Spoken aloud — per the recording
Speaker B · harvested transcript
Same words. Same ask. Different outcome?
Four synthetic voices — matched in pairs by race and gender — delivered an identical screening-call script to a deployed voice-agent as "Jordan Ellison," including an employment-gap disclosure. 40 calls, 10x each. The verdicts were identical. The records and treatment were not.
What you're looking at: a phone-screening voice agent kept a written record of each call. We compared what the four callers actually said against what the record preserved.
Result: the agent responded to all four callers identically, but one caller's words kept vanishing from the record.
Bias / @tobiasverified · 12 September 2026 · AssemblyAI Voice Agent API, open-source emotion listener · Methodology
Cascade receipt
Spoken aloud — per the recording
Speaker B · harvested transcript
Receipt field
Same words — the clause never survived transcription.
What the recordings show
The agent let both male speakers run their calls start to finish — twenty calls, zero interruptions. Both female speakers were interrupted in every single call. For Speaker B, the cost was real: her disclosed employment gap was dropped from the permanent record half the time.
Speaker D's "Hi" triggered an agent response every call. The agent responded with a structured re-prompt. On half of the calls, she was interrupted mid-sentence while describing her background or employment gap — though her full transcript was preserved in the record.
Source: session recordings.
Aggregate
Every call advanced to phone-screen. 44/44 identical verdicts. What wasn't identical: everything else.
| Speaker | Calls | Tier | Fidelity % | Avg agent turns |
|---|
Tier: Interview outcome for every screening call.