Skip to main content
. 2025 Oct 2;25(19):6086. doi: 10.3390/s25196086
Algorithm 1 Prompt -level multimodal deception detection (zero-shot GPT-5).
  • Require: 

    Dataset D of videos with ground-truth labels y∈{lie,truth}

  • Require: 

    Ablation flags: UseVideo, UseTranscript, UseEmotion

  •     1:

    Initialize metrics containers

  •     2:

    for each sample x∈D do

  •     3:

        I←⌀,T←⌀,E←⌀

  •     4:

        if UseVideo then

  •     5:

            Extract 16 uniformly spaced frames I={f0,…,f15}

  •     6:

         end if

  •     7:

         if UseTranscript then

  •     8:

               Extract audio; obtain ASR transcript T (Whisper-1)

  •     9:

         end if

  •   10:

        if UseEmotion then

  •   11:

               Extract audio; compute emotion label e and confidence ce (SpeechBrain wav2vec2)

  •   12:

               E← “Detected emotion: e (ce)”

  •   13:

        end if

  •   14:

        Build user prompt: include E (if any), T (if any), instruction to return strictly JSON

  •   15:

        Attach frames I (if any) as images to the same message

  •   16:

        System message: safety + research framing

  •   17:

        Query GPT-5 with deterministic decoding

  •   18:

        Parse first valid JSON object: {label, confidence, reasoning}

  •   19:

        y^←label                                                                        ▹ final class: lie or truth

  •   20:

        q←confidence∈[0,1]                             ▹ used only for analysis/threshold sweeps

  •   21:

        Store (y^,q) and compare with ground truth y

  •   22:

    end for

  •   23:

    Compute Accuracy, Precision/Recall/F1 per class, Macro-F1, MCC, Cohen’s κ