High-definition Transcranial Direct Current Stimulation over Right Dorsolateral Prefrontal Cortex to Enhance Metacognitive Sensitivity

Summary

This study proposed an experimental paradigm that applies High-Definition Transcranial Direct Current Stimulation (HD-tDCS) to improve human metacognition in human-AI delegation tasks. The results demonstrated the effectiveness of HD-tDCS in enhancing metacognitive sensitivity.

Abstract

In human-AI collaboration, task delegation is a critical component. Ideally, if a person believes they are capable of completing a task, they should do so themselves; otherwise, the task should be delegated to the other party. Such delegation decisions are influenced by individuals' assessments of their own abilities, which are shaped by metacognition-the capacity to evaluate one's own capacities. High-Definition Transcranial Direct Current Stimulation (HD-tDCS) is a non-invasive brain stimulation technology that can regulate the excitability of the cerebral cortex to improve human performance. This protocol presents a method that employs HD-tDCS to investigate its effects on human metacognition during human-AI delegation tasks, detailing the procedure for applying anodal, cathodal, or sham stimulation to the dorsolateral prefrontal cortex (DLPFC). Stimulation was performed using a portable HD-tDCS device, using a circumferential arrangement of five electrodes, including one target electrode and four return electrodes. The stimulation lasted for a total of 21 min, and the current intensity was 2 mA. This method is expected to provide a new research path for further in-depth exploration of the neural mechanisms and behavioral regulation of humans in the process of human-AI collaborative task delegation and provide a solid theoretical basis for further optimizing human-AI collaborative tasks.

Introduction

Artificial intelligence (AI) is becoming increasingly intertwined with both our everyday lives and the workplace. However, in many real-world contexts, AI is not yet capable of fully automating decision-making processes¹. Certain tasks remain particularly challenging for AI due to current technological limitations, as well as ethical and legal considerations. As a result, granting full autonomy to AI in specific domains is often impractical or inadvisable^2,3. Humans, therefore, continue to play an indispensable role in the workplace, and human-AI collaboration is expected to become a dominant mode of production and service delivery in the future. Research in areas such as healthcare, hospitality, and human resources has shown that collaborative efforts between humans and AI can reduce costs and, in some cases, enhance efficiency^4,5,6.

In human-AI collaboration, each party is expected to undertake tasks that align with its respective strengths, while delegating less suitable tasks to the other. Therefore, effective task delegation is essential for successful collaboration and optimal performance outcomes. Adam et al. found that when the delegation process is fully automated and handled solely by AI without human involvement, humans tend to experience heightened perceptions of self-threat⁷. Fuegener et al. further reported that during the process of delegating tasks to AI, people often struggle to accurately assess their own capabilities or evaluate task difficulty, reflecting a low level of metacognition⁸. These findings highlight the importance of enhancing metacognition in human-AI delegation scenarios to improve the efficiency and quality of collaboration.

Metacognition refers to an individual's ability to evaluate their own capabilities⁹. In the context of mutual delegation between humans and AI, it is essential not only to form an accurate understanding of the other party's capabilities but also to evaluate one's own abilities to determine when delegation is appropriate. Jussupow et al. found that human-AI collaboration is most effective when humans demonstrate high levels of metacognition-that is, when they can accurately assess both their own abilities and those of the AI¹⁰. Similarly, Atchley et al. found that fostering metacognition in educational settings can promote deeper and more effective collaboration between humans and AI systems¹¹. Previous studies on human-AI collaboration have often assessed metacognition through self-report questionnaires or by using human confidence ratings as indirect measures. However, in psychological research, a more established and rigorous measure of metacognition is metacognitive sensitivity (meta-d')¹². In the present study, we use meta-d', following the approach described by Maniscalco et al. to assess metacognitive sensitivity¹³.

Transcranial direct current stimulation (tDCS) has been widely used in research aimed at enhancing metacognition¹⁴. Core aspects of metacognition are closely associated with the prefrontal cortex. tDCS delivers a weak direct current to specific brain regions, modulating neuronal excitability and synaptic plasticity in a non-invasive and non-intrusive manner, thereby allowing researchers to improve various cognitive functions¹⁵. Harty et al. found that anodal stimulation over the right dorsolateral prefrontal cortex (DLPFC) enhanced participants' error detection performance in an error awareness task, thereby improving their metacognitive monitoring of cognitive performance¹⁶. Bona and Silvanto reported that anodal tDCS affected metacognitive bias in a working memory task¹⁷. Mattes et al. showed that while tDCS influenced participants' confidence ratings in both metacognition and mentalizing tasks, it did not alter their metacognitive sensitivity¹⁸. Temporary disruption or damage to the prefrontal cortex-specifically Brodmann areas 46 and 10-in both humans and monkeys has been shown to induce changes in metacognitive processes¹⁴. In addition, previous studies have provided causal evidence linking stimulation of the DLPFC to metacognitive abilities^19,20. The right DLPFC, in particular, plays a key role within the neural network underlying cognitive control^21,22, which is closely related to the demands of our experimental task. Therefore, we selected the right DLPFC as our target site for stimulation. Consistent with established protocols in tDCS studies on metacognition, we applied a 2 mA current for 21 min, including a 1 min ramp-up and ramp-down phase^17,23. Currently, the highest current intensity used in tDCS is 4 mA, which has been shown to be tolerable for participants in the study conducted by Workman et al.²⁴.

High-definition transcranial direct current stimulation (HD-tDCS) is a high-precision modification of conventional tDCS, offering more targeted neuromodulation through specialized electrode montages. Compared to conventional tDCS, which uses two large pad electrodes, HD-tDCS allows for more targeted and focal stimulation of specific brain regions, resulting in more sustained effects^25,26. In contrast to neurofeedback techniques that also aim to enhance cognitive performance²⁷, HD-tDCS offers several practical advantages, including a shorter activation period, simpler application procedures, and lower implementation costs²⁸. From the perspective of human-AI task delegation, HD-tDCS offers higher spatial precision, ease of operation, and cost-effectiveness. Moreover, it allows participants to engage in sustained tasks involving delegation and reflection within contexts that approximate real-life or workplace scenarios, making it especially suitable for probing metacognition in human-AI delegation. In this study, HD-tDCS was employed to deliver more targeted stimulation to the DLPFC.

Experimental design

The experimental materials were sourced from the ImageNet Large Scale Visual Recognition Challenge (ILSVRC)²⁹. The AI model was GoogLeNet Inception v3, one of the top-performing models in image classification tasks³⁰.

The experimental task was a human-AI delegation task, specifically an image classification task. Image classification involves assigning a target image to a category, and it is a general task that does not require any domain-specific training or prior expertise. Tasks based on specific situational backgrounds may compromise generalizability. For example, in high-stakes scenarios involving potential loss, people tend to delegate decisions to AI³¹, effectively making it the team's scapegoat³². The task used in this study was adapted from the human-AI collaborative classification paradigm developed by Fuegener et al. and by Gass et al.^8,33.

A preliminary experiment was conducted to assess the difficulty level of the image. Image difficulty is the average accuracy of the corresponding image in the human alone condition. As shown in Figure 1, each trial consisted of a target image and five sets of image options. Participants were asked to choose the option set that best matched the target image. A total of 217 participants completed the preliminary study. All participants in the preliminary study were excluded from recruitment for the main experiment.

Figure 1: Preliminary experiment Interface. Interface shown to participants in the preliminary experiment, where they selected one response from five options. Please click here to view a larger version of this figure.

In the formal experiment interface (see Figure 2), a target image is displayed at the top of the screen. Additionally, five image group options and one delegation option (located at the top right of the screen) are presented. Participants need to decide between delegating the task to the AI and completing it themselves. In the latter case, they select one of the five image category options that best matches the target image. After each decision, the interface advances to the next page, where participants are asked to rate their confidence in their delegation decision on a scale from 1 to 4. Before the formal experiment began, participants were informed that the AI system had an average accuracy of 76%.

Figure 2: Formal experiment Interface. Main interface shown to participants during the task. (A) The task interface; (B) the confidence rating interface. Please click here to view a larger version of this figure.

A total of 62 right-handed participants were recruited for the main experiment. All participants were graduate students who had no prior experience with similar experimental tasks. The exclusion criteria were as follows: participants must not have taken any sedative-hypnotic drugs or psychoactive substances within 24 h prior to the experiment, must have no cranial defects (e.g., holes or fractures), and must not have any metal implants in the skull (e.g., metal plates). Two participants were excluded: one due to dermatological inflammation that prevented HD-tDCS stimulation, and one for ingesting a caffeinated beverage prior to the experiment. The remaining 60 participants (mean age = 22.68 ± 0.91 years) completed the experiment. They were randomly assigned to one of three stimulation conditions-anodal, cathodal, or sham-each with 20 participants. The electric field simulations are shown in Figure 3. Each participant performed the image classification task once before and once after stimulation, yielding 120 data sets for analysis. The HD-tDCS procedure is illustrated in Figure 4, and the experimental workflow is summarized in Figure 5. We conducted an a priori power analysis using G*Power 3.1³⁴, assuming a power of 90% (1 - β = 0.90), an effect size of f = 0.25, and an alpha level of 0.05. The analysis indicated that a minimum of 18 participants per group would be required. Our final sample size met this requirement.

Figure 3: Electric field simulation. Lateral view of the electric field distribution (right side). (A) Anodal stimulation condition. (B) Cathodal stimulation condition. Please click here to view a larger version of this figure.

Figure 4: HD-tDCS experimental setup. Photograph of the HD-tDCS experimental setup, illustrating the electrode placement and device position during stimulation. Please click here to view a larger version of this figure.

Figure 5: Experimental flowchart. The experiment comprised three phases, including image classification tasks conducted before and after the stimulation. Abbreviations: HD-tDCS = high-definition transcranial direct current stimulation; rDLPFC = right dorsolateral prefrontal cortex. Please click here to view a larger version of this figure.

Protocol

This study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Institutional Review Board of Jiangsu University of Science and Technology. We obtained the informed consent of the participants to use and release their data.

1. Preparation

1. Prepare for the experiment task.
   1. Download image data from the ImageNet website (https://image-net.org/).
      NOTE: The downloaded images are categorized, and each category contains multiple images.
   2. Run the Python code related to the GoogLeNet Inception v3 model to classify images (see Supplemental File 1). Save the top five predicted categories for each image, along with the Top-1 classification accuracy.
      NOTE: This Python code performs image classification by loading the Inception V3 model. For each target image, it outputs the top five predicted categories. The image sets corresponding to these five predicted categories are presented to participants as response options in the experiment. If the correct category is not among the top five predictions, one of the predicted categories is replaced with the correct one to ensure its inclusion among the options.
   3. Use Python or any random sampling program to select target image from the images classified by the model.
   4. For each target image, extract 10 images for each of the five categories to serve as response options. Resize all images to ensure uniform dimensions.
   5. Conduct a preliminary study and invite participants to take part in the experiment.
      NOTE: The preliminary study was similar to the formal experiment, except that participants were required to complete all trials themselves, without the option to delegate and without a confidence rating process.
   6. Record each participant's accuracy in responding to each image and calculate the accuracy rate for each image. Use the accuracy scores as an index of image difficulty to classify the images accordingly.
      NOTE: All images are sorted by accuracy, with higher accuracy indicating lower difficulty. For example, when 10% of participants correctly respond to an image, the difficulty of that image is set at 90%.
   7. Divide the images into two sets of comparable difficulty based on their difficulty scores. Use the two sets of images as experimental materials to create the formal experimental program through the E-prime 3 (see Supplemental File 1).
      NOTE: This experimental program was designed to present the image classification task.
2. Ensure that the participant has no contraindications to HD-tDCS and does not meet any exclusion criteria. Introduce the experimental procedure to participants and obtain written informed consent after confirming their understanding of the study.
   NOTE: Ask the participant about their dominant hand and confirm it by observing which hand they use to write.
3. Measure the participant's head circumference using a measuring tape and select the appropriate HD-tDCS neoprene headcap size (see Table of Materials) accordingly.
4. Verify that the HD-tDCS device is fully charged. Ensure that all accessories-including electrodes, electrode gel, syringe provided with the instrument, electrode cables, and the computer used to operate the HD-tDCS system-are ready for use.
5. Randomly select one of the two experimental tasks as the pretest condition. Instruct the participant to complete the image classification task.
   NOTE: A random number generator can be used to randomly assign experimental conditions.
   1. Open the program.
   2. Click on RUN, enter the subject number and session number, confirm the information, and click YES to start the experimental task.

2. HD-tDCS stimulation

1. Insert the bottom parts of the five electrodes into the corresponding positions on the neoprene headcap. These five positions serve as the stimulation sites.
   NOTE: To target the right DLPFC, use the following five electrode positions: F4, Fp2, Fz, C4, and F8.
2. Instruct the participant to sit comfortably in a chair. Position the front edge of the neoprene headcap approximately 1-2 cm above the eyebrows. Ensure that Cz is aligned with the midpoint between the two preauricular points. Fit the cap securely on the participant's head, ensuring proper alignment.
   NOTE: When fitting the headcap, be sure to locate Cz, the inion, nasion, and left and right preauricular points according to the international 10-20 EEG system.
3. Apply electrode gel (see Table of Materials) to each stimulation site using the syringe provided with the instrument to reduce impedance between the electrodes and the scalp.
   NOTE: The electrode gel is a highly conductive and water-soluble gel.
   1. Fill the syringe with electrode gel.
   2. For each electrode, part the hair with the syringe to expose the scalp.
   3. Inject a sufficient amount of electrode gel onto the exposed scalp inside the bottom parts of all the electrodes.
   4. Screw together the top and bottom parts of all the electrodes.
4. Connect the HD-tDCS device.
   1. Turn on the HD-tDCS device and launch the corresponding control software (see Table of Materials).
      NOTE: The computer running the control software must meet at least the following requirements: Processor: 1.6 GHz; RAM: 2 GB; Interface: USB or WiFi; Screen resolution: 1,280 x 768.
   2. Enable Bluetooth on the computer. In the control software, select USE Wi-Fi Devices to begin the connection process.
   3. Wait for the device's MAC address to appear in the list. Once it appears, select the device.
   4. In the Line Noise Filter setting, select 50 Hz to eliminate electrical interference.
      NOTE: The selected frequency should match the AC power frequency used in the country where the experiment is conducted.
   5. At the bottom of the settings panel, click on the Use This Device button. The software will then establish a connection with the device.
5. Configure the HD-tDCS stimulation protocol.
   1. Click the head-shaped icon on the left panel of the control software. Then, click ADD NEW PROTOCOL to create a new stimulation protocol.
   2. Set the protocol duration to 21 min, with a 30 s ramp-up and 30 s ramp-down period for current delivery.
      NOTE: If the protocol is a sham stimulation, enable the Sham option.
   3. Drag the F4 electrode into the Stimulation column on the right. Set the channel to stimulation mode, enter a current intensity of 2,000 µA³⁵, select anodal stimulation.
      NOTE: The designation of anode and cathode depends on the intended stimulation type.
   4. Drag electrodes Fp2, Fz, C4, and F8 into the Stimulation column. Set each channel to return mode, assigning 25% of the total current to each. Save the protocol.
   5. Select the configured stimulation protocol. Attach the HD-tDCS device to the Velcro strap at the back of the headcap. Use the electrode cables to connect the device to the electrode hardware.
      NOTE: Ensure that each cable is connected to the correct electrode position, matching the configuration in the software.
   6. Confirm that all cables are properly connected. Click the LOAD PROTOCOL button in the bottom right corner.
6. Click the CHECK IMPEDANCE button to perform an impedance check.
   NOTE: Ensure that the connection status indicators for all selected electrodes are green (0-10 kΩ) or orange (10-15 kΩ). If any indicator appears red (>15 kΩ), reapply the electrode gel or adjust the hair to expose more scalp to reduce impedance.
7. Click the start button (right-pointing triangle icon) to start the stimulation.
8. Instruct the participants to report their sensations at multiple time points during the stimulation session.
   NOTE: If a participant reports discomfort beyond normal sensations, immediately click the STOP button in the software, remove the device from the participant, and check the condition of their scalp.

3. Post stimulation

1. Turn off the control software. Disconnect the cables from the electrodes and carefully remove the HD-tDCS device.
2. Remove the headcap from the participant's head and detach all electrodes.
3. Select the other experimental program that has not yet been completed. Instruct the participant to complete the image classification task as the post-test.
4. Provide the participant with supplies to clean their hair.
5. Instruct participants to complete a sensation questionnaire after the HD-tDCS session³⁶. Use the questionnaire to assess participants' tolerance and subjective experience of the HD-tDCS stimulation.
6. Clean the headcap and electrodes.
   1. Rinse the headcap thoroughly with water. Hang the headcap to air dry.
   2. Use tap water to gently rinse the gel from the electrodes.
      NOTE: Use warm water to dissolve the gel more quickly.
   3. Leave the electrodes to air dry on a paper towel or cloth away from direct sunlight. Once dry, place them in their packaging for long-term storage.

Results

The key result of this study includes the measurement of metacognitive sensitivity using the meta-d'. A normality test was conducted on the metacognitive sensitivity scores of participants across different stimulation types, and the results showed that normality was met under all conditions. A repeated measures ANOVA was performed to compare metacognitive sensitivity across different stimulation types, with the results shown in Table 1. Although the main effects of time a...

Discussion

To ensure the reliability of the experimental results, several factors must be considered during the procedure. First, when connecting the cables to the HD-tDCS electrode hardware, it is crucial to verify that each cable corresponds precisely to the designated electrode position, as incorrect connections may cause unintended physiological effects. Second, since this is a single-blind study involving a sham stimulation condition, participants must not be able to view the operation interface during the HD-tDCS procedure. T...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Electrode cables	Neuroelectrics	10 electrode cable	The electrode cable has 10 electrode medical sockets.
Electrode gel	Parker Laboratories	Signagel 250g NE016b	It is used to decrease the impedance and improve the signal quality.
Electrodes	Neuroelectrics	NG PISTIM NE029	The NG Pistim consists of two pieces: the fastener (top part) and the threaded washer (bottom part).
Experimental software	PSYCHOLOGY SOFTWARE TOOLS	E-prime3.0 Build 3.0.3.31	This is the software used to make the experimental program.
HD-tDCS device	Neuroelectrics	Starstim 8	This is the device used to deliver HD-tDCS stimulation.
HD-tDCS software	Neuroelectrics	NIC2 V2.1.3.6	It is the application that allows a computer to interact with the HD-tDCS device.
Laptop	Dell	G5 5500	This computer runs the software used in the experiment.
Neoprene headcap	Neuroelectrics	M	It is a headcap to precisely place the electrodes on the scalp.Sizes:S (49 cm), M (54 cm), L (57 cm), XL (60 cm)
Statistical tool	IBM	SPSS V26	It is used for data analysis.
Syringe	-	NE014	It is used to inject electrode gel.