AI Session Annotator (English)

About the Role

Mercor is hiring English-speaking reviewers to evaluate recorded sessions with a voice-and-camera AI assistant. You will watch each session, compare what the assistant said against what was actually visible in the video, and score it against a detailed rubric.

Your written comments are the deliverable. What you flag as a failure is what gets fixed first. This is judgment work, not volume work.

Key Responsibilities

• Review recorded sessions in which a user speaks to an AI assistant while their camera streams whatever is in front of them.

• Compare the assistant's responses against what was actually visible in the video feed, and identify hallucinated or fabricated visual detail.

• Score each session across visual interpretation, accuracy, relevance, response timing, and whether the assistant used spatial and directional language ("to your left", "move your finger up") rather than purely visual description.

• Judge whether the assistant appropriately warned the user when its answer could affect their health, physical safety, or financial security, and whether that warning was timely and proportionate.

• Write detailed open comments that support every score with concrete evidence from the session.

• Capture the conversation turn by turn in the intake form.

Required Qualifications

• Native or native-level English, spoken and written.

• Clear, precise written English; the scoring guide is written in English.

• An Android phone with a working camera and microphone. This is mandatory for the project.

• Attention to detail and the discipline to apply a rubric consistently, without substituting your own criteria along the way.

Preferred Qualifications

• Prior data annotation, linguistic QA, content moderation, or AI model evaluation experience.

• Training in linguistics, translation or interpreting,

• An ear for regional register: recognizing when the assistant's English sounds unnatural or non-native.

Additional Information

• Start date: Immediate. High volume of work between 26 to 31 Aug is expected.

• Content note: some sessions involve situations where a wrong answer would carry real consequences, such as identifying medication, checking whether food has spoiled, crossing a street, or reading bank card details. There is no graphic or violent content, but you will evaluate cases with safety implications, and part of the work is flagging when the assistant failed to warn about the risk.

Application Process

• Submit your resume or relevant background to get started.

• Qualified applicants may be asked to complete a short calibration task on sample sessions, assessed on scoring consistency and the usefulness of written comments. We are not assessing speed.