2.1.1. Process-based adaptation principle

The process-based adaptation principle refers to consciously adapting AI ethics during the technology design process. This mainly involves the development stage of technical staff. It requires technical staff to balance technology and ethics consciously, and also requires external regulation and incentives for them. External regulation of technical staff in the process mainly includes educational standards, reward and punishment mechanisms, and public accountability.

First is the educational standards. Some scholars suggest embedding ethics into AI teaching to help students reflect on the possible social impacts of AI and to improve moral awareness (Horton et al., Reference Horton, McIlraith, Wang, Majedi, McClure and Wald2022; Williams et al., Reference Williams, Ali, Devasia, DiPaola, Hong, Kaputsos and Breazeal2023; Reference Zhang, Lee, Ali, DiPaola, Cheng and BreazealZhang et al., Reference Zhang, Lee, Ali, DiPaola, Cheng and Breazeal2023). These practices mean that engineers can be taught ethical principles during learning or practical application to prevent them from unknowingly violating basic human rights.

In addition, process-based balance solutions also include rewards and punishments for technical staff. To prevent these regulations from overly restricting technological innovation, the government can introduce bonus incentive policies for companies that prioritise ethics, transparency, and safety. For example, some scholars advocate that government policies for responsible artificial intelligence innovation should include ethical incentives for technical staff, mechanisms to reward whistleblowers, and adaptive incentive programmes (Baldassarre et al., Reference Baldassarre, Caivano, Nieto and Ragone2025).

Finally, there is public accountability (Loi and Spielkamp, Reference Loi and Spielkamp2021), which reflects democratic governance. Therefore, implementing public accountability can also serve as an external regulation method for technical staff.

It is reasonable to provide internal and external regulation for technical staff because, from their perspective, the importance of balancing innovation and governance is proportional to the level of public reliance on AI. In addition, according to Lawrence Lessig, the Internet creates a form of privilege and makes this privilege possible (Dávila et al., Reference Dávila, De, Lessig and Perrone2023). Some studies also show that algorithmic bias comes from limited data sets and biased algorithm designers (Chen, Reference Chen2023). Therefore, given that technical staff have the privilege to set network rules and can produce significant impacts, it is necessary to regulate this privilege during the AI design process.

2.1.2. Top-down adaptation principle

The top-down adaptation principle mainly comes from government regulatory proposals. Its balanced approach is that regulatory mechanisms are updated as technology evolves. This adaptability is expressed in two forms: specific regulation and uniform regulation. Specific regulation is based on the analysis of each situation. It is targeted and includes risk-proportionate regulation (Akinsuru and Iseoluwa, Reference Akinsuru and Iseoluwa2025), targeted regulation (Ghimire and Edwards, Reference Ghimire and Edwards2024), and interdisciplinary regulation (Bringas et al., Reference Bringas, Nannini, Rieger, Scott, Zhao, Patro and Kinder2022; Pfeiffer et al., Reference Pfeiffer, Gutschow, Haas, Möslein, Maspfuhl, Borgers and Alpsancar2023; Patel et al., Reference Patel, Gu, Esper, Maeser and Maeser2024).

Uniform regulation has compatibility. First, it involves active participation in global and cross-field dialogue. For example, regulatory agencies continuously reflect, adapt, and improve while cooperating to develop AI guidelines (Akinrinola et al., Reference Akinrinola, Okoye, Ofodile and Ugochukwu2024). This can be considered a complex governance situation or establish a collaborative network that combines government authority with industry expertise, creating highly adaptive and responsive dynamic laws (Walter, Reference Walter2024). Second, it provides unified or sustainable solutions. For example, it may suggest achieving uniform international AI regulations, emphasising public participation and international cooperation (Kashefi et al., Reference Kashefi, Kashefi and Ghafouri2024), building sustainable AI regulations (Hacker, Reference Hacker2024), or adopting a joint AI governance approach that places bioethical principles at the core of AI rules (Fessenko and Jasperse, Reference Fessenko and Jasperse2025).

Although these principles highlight the key issue that AI governance should be flexible and adaptive, the characteristics of such technology governance and AI supervision solutions are mainly based on traditional human governance. They may still face problems such as long governance cycles, neglect of key issues, delays, and a lack of dynamic tracking. This is determined by the long governance cycles and the uncertainty of risks brought by emerging technologies.

2.2.1. The conflict between abstract principle requirements and specific practical applications

This mainly appears in two aspects. First, there are difficulties in converting abstract principles into specific applications in terms of completeness, translatability, and explainability. One, embedded ethics is practice-oriented rather than normative and may have some incompleteness (Bleher and Braun, Reference Bleher and Braun2023). Two, it is questioned whether morality can be applied in programming. Three, governance principles and technical practices lack a communication bridge, which leads to few ethical explanations of technology or inaccurate translation of abstract principles into specific practice. Therefore, training interdisciplinary communicators as translation bridges for ethical engineering may help resolve this conflict.

Second, there is no unified standard for guiding principles in ethical engineering practice. On the one hand, this is because moral concepts themselves are uncertain. As Irina Shklovski stated, ethical methods cannot be finalised. Efforts in AI ethics do not end with a clear checklist or a set of principles. They must continuously address new concerns as part of the challenging task of cultivating certainty (Shklovski and Némethy, Reference Shklovski and Némethy2023). This means that ethical concepts are still changing. On the other hand, many ethical concepts are vague, contradictory, and yet similar in meaning. For example, respecting “openness” may conflict with respecting “privacy” (Kazim and Koshiyama, Reference Kazim and Koshiyama2021). The meaning of “explainability” and how to implement it is also unclear (Prem, Reference Prem2025). These uncertainties increase the difficulty of converting principles into practice.

Therefore, it is necessary to ease the conflict between principles and practice. This is a prerequisite for guiding practice. In this regard, governments should further develop clear communication channels and decision-making protocols to propose appropriate ways to mitigate this conflict.

2.2.2. Increased technical costs

Even if the conflict between the principles and technical practices is resolved, there may still be an issue of higher technical costs (Akash et al., Reference Akash, Lessard, Reza and Islam2024). This is because additional technical work is needed. This will increase both the time and financial costs of implementing the technology. For example, the costs of updating and maintaining tools, testing for vulnerabilities by external auditors, salaries, benefits, and training for technical staff, fixing vulnerabilities, upgrading and improving systems, and managing risks (Das et al., Reference Das, Banerjee, Chatterjee and Ghosh2023).

2.2.3. Reduced technical efficiency

For example, consider differential privacy technology. Differential privacy technology also faces some technical problems. This technology usually needs to balance utility and privacy. Generally, the two are inversely related. If accuracy increases, privacy protection decreases, and the opposite is also true. Although it helps prevent linkage attacks and protect privacy, it lacks effective sensitivity values and may affect the reliability of results (Husnoo et al., Reference Husnoo, Anwar, Chakrabortty, Doss and Ryan2021).

It can be seen that although embedding governance principles in the engineering design process is necessary and helps to address the possible delay of principle-based schemes, the ethical engineering approach still faces a fundamental challenge. This challenge is how to balance improving efficiency with embedding AI governance values and principles. Therefore, a neutral approach between technology developers and government officials is needed. This approach needs to address: (1) the delay of AI governance principle schemes, and (2) the efficiency loss caused by considering AI governance in the engineering field.

3.4.1. Defining risk variables based on the EU AI Act

The table is shown in Table 1.

Table 1.

Computable variables and regulatory measures mapping table

Table 1. Long description

The table consists of four columns and seventeen rows of data.

* Row 1: Manipulation; Whether the system influences behaviour by using cognitive bias; Art. 5; Ban on release or removal from the market, fines, mandatory correction.

* Row 2: Vulnerability exploitation; Whether the system targets children or disabled groups; Art. 5; Prohibition, compliance review.

* Row 3: Severe rights impact; Whether the system directly harms equality rights or personal rights; Art. 5; Prohibition, legal responsibility.

* Row 4: Health and safety impact; Accident rate, misdiagnosis rate, failure rate; Art. 15; Safety testing, robustness assessment, fail safe protection.

* Row 5: Rights impact; Employment or credit discrimination impact, complaint rate; Art. 9; Risk assessment, rights impact assessment.

* Row 6: System risk; Failure rate, success rate of adversarial attacks; Art. 15; Security strengthening, red team testing.

* Row 7: Scale; Number of affected people normalised; Annex I I I (indirect); Graded deployment, staged release.

* Row 8: Severity; Possible property or rights loss; Art. 9; Preventive control, emergency plan.

* Row 9: Data bias; Fairness difference, suitability, completeness, equality; Art. 10; Bias reduction, reweighting, fairness constraints.

* Row 10: Data quality; Relevance, originality, availability; Art. 10; Data cleaning, label review.

* Row 11: Representativeness; Insufficient coverage index; Art. 10; Expanded sampling, stratified sampling.

* Row 12: Transparency; Disclosure coverage, including whether A I involvement and use limits are explained; Art. 13; Additional instructions for use, mechanism description.

* Row 13: Explainability; Ratio of explainable outputs or explanation success rate; Art. 13; Explanation interface, reports.

* Row 14: Human oversight; Ratio of human review or ability to stop the system; Art. 14; Threshold-based review, one-click stop.

* Row 15: Logging; Coverage rate of key event records; Art. 12; Full logging, audit interface.

* Row 16: Documentation; Completeness rate of documents; Art. 11 forward slash Annex I V; Technical documents, update records.

* Row 17: Risk management; Risk assessment frequency or completion rate; Art. 9; Establish a continuous risk management system.

3.4.2. Taking an intelligent recruitment screening system as an example

This example strictly follows the compliance requirements of the EU AI Act. It presents the full process from legal rule analysis and numerical risk reasoning to the automatic generation of governance suggestions. Relevant calculation steps are also provided to ensure that the results can be reproduced and verified.

First, it is necessary to determine the basic condition for starting the whole risk assessment process. According to Annex III of the EU AI Act, AI systems used for recruitment, staff screening, and promotion are clearly defined as high-risk systems. Therefore, the risk assessment needs to consider risks related to discrimination, non-discrimination, and fairness. More specifically, requirements need to be set for the quality of training, validation, and test datasets. This includes checking historical bias, ensuring data representativeness, adding suitable human review procedures, and taking necessary corrective measures.

Second, the Bayesian network nodes and their parameters used in this example are given in Table 2.

Table 2.

Bayesian network nodes and their parameters

Table 2. Long description

The table consists of four columns and five data rows.

* Row 1: Risk node is Historical data bias (H B). State definition is High or Low. A I A legal rule is Art. 10: Check historical data bias. Prior probability is P(H B equals High) equals 0.60.

* Row 2: Risk node is Data representativeness (D R). State definition is Poor or Good. A I A legal rule is Art. 10: Data should represent the target group. Prior probability is P(D R equals Poor) equals 0.70.

* Row 3: Risk node is Bias mitigation effectiveness (B M). State definition is Ineffective or Effective. A I A legal rule is Art. 10: Appropriate bias mitigation measures should be taken. Prior probability is P(B M equals Ineffective) equals 0.50.

* Row 4: Risk node is Discriminatory outcome (D O). State definition is Likely or Unlikely. A I A legal rule is Art. 9: Risk assessment should consider discrimination risk. Prior probability is Defined by C P T.

* Row 5: Risk node is High-risk determination (H R). State definition is Yes or No. A I A legal rule is Art. 6 and Annex III: High-risk system determination standard. Prior probability is Defined by C P T.

Among these nodes, $ HB $ , $ DR $ , and $ BM $ are root nodes. Their prior probabilities are set based on common bias problems in the industry. $ DO $ and $ HR $ are non-root nodes. Their probabilities are determined by their parent nodes and are quantified through conditional probability tables (CPTs). For example, the CPT of the intermediate node “Discriminatory Outcome” is set as follows. When $ HB= High $ , $ DR= Poor $ , and $ BM= Ineffective $ , the probability of a discriminatory outcome is the highest and is set at 0.95. When all three conditions are in ideal states, the probability is the lowest and is set at 0.10. This setting reflects the real situation in which multiple factors together can sharply increase risk.

Finally, conduct a prior risk assessment. Under the default condition, it is necessary to assess the possibility that this intelligent recruitment system may contain potential risks. The state of $ HR $ is determined only by $ DO $ . Therefore, the prior probability of the intermediate node “Discriminatory Outcome,” namely $ P\left( DO= Likely\right) $ , is first calculated. Then, $ P\left( HR= Yes\right) $ is calculated based on this result.

Since $ DO $ has three parent nodes, namely $ HB $ , $ DR $ , and $ BM $ , and each node has two states, there are $ {2}^3=8 $ possible combinations. It is necessary to examine all combinations, calculate the probability of each combination, and calculate the conditional probability that $ DO $ is Likely under each combination. After that, all results are added together. The calculation process is as Table 3.

Table 3.

The calculation process of the prior probability that the intelligent recruitment screening system produces discriminatory results

Table 3. Long description

The table consists of seven columns: Combo I D, H B, D R, B M, P (Combo), P (D O = Likely | Combo), and Contribution.

* Row 1: Combo 1, H B High, D R Poor, B M Ineffective, P (Combo) 0.6 times 0.7 times 0.5 = 0.21, P (D O = Likely | Combo) 0.95, Contribution 0.21 times 0.95 = 0.1995.

* Row 2: Combo 2, H B High, D R Poor, B M Effective, P (Combo) 0.6 times 0.7 times 0.5 = 0.21, P (D O = Likely | Combo) 0.70, Contribution 0.21 times 0.70 = 0.1470.

* Row 3: Combo 3, H B High, D R Good, B M Ineffective, P (Combo) 0.6 times 0.3 times 0.5 = 0.09, P (D O = Likely | Combo) 0.60, Contribution 0.09 times 0.60 = 0.0540.

* Row 4: Combo 4, H B High, D R Good, B M Effective, P (Combo) 0.6 times 0.3 times 0.5 = 0.09, P (D O = Likely | Combo) 0.40, Contribution 0.09 times 0.40 = 0.0360.

* Row 5: Combo 5, H B Low, D R Poor, B M Ineffective, P (Combo) 0.4 times 0.7 times 0.5 = 0.14, P (D O = Likely | Combo) 0.50, Contribution 0.14 times 0.50 = 0.0700.

* Row 6: Combo 6, H B Low, D R Poor, B M Effective, P (Combo) 0.4 times 0.7 times 0.5 = 0.14, P (D O = Likely | Combo) 0.30, Contribution 0.14 times 0.30 = 0.0420.

* Row 7: Combo 7, H B Low, D R Good, B M Ineffective, P (Combo) 0.4 times 0.3 times 0.5 = 0.06, P (D O = Likely | Combo) 0.30, Contribution 0.06 times 0.30 = 0.0180.

* Row 8: Combo 8, H B Low, D R Good, B M Effective, P (Combo) 0.4 times 0.3 times 0.5 = 0.06, P (D O = Likely | Combo) 0.10, Contribution 0.06 times 0.10 = 0.0060.

* Total Row: P (Combo) sum is 1.00, Contribution Sigma equals 0.5725.

Thus, we obtain the prior probability that the system produces a discriminatory outcome without additional information as:

$$ P\left( DO= Likely\right)\times \Sigma P\left( DO= Likely| HB, DR, BM\right)\times P(HB)\times P(DR)\times P(BM)=0.5725 $$

Therefore, in the initial state, the recruitment system has a 57.25% probability of producing a discriminatory outcome.

Next, we calculate the prior probability of the top level node, which is the high-risk decision P(HR = Yes). The same law of total probability is used:

$$ {\displaystyle \begin{array}{c}P\left( HR= Yes\right)=P\left( HR= Yes| DO= Likely\right)\times P\left( DO= Likely\right)\\ {}+P\left( HR= Yes| DO= Unlikely\right)\times P\left( DO= Unlikely\right)\end{array}} $$

Substituting the known values, where $ P\left( HR= Yes| DO= Likely\right)=0.90 $ , $ P\left( HR= Yes| DO= Unlikely\right)=0.10 $ , and $ P\left( DO= Unlikely\right)=1-P\left( DO= Likely\right)=1-0.5725=0.4275 $ , we have:

$$ P\left( HR= Yes\right)=0.90\times 0.5725+0.10\times 0.4275=0.51525+0.04275=0.5580 $$

From this result, we obtain that without any additional evidence, the prior probability that the system is classified as “high risk” is 55.8%. This value is already higher than most internal control thresholds in companies, for example, 50%. It shows that the system has a clear compliance risk even before deployment.

3.4.3. Dynamic update of risk probability

The following section demonstrates how to use dynamic evidence to drive Bayesian updating in the model and then carry out risk attribution analysis on this basis.

(1) Assume that a third-party organisation conducts a bias audit on the recruitment system and finds serious issues. The specific evidence E is that the four-fifths rule is not satisfied. The hiring rate for female candidates compared with male candidates is 61.5%. This is below the legal threshold of 80%. This value is the key input that drives the model update.

(2) The likelihood function is set. It describes the probability of observing the evidence under different states of “Discriminatory Outcome” as $ P\left(E| DO= Likely\right)=0.92 $ . If the system does have a discriminatory outcome, then the audit has a 92% chance to detect this bias: $ P\left(E| DO= Unlikely\right)=0.15 $ . If the system does not have a discriminatory outcome, the audit still has a 15% chance of giving a false positive or reporting bias. This non-zero probability takes into account possible sampling errors or other limits in the audit process.

(3) Bayesian update. We calculate the posterior probability $ P\left( DO= Likely|E\right) $ . We use Bayes’ theorem to compute the posterior probability of “Discriminatory Outcome” given the observed audit failure evidence $ E $ . The formula of Bayes’ theorem is as follows:

$$ P\left( DO= Likely|E\right)=\frac{P\left(E| DO= Likely\right)\times P\left( DO= Likely\right)}{P(E)} $$

In this equation, the denominator $ P(E) $ is the total probability of observing the evidence $ E $ . It can be calculated using the law of total probability:

$$ {\displaystyle \begin{array}{c}P(E)=P\left(E| DO= Likely\right)\times P\left( DO= Likely\right)\\ {}+P\left(E| DO= Unlikely\right)\times P\left( DO= Unlikely\right)\end{array}} $$

Substituting the values:

$$ P(E)=0.92\times 0.5725+0.15\times \left(1-0.5725\right)=0.5267+0.064125=0.590825 $$

The posterior probability is:

$$ P\left( DO= Likely|E\right)=\frac{0.92\times 0.5725}{0.590825}\approx 0.8915 $$

After the impact of this single piece of evidence from the audit failure, the prior probability of 57.25% for a discriminatory outcome is significantly increased to a posterior probability of 89.15%.

3.4.4. Reverse inference: Identifying key risk sources $ P\left( HB= High|E\right) $

We conduct a reverse inference to calculate the posterior probability that the root node “Historical Data Bias” is in the state High under the evidence of audit failure. This is written as $ P\left( HB= High|E\right) $ . This step is used for risk attribution.

Here, we calculate the contribution weight of key combinations. For example, $ {Combo}_1\left( HB= High, DR= Poor, BM= Ineffective\right) $ contributes to the explanation of evidence $ E $ as follows:

$$ {\displaystyle \begin{array}{c}{Contrib}_1=P\left(E| DO= Likely\right)\times P\left( DO= Likely|{Combo}_1\right)\times P\left({Combo}_1\right)\\ {}=0.92\times 0.95\times 0.21=0.18354\end{array}} $$

The weight of this combination in the total probability of evidence $ P(E) $ is:

$$ {Weight}_1=\frac{Contrib_1}{P(E)}=\frac{0.18354}{0.590825}\approx 0.3107 $$

Using the same method, we calculate the weights of all combinations where $ HB= High $ , which include $ {Combo}_1 $ , $ {Combo}_2 $ , $ {Combo}_3 $ , and $ {Combo}_4 $ . We then sum these values. The total contribution of $ HB= High $ is about 69.5%. This value is much higher than its prior probability of 60%.

Attribution conclusion: under the audit failure evidence, historical data bias is the main cause of the current high risk state.

$$ {\displaystyle \begin{array}{c}P\left( HR= Yes|E\right)=P\left( HR= Yes| DO= Likely\right)\times P\left( DO= Likely|E\right)\\ {}+P\left( HR= Yes| DO= Unlikely\right)\times P\left( DO= Unlikely|E\right)\\ {}=0.90\times 0.8915+\left(1-0.9\right)\times \left(1-0.8915\right)=0.8132\end{array}} $$

The probability of high-risk classification increases from 55.8% to 81.32%. This clearly triggers the mandatory requirement in Article 16 of the EU AI Act that corrective measures must be taken for high-risk systems.

3.4.5. Generation of automated governance recommendations

We can build a simple rule engine. This engine maps the numerical outputs of the model directly to specific governance actions and related legal provisions. The process is shown in Table 4.

Table 4.

Automated governance recommendations

Table 4. Long description

The table consists of three columns: Trigger condition, Corresponding A I A article, and Governance recommendation.

* Row 1: Trigger condition is P (H R equals Yes given E) greater than 80 percent. Corresponding A I A article is Art. 16. Governance recommendation is to immediately stop the automatic decision function of the system in the production environment until the correction work is completed.

* Row 2: Trigger condition is P (H B equals High given E) greater than 65 percent. Corresponding A I A article is Art. 10. Governance recommendation is to launch a special audit on bias in historical data, focusing on differences in hiring rates across dimensions such as gender and age.

* Row 3: Trigger condition is P (B M equals Ineffective given E) greater than 50 percent. Corresponding A I A article is Art. 10. Governance recommendation is to redesign the bias mitigation plan and introduce adversarial debiasing techniques.

* Row 4: Trigger condition is P (D O equals Likely given E) greater than 85 percent. Corresponding A I A article is Art. 14. Governance recommendation is to enable a mandatory human review routing system where all low confidence predictions must be manually checked by the human resources department.

3.4.6. Explanation output

Table 5 is an example of a compliance risk summary designed for regulators.

Table 5.

Example of a compliance risk summary

Table 5. Long description

The table consists of two columns titled Item and Content, with four rows of data.

* Row 1: Item is Risk level. Content is High risk with a confidence of 81.3 percent.

* Row 2: Item is Main cause. Content is Historical data bias with a contribution of 69.5 percent.

* Row 3: Item is Evidence support. Content states that a third-party audit found that the four-fifths rule is not satisfied. The female hiring rate is 61.5 percent, which is below the 80 percent threshold.

* Row 4: Item is Legal link. Content specifies a violation of data governance requirements under Art. 10(2)(g) of the E U A I Act.

Recommended actions: (1) Immediately suspend the automatic screening function (based on Art. 16). (2) Submit a bias root cause analysis report within 72 hours (based on Art. 17). (3) Retraining must pass review by an independent ethics committee before it starts (based on Art. 9).

3.4.7. Continuous iteration and governance

The results of governance actions can be used as new evidence. This feedback is sent back into the network and drives model iteration and development. For example, if the company adopts the recommendation and takes the following actions:

1. (1) Data resampling is completed. This means that the situation of historical data bias and data representativeness is improved. We update $ P\left( DR= Good\right) $ from $ 0.3 $ to $ 0.7 $ .

2. (2) An adversarial debiasing algorithm is deployed. This improves the effectiveness of bias mitigation measures. We update $ P\left( BM= Effective\right) $ from $ 0.5 $ to $ 0.8 $ .

Next, we input the updated prior probabilities into the network and run the inference process again. The new calculation will show whether $ P\left( HR= Yes\right) $ has fallen to an acceptable level, for example below 30%. If the risk is reduced, it shows that the governance measures are effective. If the risk is still high, it is necessary to find new evidence, such as new monitoring logs, and apply deeper governance actions. This forms a spiral and continuous risk management loop. This mechanism ensures that compliance of the AI system is not static. It is a dynamic ability that can adapt to changes in internal and external environments.

Therefore, the Bayesian network, together with dynamic evidence-based probability updates and an automatic rule engine, can convert the abstract legal requirements of the EU AI Act into a risk governance framework for an intelligent recruitment system. This framework is quantifiable, explainable, and executable.

3.5.1. The interpretable and adaptive AI governance framework helps to provide an inclusive governance path

At present, there is a global competition in artificial intelligence regulation (Neuwirth, Reference Neuwirth2024). However, balancing AI governance and AI innovation is a common challenge faced worldwide. The interpretable and adaptive AI governance framework has general applicability.

First, the interpretable and adaptive AI governance framework is scalable and can be applied in different countries. Different national risk quantification schemes can be entered into causal models, such as Bayesian models, making it possible to promote their use across countries. Second, the interpretable and adaptive AI governance framework can balance AI innovation and AI governance. On one hand, companies that follow this framework can shorten the time for technology release and application, which helps to avoid profit loss caused by long review periods. On the other hand, governments can use this framework to address the potential delay and static nature of traditional governance schemes.

This requires:

(1) active dialogue between countries to discuss and establish reasonable quantitative standards for artificial intelligence risk assessment; (2) regulators and experts providing quantitative standards for ethical and compliance judgements to technical personnel; and (3) technical personnel actively understanding the relevant quantitative standards and embedding them into causal models.

In conclusion, this depends on cross-disciplinary cooperation, mainly involving technical staff, evaluators (experts), and other stakeholders. The purpose is to promote the optimisation of relevant causal models, especially in dealing with new ethical situations. A well-trained causal model can infer potential or similar ethical causes based on risk assessment results. This can reduce potential ethical problems that may arise from AI innovation.

Therefore, the explainable adaptive AI governance framework has the advantage of universality and can serve as an inclusive governance approach.

3.5.2. The explainable adaptive AI governance framework can serve as a supplement to static governance strategies

AI dynamic governance and static governance can complement each other. Because technology changes quickly, dynamic governance is necessary. Since regulatory uncertainty and ethical rules are hard to formalise, governance should combine static and dynamic strategies for the long term. Expert assessment should be the main approach, supported by machines.

This requires: (1) regulators to develop multiple dynamic governance plans, for example, combining stage-based supervision with full-cycle supervision; (2) technical staff to develop reliable assessment and decision tools for different purposes; and (3) companies to conduct self-inspections according to laws and regulations.

3.5.3. The explainable adaptive AI governance scheme helps to automatically generate value alignment strategies

The explainable adaptive AI governance scheme produces risk probability values and risk types that can generate appropriate value-based decision suggestions. For example, when the causal model shows that a technology has a high risk of opacity, it will suggest increasing transparency and seeking decision suggestions from the decision database that align with values. For this, it is essential to input complete value strategies into the causal model to provide correct suggestions. In addition, when new technologies may cause ethical issues that go beyond the existing decision database, experts are needed to assess and analyse the new ethical problems, provide new decision suggestions, and input them into the causal model.

This requires: (1) experts to continuously predict and publish potential ethical issues in AI development; (2) technical staff to update the causal model structure in time according to expert suggestions; and (3) regulators to act as a bridge by classifying existing ethical decisions and promptly communicating new ethical decisions to companies.