Andrii Bidochko

• Updated: July 1, 2026
• 7 min read

Using Biometrics to Understand AI-Assisted Coding Performance and its Perception

Direct Answer

The paper Using Biometrics to Understand AI‑Assisted Coding Performance and its Perception presents the first large‑scale, multisite biometric study that compares developers’ cognitive and physiological states when they code with a generative AI assistant versus when they code solo. It shows that AI‑assisted programming triggers distinct neuro‑physiological patterns—lower EEG θ/α ratios and higher blink rates—indicating a shift in mental workload, and that these shifts are not simply a matter of speed or experience.

Background: Why This Problem Is Hard

AI‑driven code assistants such as GitHub Copilot, Tabnine, and emerging large‑language‑model (LLM) copilots promise to accelerate software development. Yet, the industry lacks concrete evidence about how these tools reshape the developer’s cognitive workflow. Existing research typically relies on self‑reported surveys or performance metrics (e.g., time‑to‑completion), which capture only the surface level of productivity.

Two fundamental bottlenecks have limited progress:

• Invisible mental load: Developers may offload syntactic generation to the model while still expending effort on verification, debugging, and integration. Traditional metrics cannot differentiate these hidden processes.
• Lack of physiological grounding: Without objective biometric data—EEG, eye‑tracking, electrodermal activity (EDA), heart‑rate variability (HRV)—researchers cannot map the subjective experience of “cognitive ease” to measurable brain and body signals.

Understanding these hidden dimensions is critical for enterprises that plan to embed AI assistants into high‑stakes development pipelines, where safety, code quality, and developer well‑being are non‑negotiable.

What the Researchers Propose

The authors introduce a comprehensive, within‑subjects crossover framework that captures a suite of biometric signals while participants solve realistic programming tasks under two conditions: with AI assistance and without. The core idea is to treat the AI assistant as a “cognitive partner” and to observe how the partnership reshapes neuro‑physiological markers of attention, workload, and stress.

Key components of the proposed framework include:

• Multimodal biometric suite: Simultaneous recording of EEG (θ/α ratio), eye‑tracking (blink and gaze metrics), EDA (skin conductance), and HRV (autonomic balance).
• Performance rubric: A task‑specific scoring system that evaluates code correctness, completeness, and style, providing an objective ground truth.
• Subjective workload assessment: The NASA‑TLX questionnaire administered after each task to capture perceived mental, physical, and temporal demands.
• Crossover design: Each participant experiences both conditions in a counterbalanced order, eliminating between‑subject variability.

How It Works in Practice

In a typical session, a developer sits at a workstation equipped with a 32‑channel EEG cap, a Tobii eye‑tracker, a skin‑conductance sensor, and a chest‑strap HRV monitor. The workflow proceeds as follows:

1. Task briefing: The researcher presents a coding problem (e.g., implementing a sorting algorithm) and records the baseline biometric state for 30 seconds.
2. Condition activation: The participant either enables the AI assistant (which suggests code snippets in real time) or works without any assistance.
3. Live coding: As the developer writes code, the system continuously streams biometric data to a synchronized logging server.
4. Performance capture: Upon completion, the code is evaluated against the rubric, and the participant fills out the NASA‑TLX.
5. Switch & repeat: After a short break, the participant repeats the process under the opposite condition, ensuring that each individual serves as their own control.

This design isolates the effect of AI assistance from individual skill differences, allowing the researchers to attribute observed physiological changes directly to the presence of the code assistant.

Evaluation & Results

The study recruited 48 participants across two universities (Bari, Italy, and Copenhagen, Denmark), split evenly between undergraduate and graduate computer‑science students. The evaluation focused on four hypotheses:

1. Physiological differences between AI‑assisted and solo coding

• EEG θ/α ratio: Significantly lower during the first AI‑assisted task, suggesting reduced frontal‑midline theta activity associated with sustained attention.
• Gaze blink rate: Higher during the second AI‑assisted task, indicating a relaxation of visual focus when the model supplies code suggestions.

2. Moderating role of developer experience

Both undergraduate and graduate participants exhibited the same biometric trends; experience level did not moderate the effect.

3. Association between physiology and performance

• In the non‑AI condition, higher EDA correlated with better rubric scores, implying that increased arousal supported deeper problem solving.
• Under AI assistance, this correlation vanished, suggesting that the assistant decouples physiological arousal from performance outcomes.

4. Alignment between subjective perception and objective measures

Only the “Physical demand” dimension of NASA‑TLX predicted performance in the solo condition; none of the self‑reported dimensions aligned with performance when AI assistance was present.

Collectively, these findings demonstrate that AI‑assisted programming is not merely a faster version of solo coding. Instead, it creates a cognitively distinct mode where developers offload generative effort, leading to measurable shifts in brain rhythms and eye behavior.

Why This Matters for AI Systems and Agents

For practitioners building AI‑driven development tools, the study offers three actionable insights:

• Design for cognitive ergonomics: Assistants should surface suggestions in a way that minimizes unnecessary visual scanning, perhaps by integrating inline completions directly into the IDE’s code view.
• Personalized assistance strategies: Since experience level did not affect biometric responses, adaptive models could rely on real‑time physiological feedback (e.g., blink rate) to modulate suggestion frequency, reducing overload for any user.
• Biometric‑aware evaluation: Traditional benchmarks (e.g., pass@k) miss the nuanced impact on developer mental state. Incorporating biometric metrics into product testing can reveal hidden trade‑offs between speed and cognitive strain.

These considerations align with emerging enterprise AI platforms that aim to embed intelligent agents into complex workflows. For example, the Enterprise AI platform by UBOS can leverage biometric signals to trigger context‑aware assistance, ensuring that AI agents augment rather than overwhelm developers.

Moreover, the findings support the integration of AI assistants into broader automation pipelines. The Workflow automation studio could orchestrate code generation, testing, and deployment while monitoring developer stress levels, automatically pausing or escalating tasks when physiological markers indicate fatigue.

What Comes Next

While the study establishes a solid baseline, several limitations open avenues for future research:

• Real‑world development environments: The experimental tasks were constrained to short coding problems. Extending the protocol to large‑scale, multi‑module projects would test whether the observed biometric patterns scale.
• Longitudinal monitoring: Repeated exposure to AI assistance over weeks or months could reveal adaptation effects, such as habituation of EEG signatures or changes in perceived workload.
• Closed‑loop assistance: Integrating real‑time biometric feedback into the assistant itself—e.g., reducing suggestion density when blink rate spikes—remains an open engineering challenge.

Addressing these gaps could enable a new class of “cognitively aware” development tools that dynamically adjust their behavior based on the developer’s mental state. The AI marketing agents showcase how such feedback loops can be applied beyond coding, suggesting a broader ecosystem where biometric data informs AI behavior across domains.

For startups eager to experiment with biometric‑enhanced AI assistants, the UBOS for startups page offers resources and sandbox environments to prototype such integrations. Meanwhile, SMBs can explore the UBOS solutions for SMBs to embed lightweight monitoring into existing development pipelines without heavy infrastructure overhead.

In summary, the research marks a pivotal step toward understanding the hidden cognitive dimensions of AI‑augmented software engineering. By marrying biometric science with AI assistant design, developers and enterprises can build tools that are not only faster but also healthier and more aligned with human cognition.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.