scored. Aethel keeps only the text and does not save audio recordings. In the backup interview, your browser turns your speech into text instead; in some browsers, such as Chrome, the browser maker processes that audio to do this. You can always type instead.

What the interviewer can and cannot do

It can

  • Ask about real work you have done, one question at a time.
  • Follow up on details from your earlier answers.
  • Ask a question again in simpler words if you get stuck. It never suggests an answer.
  • Ask questions a company added to its interview link.
  • Listen while your microphone is on. It turns on by itself when it is your turn, and you can mute it. What you say is turned into text as you speak, and the interviewer replies to that text. Your audio is not recorded or saved.

It cannot

  • See you. Your camera, if you turn it on, stays in your browser.
  • Judge your voice, accent, or tone. The AI that writes its questions receives only the text of what you said.
  • See your screen, your other tabs, or your keystrokes.
  • Look you up online or see other candidates.
  • Make a hiring decision, or reject or advance you.

The four areas for each field

Every interview is scored on four areas. The areas depend on your field, so a lawyer is judged on legal work and a designer on design work. These are the areas today:

  • Software engineering

    The four areas
    Problem solving, Architecture, Communication, Collaboration
  • Data and machine learning

    The four areas
    Problem solving, Modeling and analysis, Communication, Collaboration
  • Design

    The four areas
    Design problem solving, Craft and design decisions, Communication, Collaboration
  • Product management

    The four areas
    Product judgment, Prioritization and strategy, Communication, Stakeholder collaboration
  • Finance and accounting

    The four areas
    Financial analysis, Accuracy and controls, Communication, Stakeholder judgment
  • Law and compliance

    The four areas
    Legal reasoning, Issue spotting, Communication, Client and stakeholder judgment
  • Healthcare and life sciences

    The four areas
    Clinical or scientific reasoning, Safety and standards, Communication, Patient and team collaboration
  • Healthcare support and caregiving

    The four areas
    Noticing and responding, Safety and procedures, Communication, Care team teamwork
  • Science and research

    The four areas
    Scientific reasoning, Research design and rigor, Communication, Collaboration
  • Writing, editing and languages

    The four areas
    Editorial judgment, Craft and accuracy, Communication, Collaboration
  • Education and training

    The four areas
    Teaching problem solving, Learning design, Communication, Learner and team collaboration
  • Childcare and early education

    The four areas
    Problem solving, Safety and routines, Communication, Working with families and staff
  • Sales and customer success

    The four areas
    Problem solving, Customer understanding, Communication, Relationship judgment
  • Customer service and support

    The four areas
    Problem solving, Service judgment, Communication, Teamwork
  • Retail, food and hospitality

    The four areas
    Problem solving, Service and standards, Communication, Teamwork
  • Administration and office support

    The four areas
    Problem solving, Organization and accuracy, Communication, Working with people
  • Government and public service

    The four areas
    Problem solving, Rules and fairness, Communication, Working with the public and agencies
  • Nonprofit, church and community work

    The four areas
    Problem solving, Programs and stewardship, Communication, Community and volunteers
  • Skilled trades and facilities

    The four areas
    Troubleshooting, Safety and quality of work, Communication, Crew teamwork
  • Operations and other fields

    The four areas
    Problem solving, Planning and execution, Communication, Collaboration

Two rules apply in every field. Saying "we" or naming your team is not, by itself, evidence of collaboration: your words need to show you working with someone, such as agreeing on a plan, settling a disagreement, or helping a teammate. And answering clearly counts toward communication, but only up to 80; a higher communication score needs you to describe explaining, writing, or presenting to people. You need to describe what you actually did.

The evidence rule

After the interview, an OpenAI model reads the transcript and scores it. Then our own server checks the work.

  1. The scoring step receives only your transcript and your field. It does not receive your profile, your resume, or anything the company wrote. Your first name can appear in the transcript, because the interviewer greets you by name.
  2. For each area, the AI must copy quotes from your answers. The interviewer's questions never count as evidence.
  3. Our server checks every quote. The words must appear, in the same order, in one of your answers, the quote must be at least 8 words long, and it must show that area's skill, not just mention it. The AI marks each quote as showing the skill or only giving context, and context does not count. For collaboration, a sentence of the quote must show you working with someone: naming a team, its size, or what other people did is not enough. Writing data, such as rows into a table, is not communication. Quotes that fail are thrown out.
  4. An area with no quote left is marked "not enough evidence". It does not get a low score.
  5. The overall score is the average of the areas that have evidence, and only exists when at least two areas do. Any overall number the AI suggests is ignored.
  6. The written summary can only describe what you explained. The AI is told not to state a level, recommend hiring, mention pay, or praise beyond the evidence. Our server then removes any sentence that uses a level word such as "senior", recommends hiring, or uses a common praise phrase such as "deep expertise". This is a fixed word check, not a reading of the evidence. Any quote left in the summary must be your own words. Level, pay, and suggested-role fields are always dropped.
  7. Some replies never count as answers: short replies such as "Yes, I'm ready", "OK" or "Can you repeat that?", and any reply under 8 words. With fewer than 3 real answers, or under 60 words of answers in all, nothing is scored, and the result says so. That is not a low score.
  8. Each answer is checked against the question it answers. An answer about something else gets no credit, and the result names the question in a short note. This check is done by the AI, so it can miss. An answer the AI leaves out of its check is checked a second time on its own.
  9. For each area, the AI rates three things in your answers. The situation: no specific case, a case without its context, or a specific case with where it was and how big. Your reasoning: none, a reason for what you did, one option you weighed against another, or several options compared. The result: none or vague, approximate, measured or confirmed by someone, or followed up over time or used by other people. The ratings give 0 to 8 points in all, and the points set the area's score on a fixed scale that is the same for everyone: 8 points is 95, 7 is 91, 6 is 87, 5 is 80, 4 is 76, 3 is 72, 2 is 66, 1 is 60, and 0 is 50. Without a specific case, named with where it was and how big, an area is at most 60, whatever the other two ratings are.
  10. These ratings are made in a separate step that never sees how you write or speak. First, an AI copy-edits your answers into standard English grammar, keeping every fact. In a Spanish interview, it translates them into English, keeping every fact. Then an AI rates the copy-edited answers 14 times, and each area's score is the average of those 14 ratings. The copy-edited text is used only for this. It is not saved, not shown to anyone, and never quoted as your words: every quote is your own words. If this step fails, the ratings of the AI that picked the quotes are used.
  11. Each rating also gives the AI's own number for the area. When that number is more than 5 below the score the scale gives, that rating counts as the lower number, because the AI judged the answers weaker than the three things show. A rating never counts above the scale, so small differences in how polished an answer sounds do not decide the score.
  12. A score is only as strong as the quotes shown with it. An area can score 90 or more only when one of its quotes has at least 20 words and does not break off mid-phrase (it ends at a full stop, comma, colon, or bracket in your answer). Otherwise the area is at most 89, and when its best quote has fewer than 12 words, at most 74. For collaboration and communication, only the quote's sentences about working with people, or about explaining, writing, or presenting, count toward those words. A clear answer still counts toward communication up to 80.
  13. Communication is judged on whether you answered the question, in a clear order, with a concrete example, and on what you said about explaining things to others. Never on grammar, word choice or how fluent your English is. Communication shown only by how you answered is capped at 80.
  14. Spoken answers are written down by speech-to-text, which can mishear technical words. A misheard word is never shown as yours: the AI picks another quote, or shows the likely word in brackets.
  15. The AI is told to treat your answers as evidence only, and to ignore any instructions written inside them.

For an area that has evidence, the three ratings give these scores on the fixed scale. Each row is one example. With a specific case, other ratings that add up to the same points get the same score.

  • 95

    What the answers for that area showed
    A specific case, several options compared, and a result followed up over time or used by other people.
  • 91

    What the answers for that area showed
    A specific case, several options compared, and a measured or confirmed result.
  • 87

    What the answers for that area showed
    A specific case, one option weighed against another, and a measured or confirmed result.
  • 80

    What the answers for that area showed
    A specific case, a reason for what you did, and a measured or confirmed result.
  • 76

    What the answers for that area showed
    A specific case, a reason for what you did, and an approximate result.
  • 72

    What the answers for that area showed
    A specific case and a reason for what you did, with no clear result.
  • 66

    What the answers for that area showed
    A specific case, with no reason and no clear result.
  • 60

    What the answers for that area showed
    No specific case named with where it was and how big. Without one, this is the most an area gets, however good the reasoning and the result.
  • 50

    What the answers for that area showed
    No specific case, no reason and no clear result.

An area's score is the average of the 14 ratings, so it can fall between these numbers. The AI's own number and the quote rules above can lower it.

What is never scored

Only the text of your answers reaches the scoring step. So these are never scored, and most are never collected at all:

  • Your voice, accent, or tone. We do not record audio.
  • Your face, appearance, or video. We do not record or upload video. Your camera preview stays on your device, so no face recognition or face analysis is possible, and Aethel uses none.
  • Your voice as an identifier. Aethel does not use voice identification or keep voiceprints. Spoken answers are only turned into text.
  • Your background or surroundings on camera.
  • Your typing speed, or how long you take to answer.
  • Your screen, mouse movements, keystrokes, or other tabs. We do not collect them. In the live interview room, Aethel counts how many characters are pasted into the answer box, never what was pasted, so each answer can show as spoken, typed, or pasted. That mark is not part of any score.
  • Your age, gender, race, or other personal characteristics. We do not ask for them.
  • Your grammar, word choice, or how fluent your English is.
  • Your school, past employers, or resume, unless you talk about them in an answer.

How long the interview took, and how many times you left the room and came back, are shown as plain facts but are never scored. See how the interview ran.

How the interview ran

Your results page shows a few plain facts about how a live video interview ran. A company that gets the interview through its own link sees the same facts with your results.

  • Active time: Start to finish, minus any time away.
  • Start to finish: From the first saved question or answer to the last one.
  • Reconnects: How many times you left the interview room and came back after the interview started, for example after a page reload or a dropped connection, and how long those breaks took in total. Each break counts from the last saved question or answer before it to the first one after it.
  • Interviewer video: Shown only when the AI interviewer's video stopped during the interview (how many times, and whether the interview went on by voice only), or when the interview ran by voice only from the start: "Voice only, video was not available" when Aethel had no video to give, for example because the month's interviewer video was used up, or "Voice only (voice interview link)" when the link was set to voice. Video that dropped or was not available is a problem on Aethel's side, not something you did.

They come from what our server already records: when each interview room was opened, when each question and answer was saved, and whether the AI interviewer's video dropped. Nothing about your screen, keystrokes, mouse, or other tabs is used. These facts are not part of any score. A reconnect is often just a dropped connection. So it is shown as a plain fact, not as a warning.

Older results, and backup interviews, do not show these facts. A recent interview you share with "Use my recent interview" is shared without them.

The human decision

The AI review points a company's team to the evidence in an interview. It does not decide anything. Aethel never rejects or advances a candidate on its own. The company's team reads the transcript and the quotes, sets its own decision by hand, and contacts candidates directly.

Ask for an accommodation or another way to be assessed

Need an accommodation, or another way to be assessed? Before an interview in our live interview room, you can choose more time per answer, more time overall, type only, or larger captions on the consent screen, without asking anyone. For anything else, tell us before you start. We reply within 2 business days. Replies come from Aethel's AI support assistant. If a company sent you the interview link, that company decides whether to offer another way, so ask them directly too.

Some things are built in for everyone: you can type every answer, the camera and microphone are optional, and there is no countdown on your answers. In the live interview room, a spoken answer ends when you pause, and you can choose more time per answer on the consent screen. For anything else, send us a message or email support@aethel.pro. You do not need to share a diagnosis, and your request is not part of your results.

Delete your interview

You can delete your interview and your account at any time from your profile page, and it happens right away. If you ask us by email, we delete it within 30 days.

Deleting removes your transcripts, scores, and consent records, and a company you interviewed with through its link loses access to them in Aethel. A company that already downloaded your results, or received them in its own systems, keeps that copy until it deletes it, so ask the company too. See what gets deleted.

Known limitations

  • The AI can misread an answer or miss context. That is why every score shows its quotes, and why people decide.
  • Speech recognition can get words wrong, especially with background noise, some accents, or technical terms. In the live video room a spoken answer is sent when you pause or press "Done answering", so if something was misheard, correct it in your next answer, or type instead.
  • Interviews run in English or Spanish. For Spanish, the candidate chooses "Interview in Spanish" on the consent screen, just before the interview starts. In a Spanish interview, an AI translates each answer into English before the step that rates it. The quotes behind each score stay in the candidate's own Spanish words. After a Spanish interview, the candidate's results page is in Spanish, and the candidate can switch it to English. Its PDF is in Spanish too, and the page they share by link opens in Spanish, with the same switch. The written feedback is in Spanish; the summary from the AI review is in English. The debrief, the voice practice, the company's own view of the result and the verification page are in English. The free practice is also in Spanish. The backup text interview stays in English. The short add-on at a company's link is in Spanish when the candidate reads that link's page in Spanish, and in English otherwise. Spanish scoring has not been in a fairness test yet. In the live video room, speech recognition is set to the interview's language, English or Spanish. In the backup interview, your browser's speech recognition is set to US English. The check that a quote is about an area looks for English words; for a Spanish quote it reads an English translation of it, and the quote shown stays in your words.
  • The check that a quote is "about" an area, and that a collaboration quote shows you working with someone, uses word lists for each area. A real example described in unusual words can be missed and marked "not enough evidence", or scored lower. The AI's own judgment of whether a quote shows the skill can also be wrong.
  • The AI is not perfectly consistent. The same answers scored again can get a different number. In our latest test of 69 made-up interviews, each scored 3 times, 9 in 10 stayed within 3 points. The largest difference, 6 points, was for answers with no real examples.
  • An interview of about 15 minutes covers only part of what you can do. Results describe what you said in that interview. We have not tested whether they predict job performance.

How we test it

Our code includes automated tests that check the scoring rules above. They check that:

  • Quotes must be your exact words. A made-up or reworded quote is rejected, and its area is left unscored.
  • An area without a valid quote shows "not enough evidence", never a low score.
  • A quote must show its area's skill: fragments under 8 words, quotes the AI marks as context, collaboration quotes that only name a team, and data written to a table are thrown out.
  • No area scores 90 or more without a complete quote of at least 20 words about that area, and no area scores above 95.
  • Without a specific case, named with where it was and how big, an area scores at most 60.
  • The overall score is the average of the areas with evidence, and needs at least 2 of them. Any overall number the AI suggests is ignored.
  • Sentences that use a level word (such as "senior"), recommend hiring, or use common praise phrases (such as "deep expertise") are removed from the summary. This is a fixed word check, not a reading of the evidence.
  • Each field gets its own question and scoring instructions, and the list of fields is the same on our website and our server.
  • A company can only open results from its own interview links, never another company's.
  • Our deletion code covers your account, profile, and interview records, and new interview records carry a deletion date.

These tests use fixed, made-up AI responses, so they check our rules, not the live AI's judgment. They do not yet run automatically on every change; today we run them by hand. None of them measures whether results differ between groups of people.

Bias audit

No independent bias audit has been done.

Aethel runs its own tests on made-up interviews. They are not an audit, and they do not measure outcomes for real applicants.

  • On October 5, 2026, with the scoring used then (version 2026-10-05-f1): 798 scorings of made-up candidates, each changing one thing that should not matter. A name, a stated disability, a hint about age, and speech-to-text mistakes made no difference we could measure. Non-native English grammar lowered strong and average answers by about 2 points out of 100, and a follow-up test found up to 7 points in some fields. We then changed the scoring so answers are copy-edited into standard English before they are rated.
  • On October 21, 2026, with the scoring used now (version 2026-10-21-d4): one strong made-up interview in each of 8 of our 20 fields, once with non-native grammar and once in fluent English, each scored 3 times. The two came within about 1 point of each other in every one of those fields. This smaller test did not cover average answers or the other 12 fields.
  • On October 27, 2026 we repeated the October 5, 2026 test on the scoring used now (version 2026-10-21-d4), with 422 scorings of the same made-up candidates. On average, a name, a stated disability, a hint about age or years of experience, non-native English grammar, and speech-to-text mistakes each moved the overall score by less than 2 points out of 100. One exception, by answer level: speech-to-text mistakes lowered the overall score of weak answers by 2.1 points out of 100 on average (95% interval: 0.2 to 3.9 points lower). For scale, two scorings of the very same interview differed by 0.7 points on average. The 8 everyday-work fields added on October 27, 2026 and interviews in Spanish were not part of it.
  • Not tested yet: the 8 everyday-work fields added on October 27, 2026, interviews in Spanish, and results from real candidates.

We cannot yet measure differences between groups of real candidates, because we do not ask for age, gender, race, or other personal characteristics.

To check whether results differ between groups of real candidates, candidates would first need a way to tell us about themselves if they choose, kept apart from scoring and never shown to companies. We have not built that, so no such check has started. When one has, we will publish what we find on this page.

Some places, such as New York City, Illinois, Colorado, California, and the European Union, have rules for AI in hiring. New York City's Local Law 144, for example, asks an employer to have an independent bias audit before it uses a tool like this to substantially help decide who moves forward. Aethel does not have one, and we tell companies so before they use it. Companies that use Aethel are responsible for the notices and audits the laws where they hire require. Read how Aethel supports AI hiring laws, law by law, and what we tell companies, or email support@aethel.pro.