Designing an AI interviewer that cannot lead the witness

1The problem: interviews contaminate evidence

Human memory is not a recording. Loftus and Palmer showed in 1974 that changing a single verb in a question ("how fast were the cars going when they smashed into each other" versus "when they hit each other") changed both the speed witnesses estimated and what they claimed to have seen a week later. The misinformation effect this exposed has been replicated for fifty years: details embedded in questions migrate into memory, and once there, they are indistinguishable from the original event.

Police interviewing has responded with structured protocols. The NICHD Investigative Interview Protocol and the Cognitive Interview both work by front-loading free recall, delaying focused questions, and forbidding suggestion. They demonstrably improve the quality of testimony. The difficulty is adherence: protocol discipline decays under time pressure, case familiarity, and fatigue. A human interviewer who already believes they know what happened finds it genuinely hard not to ask about the red car.

Themis starts from a simple observation: a machine has no beliefs about the case to leak. If the software is built so that suggestive questions are structurally difficult to produce, protocol adherence stops being a matter of training and starts being a property of the system.

2Interview structure

Every Themis interview runs through six phases in a fixed order: introduction and ground rules, narrative practice on a neutral topic, sensory context reinstatement, free narrative account, clarification, and closing. Each phase has a minimum and a maximum number of exchanges, enforced by the engine rather than suggested to the model. The model may propose advancing early; the engine refuses until the minimum is met. The model may want to linger; the engine forces progression at the maximum.

Two rules matter more than the rest. Free narrative always precedes probing, and probes may only reference details the witness has personally raised. The second rule is enforced twice: once in the interviewer\'s instructions, and again by an independent screening model that sees only what the witness actually said and rejects questions that import anything else.

3The turn engine

Themis is not a chat loop. Each conversational turn is a decision procedure with five stages, and the model\'s output is a structured decision, not free text.

When the witness finishes speaking, the transcribed answer is assessed in a single structured call that returns four things: a safety read (distress level, disclosure of ongoing danger, self-harm risk), an intent classification (a substantive answer, a request for a break or a repeat, a request to stop, a question about the process), a chosen action, and a one-sentence internal observation that is stored for the case officer but never spoken.

The action space is small and closed: ask the next question, advance to the next phase, offer a break, ask the witness to confirm they want to stop, end the interview, or escalate for safety. The engine then validates the chosen action against the phase rules, so a model error cannot skip free recall or trap a witness in an endless phase. Two degenerate behaviours discovered in testing are handled in code: a question identical to the previous one is treated as evidence the phase is exhausted, and a screening rewrite that arrives wrapped in meta-instructions is stripped down to the question itself before it can be spoken.

Voice, and why the microphone is managed

The interview is spoken. Questions are delivered by speech synthesis and answers are transcribed live. The microphone is closed the entire time the system is speaking and reopens only after synthesis ends, so the interviewer\'s own voice can never be transcribed as the witness\'s answer. Witnesses control the room by voice: asking for the question again is handled instantly on the device, and requests for a break or to stop are understood by the engine in whatever words the witness uses. On browsers without speech recognition the interview degrades to spoken questions with typed answers rather than refusing to run.

4Question screening

Before any generated question is spoken, a second model with no stake in the conversation examines it against ten categories drawn from the suggestibility literature. The screening model sees the candidate question and the witness\'s own statements, nothing else. If the question fails, the screener proposes a rewrite, which is checked again; after two failed rewrites a neutral open prompt is used instead. The categories:

Screening is defence in depth, not the only defence. The interviewer\'s own instructions forbid the same categories, and the phase structure removes most opportunities to violate them. In testing, the most common screening intervention is softening a focused probe that arrived one phase too early.

5The safety model

An interview with no human present needs an explicit answer to a hard question: what happens when the witness is in trouble? Themis treats this as a first-class part of every turn rather than an exception path.

Every safety read distinguishes distress from danger. Elevated distress prompts the interviewer to offer a pause. Critical distress, a disclosure of ongoing danger, or any indication of self-harm risk stops the interview immediately: the witness sees emergency contact information and a plain statement that help is being arranged, the recording is preserved, and the case officer\'s dashboard is flagged within seconds. Escalation is not reversible from inside the interview; only the officer can decide what happens next.

One asymmetry is deliberate. A calm, clear request to stop is honoured as a right, not treated as an emergency. The witness is asked once, without pressure, whether they want to end; a second confirmation ends the interview with a partial record preserved. Early testing showed language models tend to label any stop request as critical distress, so the engine explicitly refuses to escalate on distress signals alone when the witness is simply exercising the right to leave.

6What the analysis does and does not claim

After an interview ends, the transcript is analysed for eight linguistic indicator types: hedging concentration, over- and under-specificity, internal timeline consistency, formulaic repetition, emotional mismatch, spontaneous self-correction, and unprompted peripheral detail. The last two are positive indicators and are always reported alongside concerns. When a case holds more than one interview, statements are compared across sessions and contradictions are quoted verbatim with topic and severity.

Everything the analysis produces is labelled a computational indicator. The system makes no claim about truthfulness, produces no credibility score of a person, and its reports state on their face that professional review is required. This is not a hedge. Linguistic patterns have documented associations with recall quality, but they are weak, contested signals, and treating them as verdicts would repeat the polygraph mistake with better typography.

7Limitations

Themis inherits the limits of its parts. Speech recognition degrades with accent distance, background noise, and code-switching, and a transcription error can survive into analysis; the recording exists partly so errors can be caught. The language model can misclassify intent or produce a clumsy question; screening and phase enforcement bound the damage but do not eliminate it. Urdu support relies on browser speech engines that vary in quality across devices, which is why every language path keeps a typed fallback.

There are also limits we chose. The system interviews one witness at a time, does not interview suspects, does not analyse video or voice affect, and does not integrate with case-management systems during the pilot. Each of these is a scoping decision, revisited deliberately rather than drifted into.

8References

• Fisher, R. P., & Geiselman, R. E. (1992). Memory-enhancing techniques for investigative interviewing: The cognitive interview. Charles C Thomas.
• Lamb, M. E., Orbach, Y., Hershkowitz, I., Esplin, P. W., & Horowitz, D. (2007). A structured forensic interview protocol improves the quality and informativeness of investigative interviews with children. Child Abuse & Neglect, 31(11-12), 1201-1231.
• Loftus, E. F., & Palmer, J. C. (1974). Reconstruction of automobile destruction: An example of the interaction between language and memory. Journal of Verbal Learning and Verbal Behavior, 13(5), 585-589.