Methods, systems, devices, and media for predicting human trust behavior towards artificial intelligence

Abstract

This invention discloses a method, system, device, and medium for predicting human trust behavior towards artificial intelligence (AI), relating to the field of AI technology. The method includes: determining a participant's decision-making confidence based on a human decision-making confidence model; determining the participant's confidence in AI decisions based on a human confidence model for AI decisions; determining the expected utility of the participant trusting AI and the expected utility of the participant not trusting AI based on the participant's decision-making confidence and the participant's confidence in AI decisions; inputting the expected utility of the participant trusting AI and the expected utility of the participant not trusting AI into a human trust behavior prediction model for AI, and outputting a trust result; the human trust behavior prediction model for AI is trained using machine learning methods on a training set. This invention enables the prediction of human trust behavior towards AI.

Description

Technical Field

[0001] This invention relates to the field of human-artificial intelligence interaction technology, and in particular to a method, system, device and medium for predicting human trust behavior toward artificial intelligence. Background Technology

[0002] Artificial intelligence (AI) systems possess the ability to extract human-relevant information from vast amounts of data, enabling them to collaborate with humans in various decision-making domains. AI systems are increasingly being used to supplement human cognitive and computational capabilities, allowing them to perform tasks more efficiently and safely, and reducing human burden. However, the effectiveness of human-AI collaboration hinges critically on human trust in the AI ​​system. Trust is typically based on the system's capabilities and reliability, but human trust does not always align with the AI's capabilities. Inappropriate human trust in AI can diminish the benefits of using AI to improve efficiency and even threaten human safety. To prevent such problems, human trust in AI must be properly calibrated to prevent misuse (over-reliance), underuse (underutilization), or abuse (misapplication). Developing a quantitative method to predict human trust behavior is a necessary step in overcoming this challenge.

[0003] Trust has been defined in various forms by numerous scholars. These definitions share a common conceptual characteristic: they reflect beliefs, confidence, and expectations or attitudes towards uncertainty, reliability, and competence. Therefore, different academic paradigms have established different quantitative models, but there is a consensus on the quantification of trust—the established quantitative trust models are dynamic because human trust levels change with accumulated experience.

[0004] Despite increasing attention being paid to research on trust between humans and AI, the behavioral mechanisms by which humans accept or reject AI suggestions during decision-making in human-AI systems with AI-assisted decision-making remain unclear. Existing methods often focus on measuring the trust level variable itself, neglecting the psychological processes by which humans form trust. This can lead to significant differences between models and prevent their widespread applicability to different task scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device, and medium for predicting human trust behavior towards artificial intelligence, thereby enabling the prediction of human trust behavior towards artificial intelligence.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for predicting human trust behavior toward artificial intelligence includes:

[0008] Determine participants' decision-making confidence based on the human decision-making confidence model;

[0009] The confidence of participants in AI decision-making is determined based on a human confidence model.

[0010] Based on the participants' confidence in decision-making and their confidence in AI decision-making, determine the expected utility of participants trusting AI and the expected utility of participants not trusting AI.

[0011] The expected utility of the participants trusting AI and the expected utility of the participants not trusting AI are input into the human trust behavior prediction model for AI, and the trust result is output; the human trust behavior prediction model for AI is trained using machine learning methods based on the training set.

[0012] Optionally, the sample data in the training set includes input data and label data. The input data includes sample data of the expected utility of participants trusting artificial intelligence and sample data of the expected utility of participants not trusting artificial intelligence. The label data represents trust behaviors, which include trust in artificial intelligence decisions and distrust in artificial intelligence decisions.

[0013] Optionally, the machine learning method may employ logistic regression, support vector machine, random forest, XGBoost, or K-nearest neighbor algorithm.

[0014] Optionally, the human decision-making confidence model is represented as:

[0015]

[0016] Wherein, P(m=d) h |x) represents the participant's confidence in making a decision, m represents the stimulus state, and d h Let x represent the decision made by the participant, x represent the real stimulus with the mean, and noise be a normally distributed variable representing the participant's sensory noise. P(m=1|x) represents the participant's confidence in the decision given stimulus state m=1, and P(m=-1|x) represents the participant's confidence in the decision given stimulus state m=-1.

[0017] Optionally, the human confidence model in artificial intelligence decision-making is represented as:

[0018]

[0019] Among them, P AI (d AI =m|x) represents the participant's confidence in the AI's decision-making, m represents the stimulus state, and d AILet x represent the decision made by artificial intelligence, x represent the real stimulus with the mean, noise be a normally distributed variable representing the sensory noise of the participants, M represent the set of stimulus states, M = {1, -1}, and the parameter Θ = [0, 1], θ i Let P(mθ) represent the difficulty model of the i-th binary decision task. i |x) represents the probability of each stimulus level occurring when the condition is x, P AI (d AI =m|mθ i This indicates the capabilities and performance of artificial intelligence.

[0020] Optionally, based on the participants' confidence in their decision-making and their confidence in the AI's decision-making, the expected utility of participants trusting the AI ​​and the expected utility of participants not trusting the AI ​​are determined, specifically including:

[0021] EU accept =P(m=d) AI |x)*V accept -(1-P(m=d AI |x))*V accept ;

[0022] EU reject =P(m=d) h |x)*V reject -(1-P(m=d h |x))*V reject ;

[0023] Among them, EU accept This indicates that participants trust the expected utility of artificial intelligence, EU reject This indicates that participants do not trust the expected utility of artificial intelligence, V accept V represents the value of the incentive or penalty for participants trusting artificial intelligence. reject Participants do not trust the values ​​of incentives or penalties received from artificial intelligence, P(m=d) AI |x) represents the participants' confidence in the AI ​​decision-making process, P(m=d h |x) represents the participant's confidence in making decisions.

[0024] This invention also discloses a system for predicting human trust behavior towards artificial intelligence, comprising:

[0025] The participant decision confidence determination module is used to determine the participant's decision confidence based on the human decision confidence model;

[0026] The participant confidence in AI decision-making module is used to determine the participant's confidence in AI decision-making based on a human confidence model.

[0027] The expected utility determination module is used to determine the expected utility of a participant trusting artificial intelligence and the expected utility of a participant not trusting artificial intelligence, based on the participant's confidence in decision-making and the participant's confidence in the artificial intelligence decision-making.

[0028] The prediction module is used to input the expected utility of the participant trusting the artificial intelligence and the expected utility of the participant not trusting the artificial intelligence into the human trust behavior prediction model for artificial intelligence, and output the trust result; the human trust behavior prediction model for artificial intelligence is trained on the training set using machine learning methods.

[0029] The present invention also discloses an electronic device, including a memory and a processor, the memory for storing a computer program, the processor for running the computer program to cause the electronic device to perform the method for predicting human trust behavior in artificial intelligence.

[0030] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for predicting human trust behavior toward artificial intelligence.

[0031] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0032] This invention inputs the expected utility of participants trusting AI and the expected utility of participants distrusting AI into a prediction model of human trust behavior towards AI, and outputs the trust result. It can predict human trust behavior towards AI, thereby avoiding excessive reliance on or underutilization of AI in the process of human-AI interaction and decision-making, and appropriately calibrating human trust in AI. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic flowchart of a method for predicting human trust behavior towards artificial intelligence, provided by an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of a ball motion experiment without the participation of artificial intelligence, provided in an embodiment of the present invention.

[0036] Figure 3 This is a schematic diagram of a ball motion experiment involving artificial intelligence, provided in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram illustrating the principle of the human decision-making confidence model provided in an embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram illustrating the principle of the human confidence model in artificial intelligence decision-making provided in an embodiment of the present invention;

[0039] Figure 6 The confusion matrix diagram is shown for the random forest method provided in the embodiments of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The purpose of this invention is to provide a method, system, device, and medium for predicting human trust behavior towards artificial intelligence, thereby enabling the prediction of human trust behavior towards artificial intelligence.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Example 1

[0044] like Figure 1 As shown in the figure, this embodiment provides a method for predicting human trust behavior towards artificial intelligence, which includes the following steps.

[0045] Step 101: Determine the participants' decision confidence based on the human decision confidence model.

[0046] Step 102: Determine participants' confidence in AI decision-making based on the human confidence model.

[0047] Step 103: Based on the participants' confidence in decision-making and their confidence in AI decision-making, determine the expected utility of participants trusting AI and the expected utility of participants not trusting AI.

[0048] Step 104: Input the expected utility of the participants trusting artificial intelligence and the expected utility of the participants not trusting artificial intelligence into the human trust behavior prediction model for artificial intelligence, and output the trust result; the human trust behavior prediction model for artificial intelligence is trained using machine learning methods based on the training set.

[0049] Trust outcomes include trust in AI decisions and distrust in AI decisions. Furthermore, the prediction results of this invention can be predicted at the start of the task, without requiring real-time measurement of the operator's physiological data, thus enabling advance prediction. This invention is applied to the field of autonomous driving, adjusting AI decisions in autonomous driving based on the trust outcomes output by the AI ​​trust behavior prediction model.

[0050] This invention discloses a method for predicting human trust behavior towards AI in AI-assisted decision-making scenarios. Specifically, it calculates a person's confidence in themselves and in AI under different task scenarios to predict whether to accept AI's suggestions.

[0051] The training set sample data includes input data and label data. The input data includes sample data of expected utility of participants trusting artificial intelligence and sample data of expected utility of participants not trusting artificial intelligence. The label data represents trust behaviors, which include trust in artificial intelligence decisions and distrust in artificial intelligence decisions.

[0052] This embodiment of a method for predicting human trust behavior toward artificial intelligence also includes: establishing a difficulty calculation model for a binary decision-making task.

[0053] For different binary decision-making tasks, we analyze the characteristics of the tasks and the factors that affect their difficulty. These factors should include at least the uncertainty of the task, and for time-critical tasks, time pressure should also be considered. Then, we select an appropriate model to measure and quantify the task's difficulty.

[0054] The difficulty model of a binary decision-making task is represented as: θ=f(ρ,t).

[0055] Where θ represents the task difficulty, ρ represents the task uncertainty, t represents the time allotted to complete the task, and f() is the function for calculating the task difficulty.

[0056] Step 102 specifically includes: calculating a model of human decision-making confidence under different difficulty levels (human decision-making confidence model) based on the difficulty of the binary decision-making task. The established human decision-making confidence model should take into account the abilities of different operators.

[0057] In a task, assume there are real binary stimulus states m ∈ M = {1, -1}, for example, a stimulus to the left (m = -1) or to the right (m = 1). Then, participant s makes a decision d ∈ {1, -1}. To make this decision, participants base their decisions on the perceived level of evidence (the value of which depends on the task difficulty, θ). i Generate a noisy σ ∈Θ=[0,1]) sThe internal variable x (the level of stimulation accessible to participants) is used to calculate their confidence in their choices. Here, i = 1, 2, 3…n, and n is θ. i The quantity of x. x follows a Gaussian distribution with a mean equal to the level of the true stimulus.

[0058] x~N(mθ i ,σ s (1);

[0059] We can obtain the probability of occurrence at each stimulus level under condition x:

[0060]

[0061] From formula (1), it can be seen that when the real stimulus is (m, θ) i Given the following scenario, the probability that a participant perceives a stimulus level of x is:

[0062]

[0063] The participants' confidence levels regarding stimulus states m=1 or m=-1 are:

[0064]

[0065]

[0066] Participants make their final decision by comparing their confidence in the two options. h .

[0067]

[0068] Ultimately, confidence in a decision is defined as a participant's belief in a particular choice d. h Given the probability of being correct, combined with the choice d h Based on the decision-maker's level of belief in the relevant choices, we can provide the decision-maker's confidence level.

[0069]

[0070] Wherein, P(m=d) h |x) represents the participant's confidence in making a decision, m represents the stimulus state, and d h Let x represent the decision made by the participant, x represent the real stimulus with the mean, and noise be a normally distributed variable representing the participant's sensory noise. P(m=1|x) represents the participant's confidence in the decision given stimulus state m=1, and P(m=-1|x) represents the participant's confidence in the decision given stimulus state m=-1.

[0071] Specifically, step 102 includes: the model for calculating human confidence in AI decision-making should at least consider two parts: the difficulty of the decision-making task and the perceived capabilities of the AI. For example, if a task is relatively simple for the decision-maker, but the AI's capabilities are low and it may make mistakes, the decision-maker's confidence in the AI ​​will be low. Conversely, if a task is difficult, but the AI's capabilities are high, the participants' confidence in the AI ​​may be relatively high.

[0072] The human confidence model in AI decision-making is represented as follows:

[0073]

[0074] The first part reflects the difficulty of the task, and the second part calculates the AI's performance under the current stimuli. AI (d AI =m|x) represents the participant's confidence in the AI's decision-making, m represents the stimulus state, and d AI Let x represent the decision made by artificial intelligence, x represent the real stimulus with the mean, noise be a normally distributed variable representing the sensory noise of the participants, M represent the set of stimulus states, M = {1, -1}, and the parameter Θ = [0, 1], θ i Let P(mθ) represent the difficulty model of the i-th binary decision task. i |x) represents the probability of each stimulus level occurring when the condition is x, P AI (d AI =m|mθ i This indicates the capabilities and performance of artificial intelligence.

[0075]

[0076] Where φ is the cumulative normal distribution, and the parameter σ AI Updates can be performed using Rescorla-Wagner learning rules, or alternative learning rules can be chosen.

[0077] σ AI,t+1 =σ AI,t +α(ο t -σ AI,t (10);

[0078] Where t represents the task order, α is the learning rate (each participant has a different learning rate), and ο t This provides feedback on the AI's previous decision (correct = 1, incorrect = 0), where correct indicates the decision was correct and incorrect indicates the decision was incorrect.

[0079] Specifically, step 103 includes: combining the participants' confidence in themselves and their confidence in the AI ​​obtained in each decision from steps 101 and 102, calculating the expected utility of the participants trusting or distrusting the AI, using the following formula:

[0080] EU accept =P(m=d) AI |x)*V accept -(1-P(m=d AI |x))*V accept (11);

[0081] EU reject =P(m=d) h |x)*V reject -(1-P(m=d h |x))*V reject (12);

[0082] V accept V represents the value of the incentive or penalty for participants trusting artificial intelligence. reject Participants do not trust the values ​​of incentives or penalties received from artificial intelligence, P(m=d) AI |x) represents the participants' confidence in the AI ​​decision-making process, P(m=d h |x) represents the participant's confidence in making decisions.

[0083] EU accept and EU reject As the independent variable, the participant's trust in AI (1) or not (0) is used as the dependent variable to establish a prediction model of human trust behavior towards artificial intelligence. The prediction model of human trust behavior towards artificial intelligence is expressed as follows:

[0084] Final decision = g(EU accept EU reject (13);

[0085] Where Final decision represents the trust result, g() represents the machine learning method, and EU accept This indicates that participants trust the expected utility of artificial intelligence, EU reject This indicates that participants do not trust the expected utility of artificial intelligence.

[0086] Machine learning methods include, but are not limited to, logistic regression, support vector machine, random forest, XGBoost, and K-Nearest Neighbor algorithm.

[0087] This invention also includes collecting behavioral data from participants through experiments and estimating the parameters in the model.

[0088] Behavioral data of participants were collected through human factors experiments, and the parameter σ in the model was analyzed. s , σ AI α is estimated. Parameter estimation methods that can be used include variational Bayesian inference, the Expectation-Maximization algorithm (EM), and genetic algorithms.

[0089] Predict human trust behavior towards AI.

[0090] The parameters estimated in the model through the above steps are then substituted into the corresponding models in steps 101, 102, and 103 to predict human trust in AI. The accuracy can be calculated using the following formula:

[0091]

[0092] Where Accuracy represents the accuracy rate, and N C To predict the number of correctly predicted tasks, N represents the total number of tasks involved in the prediction. All tasks are binary decision tasks.

[0093] The trust prediction method proposed in this invention is based on human behavioral data and does not require the recording of human physiological data. In application, this method does not require participants to wear any experimental devices, offering advantages such as being non-invasive and highly efficient, and can provide a foundation for trust calibration.

[0094] Example 2

[0095] The following specific example illustrates a method of the present invention for predicting human trust behavior toward artificial intelligence.

[0096] Step S1: Establish a task difficulty calculation model for the decision-making task. For different binary decision-making tasks, analyze the characteristics of the task and the factors that affect the task difficulty. Factors affecting task difficulty should at least include task uncertainty, and for time-urgent tasks, time pressure should be included. In this embodiment, a binary decision-making experiment involving ball motion is involved. This is a binary decision-making task: among 30 moving balls, a certain proportion of the balls exhibit regular left or right motion. Participants need to detect the overall direction of motion of the moving points and make a decision. In this process, the uncertainty of the task is determined by the proportion of regularly moving balls; the lower the proportion, the higher the uncertainty of the task. This experiment does not consider the influence of time pressure. Therefore, the task difficulty is calculated by the following formula:

[0097]

[0098] Where, n 规则 n represents the proportion of regularly moving balls. 总Let θ be the total number of balls. From this formula, we can see that the smaller θ is, the more difficult the task.

[0099] This embodiment designed two experiments to collect behavioral data from participants. The experimental procedure is as follows: Figure 2 and Figure 3 As shown.

[0100] Step S2: Establish a participant decision-making confidence model. A schematic diagram of the participant decision-making confidence model is shown below. Figure 4 As shown. In the ball-movement task, there exists a true binary stimulus state m∈M={1,-1}, with the stimulus pointing left (m=-1) or right (m=1). Then, participant s makes a decision d∈{1,-1}. To make this decision, participants base their decisions on the perceived level of evidence (whose value depends on the task difficulty, θ). i Generate a noisy σ ∈Θ=[0,1]) s The internal variable x (the level of stimulation accessible to participants) is used to calculate their confidence in their choices. Here, i = 1, 2, 3…n, and n is θ. i The quantity of x. x follows a Gaussian distribution with a mean equal to the level of the true stimulus.

[0101] x~N(mθ i ,σ s ).

[0102] The confidence level of each participant in each decision can be calculated using formula (7).

[0103] The schematic diagrams of steps S3 and S4 are as follows: Figure 5 As shown.

[0104] Step S3: Establish a confidence model of participants in AI. The confidence model of participants in AI decision-making considers two parts: the difficulty of the decision-making task and the perceived capabilities of the AI. The confidence of participants in AI for each task can be calculated using formula (8).

[0105] Step S4: Establish a predictive model for participants' trust behavior towards AI. Based on formulas (11) and (12), calculate the expected utility of participants trusting and distrusting AI. In the ball-movement experiment, we set V... accept =20, V reject =5. Here, machine learning methods such as Logistic Regression (LR), Support Vector Machine (SVM), Random Forest, XGBoost, and K-Nearest Neighbor are used to predict whether humans trust AI.

[0106] Step S5: Collect behavioral data from participants through experiments and estimate the parameters in the model.

[0107] This embodiment designed two experiments to collect behavioral data from participants. The experimental procedure is as follows: Figure 2 and Figure 3 As shown,

[0108] Experiment 1: Preparatory Experiment.

[0109] In Experiment 1, participants were asked to perform 300 ball-movement experiments. The purpose of this experiment was to estimate the level of evidence perceived by different participants regarding the ball-movement task. In other words, in Bayesian decision theory, participants generate a noisy internal variable based on the perceived level of evidence, and then make decisions based on this internal variable. The goal was to estimate the perceived noise of each participant's perception of the ball-movement task.

[0110] Each task begins with a "+" on the screen. After the participant clicks a key on the keyboard, a stimulus of ball movement is displayed within a circular area (0.5s). The participant is then asked for their choice (left or right) and is required to make a choice as quickly and accurately as possible within 4s. Simultaneously, they receive feedback (0.5s) and a corresponding prompt (correct or incorrect) is given through their headphones. θ is randomly generated from (0,1). The direction of movement m is randomly selected from left to right.

[0111] Experiment 2: Human-AI Decision-Making Experiment.

[0112] In the second experiment, participants were asked to complete a ball-movement experiment with AI suggestions. Each task began with a "+" sign displayed on the screen. After participants randomly clicked a key on the keyboard, the ball-movement stimulus appeared in a circular area (0.5s), and participants were then asked for their choice (left or right). Next, the AI's suggestion was displayed on the screen (1s). Afterward, participants were asked whether they accepted the AI's suggestion; if yes, they pressed the 'Y' key on the keyboard, otherwise they pressed the 'N' key. Finally, participants received feedback on the experimental results. Three experimental blocks were set, each consisting of 100 human-AI decision-making experiments using the ball-movement task. Different accuracy rates were set for the AI ​​suggestions in these three experiments (approximately 60%, 80%, and 95%). It is worth noting that the AI's suggestions were generated according to rules; that is, the more difficult the task, the more likely the AI ​​was to make errors. However, it is believed that not using real AI would not affect the research, as there was no difference between the two for the participants. At the start of each block, participants are instructed to remove any reliance on the AI ​​developed in previous trials and to redo the task assuming the AI's capabilities are identical to theirs. Within each block, the task difficulty is presented randomly, and the AI's accuracy is also randomized for the participants based on the block order.

[0113] Furthermore, to prevent participants from over-relying on or distrusting the AI, a reward and penalty system was implemented for each task: when participants trusted the AI's advice, they would receive 20 points if the final result was correct, and deduct 20 points otherwise; when participants distrusted the AI's advice (trusting their own), they would receive 5 points if the final result was correct, and deduct 5 points otherwise. After each choice, participants would also be informed of their total score. Participants were informed that their final reward was related to their score. In short, the goal was to encourage participants to achieve the highest score by carefully considering whether to trust the AI.

[0114] Finally, variational Bayesian inference was chosen to estimate the parameters in the model. The parameter estimation results are shown in Table 1.

[0115] Table 1. Model parameters of participants

[0116]

[0117] Step S6: Estimate participants' trust behavior towards the AI. Based on the experimental data obtained in Step S5 and the estimated parameters, machine learning methods are used to predict whether each participant trusts the AI. The data is divided into a test set and a training set, and the accuracy on the test set is calculated. The final accuracy for each participant is shown in Table 2.

[0118] Table 2. Accuracy of different models predicting participants' trust behavior

[0119]

[0120]

[0121] A human trust behavior prediction model for artificial intelligence, trained using random forest, is used to predict human trust behavior towards artificial intelligence. The prediction results are as follows: Figure 6 As shown, Accept means trusting artificial intelligence, and Reject means not trusting artificial intelligence.

[0122] This invention does not directly measure trust, but rather views trust as a decision-making behavior. It predicts whether humans will trust AI based on two types of confidence in the decision-making process: human confidence in their own decisions and confidence in the AI's decisions. Furthermore, considering the variability among humans, it develops a personalized trust behavior prediction model, which is of great significance for trust calibration and improving the security and task performance of human-AI systems.

[0123] Example 2

[0124] This embodiment provides a system for predicting human trust behavior towards artificial intelligence, comprising:

[0125] The participant decision confidence determination module is used to determine the participant's decision confidence based on the human decision confidence model.

[0126] The participant confidence assessment module is used to determine participants' confidence in AI decision-making based on a human confidence model for AI decision-making.

[0127] The expected utility determination module is used to determine the expected utility of a participant trusting artificial intelligence and the expected utility of a participant not trusting artificial intelligence, based on the participant's confidence in the decision-making process and the participant's confidence in the decision-making process of artificial intelligence.

[0128] The prediction module is used to input the expected utility of the participant trusting the artificial intelligence and the expected utility of the participant not trusting the artificial intelligence into the human trust behavior prediction model for artificial intelligence, and output the trust result; the human trust behavior prediction model for artificial intelligence is trained on the training set using machine learning methods.

[0129] Example 3

[0130] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the method for predicting human trust behavior in artificial intelligence according to Embodiment 1.

[0131] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for predicting human trust behavior toward artificial intelligence as described in Embodiment 1.

[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0133] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting human trust behavior towards artificial intelligence, characterized in that, include: Determine participants' decision-making confidence based on the human decision-making confidence model; The confidence of participants in AI decision-making is determined based on a human confidence model. Based on the participants' confidence in decision-making and their confidence in AI decision-making, determine the expected utility of participants trusting AI and the expected utility of participants not trusting AI. The expected utility of the participants trusting AI and the expected utility of the participants not trusting AI are input into the human trust behavior prediction model for AI, and the trust result is output. The human trust behavior prediction model for artificial intelligence is obtained by training a training set using machine learning methods. The confidence of participants in AI decision-making is determined based on a human confidence model, which specifically includes: the human confidence model in AI decision-making should at least consider two parts: the difficulty of the decision-making task and the capabilities of AI as perceived by humans. The human confidence model in AI decision-making is represented as follows: Among them, P AI (d AI =m|x) represents the participant's confidence in the AI's decision-making, m represents the stimulus state, and d AI Let x represent the decision made by artificial intelligence, x represent the real stimulus with a mean, noise be a normally distributed variable representing the sensory noise of the participants, M represent the set of stimulus states, M = {1, -1}, and the parameter Θ = [0, 1], θ i Let P(mθ) represent the difficulty model of the i-th binary decision task. i |x) represents the probability of each stimulus level occurring when the condition is x, P AI (d AI =m|mθ i This indicates the performance capabilities of artificial intelligence. Where φ is the cumulative normal distribution, and the parameter σ AI Update using Rescorla-Wagner learning rules; s AI,t+1 =s AI,t +a(o t -s AI,t ); Where t represents the task order, α is the learning rate, and each participant has a different learning rate. t This is the feedback from the AI's previous decision. The feedback is either correct=1 or incorrect=0, where correct indicates that the decision was correct and incorrect indicates that the decision was incorrect.

2. The method for predicting human trust behavior towards artificial intelligence according to claim 1, characterized in that, The training set sample data includes input data and label data. The input data includes sample data of expected utility of participants trusting artificial intelligence and sample data of expected utility of participants not trusting artificial intelligence. The label data represents trust behaviors, which include trust in artificial intelligence decisions and distrust in artificial intelligence decisions.

3. The method for predicting human trust behavior towards artificial intelligence according to claim 1, characterized in that, The machine learning methods mentioned employ logistic regression, support vector machine, random forest, XGBoost, or K-nearest neighbor algorithm.

4. The method for predicting human trust behavior towards artificial intelligence according to claim 1, characterized in that, The human decision-making confidence model is represented as follows: Wherein, P(m=d) h |x) represents the participant's confidence in making a decision, m represents the stimulus state, and d h Let x represent the decision made by the participant, x represent the real stimulus with the mean, and noise be a normally distributed variable representing the participant's sensory noise. P(m=1|x) represents the participant's confidence in the decision given stimulus state m=1, and P(m=-1|x) represents the participant's confidence in the decision given stimulus state m=-1.

5. The method for predicting human trust behavior towards artificial intelligence according to claim 1, characterized in that, Based on the participants' confidence in decision-making and their confidence in AI decision-making, the expected utility of participants trusting AI and the expected utility of participants not trusting AI are determined, specifically including: I accept =P(m=d AI |x)*V accept -(1-P(m=d AI |x))*V accept ; I reject =P(m=d h |x)*V reject -(1-P(m=d h |x))*V reject ; Among them, EU accept This indicates that participants trust the expected utility of artificial intelligence, EU reject This indicates that participants do not trust the expected utility of artificial intelligence, V accept V represents the value of the incentive or penalty for participants trusting artificial intelligence. reject Participants do not trust the values ​​of incentives or penalties received from artificial intelligence, P(m=d) AI |x) represents the participants' confidence in the AI ​​decision-making process, P(m=d h |x) represents the participant's confidence in making decisions.

6. A system for predicting human trust behavior towards artificial intelligence, characterized in that, The system for predicting an individual's trust behavior in artificial intelligence, as described in claim 1, comprises: The participant decision confidence determination module is used to determine the participant's decision confidence based on the human decision confidence model; The participant confidence in AI decision-making module is used to determine the participant's confidence in AI decision-making based on a human confidence model. The expected utility determination module is used to determine the expected utility of a participant trusting artificial intelligence and the expected utility of a participant not trusting artificial intelligence, based on the participant's confidence in decision-making and the participant's confidence in the artificial intelligence decision-making. The prediction module is used to input the expected utility of the participant trusting the artificial intelligence and the expected utility of the participant not trusting the artificial intelligence into the human trust behavior prediction model for artificial intelligence, and output the trust result; the human trust behavior prediction model for artificial intelligence is trained on the training set using machine learning methods.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform a method for predicting human trust behavior in artificial intelligence according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method for predicting human trust behavior in artificial intelligence as described in any one of claims 1 to 5.

Images