Three major organizations are taking a different approach to AI ethics: they're building accountability into AI systems from the design phase onward, rather than trying to fix problems after deployment. Salesforce, UL Solutions, and DNV have each developed frameworks that prioritize transparency, fairness testing, and human oversight as core features of responsible AI, not afterthoughts.

Why Are Companies Struggling to Make AI Transparent?

As artificial intelligence becomes embedded in healthcare diagnoses, lending decisions, hiring systems, and critical infrastructure, the stakes for getting ethics right have never been higher. Yet many organizations deploy AI systems without clear mechanisms for understanding how those systems reach their conclusions. This opacity creates real problems: doctors can't explain why an AI recommended a particular treatment, loan officers can't tell applicants why they were denied credit, and regulators can't audit whether algorithms are discriminating against protected groups.

The challenge isn't new, but it's becoming urgent. UNESCO's 193 Member States adopted the first global standard on AI ethics in November 2021, establishing principles including transparency, fairness, environmental sustainability, and human oversight. Yet translating these principles into actual products remains difficult.

What Does Responsible AI Design Actually Look Like?

Salesforce has embedded responsible AI practices directly into its product development cycle. The company's Office of Ethical and Humane Use brings together responsible AI, ethical use policy, and accessibility teams to oversee AI systems before they reach customers. Every AI system undergoes intake, triage, review, testing, and implementation phases before deployment.

The testing process is rigorous. Salesforce conducts adversarial testing, which deliberately tries to break AI systems by feeding them tricky or malicious inputs. The company also performs content safety assessments, accessibility checks, employee trust testing, and large-scale stress testing to identify potential failure modes. At the platform level, guardrails cover privacy, accuracy, safety, autonomy, and monitoring, with additional controls designed to detect prompt injection attacks and keep AI agents within their intended scope.

"There's no question we are in an AI and data revolution, which means that we're in a customer revolution and a business revolution. But it's not as simple as taking all of your data and training a model with it. There's data security, there's access permissions, there's sharing models that we have to honour. These are important concepts, new risks, new challenges and new concerns that we have to figure out together," said Clara Shih, Senior Advisor and Former Head of Business AI at Meta.

Clara Shih, Senior Advisor and Former Head of Business AI at Meta

UL Solutions, a safety evaluation and certification company, has developed a different but complementary approach. Its UL 3115 Outline of Investigation provides a framework for assessing AI-based products before and during deployment. The framework covers technical performance, ethics, and governance, while also evaluating risks such as malfunction, misuse, malicious interference, and data compromise.

The ethical pillar of UL's framework specifically addresses fairness, bias, and data privacy. UL Solutions also assesses transparency, explainability, oversight, and lifecycle management, helping organizations understand how AI behaves in real-world applications. The company's AI algorithm reproducibility programme examines whether AI technologies can deliver predictable and repeatable outcomes, while marketing claim verification assesses whether claims made about AI-enabled products are actually supported by evidence.

How to Build Transparency Into AI Systems

Industry leaders have identified several key practices that make AI systems more transparent and accountable:

• Human Decision-Making Authority: AWS emphasizes that humans must retain final decision-making power and responsibility, often practiced through a "human-in-the-loop" approach where AI provides recommendations but humans make final calls.
• Data Protection Throughout Lifecycle: UNESCO stresses that user data must be securely safeguarded and respected throughout the entire lifecycle of the AI model, from training through deployment and monitoring.
• Algorithm Explainability: SAP advocates for users and operators to have clear insight into how an algorithm reaches its conclusions, avoiding opaque "black-box" decisions that no one can understand or audit.
• Robustness Against Harm: UNESCO requires that technologies be robust against attacks and engineered to prevent physical, psychological, or societal harm.
• Equitable Treatment: IBM emphasizes that systems must treat all individuals and groups equitably, actively working to prevent discrimination or bias stemming from training data or algorithms.

DNV, a global assurance and risk management company, takes a lifecycle-focused approach to trustworthy AI. Rather than treating compliance as a one-time checkbox, DNV emphasizes assurance throughout the technology's entire lifecycle. The company's recommended practice for AI-enabled systems provides guidance for understanding whether AI performs as intended and how legal and ethical requirements can be translated into practical assurance processes.

DNV also emphasizes the importance of identifying all stakeholders affected by AI, including people who may never directly interact with a system but experience its outcomes. In healthcare, for example, doctors may use AI to support diagnoses while patients experience the consequences of those decisions. In maritime operations, passengers, operators, developers, and regulators can have different relationships with the same AI system. DNV's approach therefore considers data, sensors, algorithms, digital twins, human oversight, and ongoing monitoring as interconnected elements of responsible AI.

"Artificial intelligence and generative AI may be the most important technology of any lifetime," said Marc Benioff, Chair, CEO, and Co-founder of Salesforce.

Marc Benioff, Chair, CEO, and Co-founder of Salesforce

Salesforce's transparency measures include AI disclosure, model cards, and evaluations focused on safety, privacy, truthfulness, and fairness. The company also incorporates human-AI handoffs and accessibility into the product experience, ensuring that when AI systems reach the limits of their capabilities, humans can smoothly take over.

These approaches represent a shift in how companies think about AI ethics. Rather than viewing transparency and fairness as regulatory burdens or public relations exercises, leading organizations are embedding these principles into the core architecture of their AI systems. The result is AI that can be audited, understood, and held accountable, even as it becomes more powerful and more widely deployed across industries that affect millions of people's lives.