Overview

A common concern from lenders and compliance teams is whether an applicant could use an AI tool, such as ChatGPT or a similar assistant, to complete the Begini psychometric assessment on their behalf, producing a score that does not reflect their genuine behavioural profile.

Begini's position is that the platform is AI-resilient by design. The goal is not to claim perfect prevention, which is neither achievable nor verifiable in an unsupervised online environment, but to make AI assistance hard to apply, easy to detect, and operationally manageable when suspected.

Key Principle: Because there is no 'correct answer' to optimise for in a psychometric game, AI assistance does not improve the score. It produces a signal that looks different from genuine human behaviour, and that difference is what Begini detects.

Layer 1 — Assessment Design

The primary defence against AI-assisted completion is the nature of the assessment itself. Begini games are designed around live behavioural interaction, not static question-and-answer formats. This reduces the value of AI assistance before any technical controls are applied.

Why game-based assessment is structurally resistant

• AI tools work by generating text responses to text prompts. Begini games require real-time motor input, taps, swipes, drags, holds, and trajectory-based decisions that cannot be proxied through a language model.
• There is no correct answer to optimise for. The psychometric signal comes from how the applicant behaves across the session, not from any individual choice. An AI trying to 'pass' has no target to aim at.
• Time-pressured games compress decision windows to a duration that is incompatible with the latency of prompt-and-respond AI loops.
• Stimulus randomisation means screenshots, shared prompts, and pre-cached strategies are not reusable across sessions.

Layer 2 — Behavioural Biometrics

Even where AI assistance could theoretically be applied, Begini captures interaction metadata at the event level. This is data that AI cannot easily reproduce and that human-assisted or scripted completion will inevitably distort.

Signals captured

• Tap and click timing, including decision latency per game and per task type
• Touch pressure, gesture arc, and hold duration (mobile SDK)
• Micro-corrections and false starts before committing to a response
• Inter-event interval variance — the natural irregularity of human pacing
• Navigation behaviour, including tab switches, pauses, and restarts

Detection Logic: AI-assisted or scripted sessions show statistically anomalous patterns: response timing that is too uniform, implausible completion speed, absence of micro-hesitation, or reduced variance where variance should naturally exist. Begini monitors for these patterns without surfacing detection logic to applicants.

Layer 3 — Session Integrity Controls

Session-level controls add a further layer of protection by making the assessment environment harder to exploit.

• Single-use session tokens with expiry windows prevent replay or redistribution of assessment links
• Single active session enforcement — the assessment cannot be opened across multiple devices or windows simultaneously
• Repeat-attempt monitoring — multiple restarts or unusual completion patterns are logged as signals, not ignored
• Device binding — session characteristics are logged and changes mid-session are flagged
• Browser environment checks — certain remote desktop and virtual environment indicators can be detected via available browser APIs

Layer 4 — The Trust Score

The cleanest commercial answer to the AI-assistance question is Begini's session confidence layer. Rather than treating every completed assessment as equally valid, each session carries a trust and confidence score alongside the psychometric output.

• A low-confidence session can be flagged for manual review rather than accepted into automated decisioning
• Lenders can configure thresholds at which step-up verification is triggered automatically
• This turns the question from 'did AI assist?' to 'how much weight should we place on this session?' and is a more operationally tractable framing

Next steps

• Fraud Detection & Trust Score — how Begini evaluates session reliability and flags suspicious behaviour
• KYC and Identity Binding — how Begini links the assessment to a verified loan applicant
• Understanding Assessment Results — how to interpret and act on score and trust outputs