Mitigating AI risk in the enterprise (Reader Forum)

Addressing AI risk is about more than compliance — enterprises need to build trust, promote fairness, and protect user privacy

With the constantly changing Artificial Intelligence (AI) landscape, enterprises must be vigilant about the risks associated with AI technologies. Mitigating risk in AI is not just about compliance, it’s about building trust, promoting fairness, and protecting user privacy. 

We continue to see the benefits and increased potential for the use of AI systems, but they do bring risks such as data privacy and security risks, model drift, compliance issues, operational risks, and ethical concerns. AI solutions can be invaluable for efficient decision-making, but they must make reliable decisions based on high quality data. When AI decisions are flawed, the consequences can harm both people and organizations. For example, if an AI system discloses proprietary information or provides erroneous advice, such occurrences can erode trust in both the organizations involved and the AI solution itself.

As AI systems become increasingly integrated into our daily lives, ensuring fairness, transparency, accountability, ethical use, and user privacy becomes crucial.

AI bias occurs when algorithms produce unfair or inaccurate results due to skewed data or faulty design. Here are some examples of how bias can skew results:

1. Facial recognition: AI systems could misidentify people if trained on insufficiently wide-ranging datasets.
2. Hiring algorithms: AI tools used to assess job candidates could mischaracterize candidates if using datasets trained on outdated historical hiring data.
3. Loan approvals: AI systems used by financial institutions could incorrectly deny loans, give adverse interest rate terms, or lower credit limits to individuals caused by insufficient training data.
4. Healthcare disparities: Medical AI algorithms could fail to account for certain patient conditions, leading to less optimal diagnoses and treatment recommendations.

When reviewing AI outputs for biases, appropriate success metrics must be used. These are some common metrics used to evaluate biases:

1. Disparate impact: Whether a model’s outcomes disproportionately affect specific groups. A value of 1 indicates no bias, while values greater or less than 1 indicate bias.
2. Demographic parity: The distribution of positive outcomes across different demographic groups, promoting an equal distribution of the model’s positive outcomes.
3. Equalized odds: Consideration of both true positive rates (sensitivity) and false positive rates (specificity) across different groups, ensuring consistent accuracy.
4. Counterfactual fairness: Evaluating how a model’s predictions change when a specific attribute of an individual is altered, helping to identify and address biases related to that attribute.

Continuous monitoring with a human-in-the-loop (HITL) approach, in which human oversight and intervention is included in the workflow, is necessary to review model outputs and identify biases. A combination of these metrics should be used to effectively identify and mitigate biases in AI models. Regular audits and updates to the models can help maintain fairness and transparency.

Evaluating a model’s fairness involves more than examining aggregate performance metrics such as precision, recall, and accuracy. These metrics can mask poor performance on smaller datasets, leading to biased predictions. Therefore, it’s critical to use bias-specific metrics for a comprehensive view of a model’s fairness.

The IEEE Standards Association (IEEE SA) aims to help organizations make their AI systems better by offering  IEEE CertifAIEd, a certification program for assessing the ethics of Autonomous Intelligent Systems (AIS) to help protect, differentiate, and grow product adoption. By aligning AI systems with the IEEE CertifAIEd criteria, organizations can make informed development decisions regarding their AI operations that are ethical, transparent, and fair. This enhances consumer trust and satisfaction while also supporting the organization’s commitment to responsible AI practices by:

• Mitigating risks: addressing biases and ensuring fair outcomes
• Improving transparency: providing clear explanations for AI decisions
• Ensuring accountability: maintaining accurate records and clear lines of responsibility
• Protecting privacy: safeguarding personal and private information
• Fostering inclusivity: ensuring equitable service for all stakeholders

The IEEE CertifAIEd program plays a key role in guiding organizations to implement trusted AI practices. The resulting certificate and mark demonstrate the organization’s effort to deliver a solution with a more trustworthy AIS experience to its users.