AI & the Future of Functional Safety Software Development: A Pragmatic Path Forward

AI promises to revolutionize functional safety software, but its path is twofold. Read on for a pragmatic roadmap for safely harnessing AI to accelerate development today while carefully navigating the risks of embedding it within critical systems.

Artificial intelligence (AI) is transforming industries at an unprecedented pace. Embedded software development for functional safety systems is no exception. From automotive to aerospace, developers are exploring how AI can enhance productivity, improve quality, and accelerate time to market. But as with any transformative technology, the path forward must be navigated with care, especially when safety is on the line.

Let’s explore two distinct dimensions of AI in the context of functional safety:

• AI in the development process and tooling automation
• AI deployed within embedded systems, for example, inference engines

While both are promising, they present vastly different risk profiles and maturity levels. AI-powered tooling and automation provides near-term promise in boosting productivity and innovation in a notoriously cautious field, while deploying AI models in safety-critical embedded environments is emerging but merits some caution and awareness of risks.

AI in the Development Process: A Productivity Catalyst

The software development life cycle for safety-critical systems is notoriously rigorous. Standards like DO-178C, ISO 26262, and IEC 61508 demand exhaustive documentation, traceability, and verification of the entire development process. This complexity, while essential for safety, often slows down innovation and increases cost.

Here’s where AI can shine—not in replacing engineers, but in augmenting them. Human-in-the-loop development processes with AI-augmented tooling promises to boost productivity by:

• Guiding and enforcing processes.
• Ensuring continuous compliance with standards.
• Simplifying the creation or referencing of relevant documentation.
• Helping developers and verification teams focus on the important tasks at hand.

Key Opportunities for Embedded Development

• Automated test generation. AI can analyze requirements and generate meaningful test cases, reducing manual effort and improving coverage.
• Requirements traceability. Natural language processing (NLP) can help link requirements to code, tests, and documentation, streamlining compliance.
• Static and dynamic analysis enhancement. AI can assist in identifying patterns in code that may lead to defects, improving early detection. For example, AI is extremely adept at boosting the signal-to-noise ratio and minimizing the dilutive impact of false positives, as well as suggesting fixes.
• Assurance argument support. AI tools can help draft structured safety arguments, like using Goal Structuring Notation, though human oversight remains essential.

These applications are deterministic, auditable, and bounded, making them well-suited for safety-critical development environments. They don’t make autonomous decisions. They assist humans in making better ones. Safety cannot be outsourced, but guided processes can yield better productivity, allowing for faster delivery cycles and more innovation.

It is important to recognize that tools used in the development of functional safety software may be subject to conformance standards to prevent potential inadvertent failures during the development process. As such, qualification may be required.

To help minimize the demands of tool qualification and support safe development for critical applications, organizations like Parasoft offer qualification kits for many embedded software tools. Looking ahead, Parasoft will offer a TÜV SÜD certificate for C/C++test CT with GoogleTest to support organizations that need to comply with various standards.

AI in Embedded Systems: A Cautionary Frontier

Deploying AI models, especially inference engines like neural networks, within embedded systems introduces a host of challenges that are not yet fully resolved.

Major Obstacles

Lack of Determinism

Functional safety demands predictable behavior. AI models, particularly deep learning systems, are inherently probabilistic. Their outputs can vary based on subtle input changes, making them difficult to validate under traditional safety frameworks.

Hallucinations and False Confidence

Large language models (LLMs) and other generative AI systems can produce outputs that sound plausible but are factually incorrect. In safety-critical contexts, this is unacceptable.

Opaque Decision-Making

AI models often operate as black boxes. Unlike rule-based systems, their internal logic is not easily interpretable. This lack of explainability undermines trust and complicates certification.

Absence of Standards

While tooling automation can be mapped to existing standards, the deployment of AI in embedded systems lacks a mature regulatory framework. Efforts are underway, but consensus and mandates are at varying stages of maturity across industries. See the FAA’s AI roadmap in avionics or emerging standards such as ISO 8800 in automotive.

Verification and Validation Complexity

Traditional V&V methods struggle to cope with the nonlinear, high-dimensional nature of AI models. Proving that an AI system will behave safely in all scenarios is a monumental challenge.

A Pragmatic Roadmap

Given these realities, the near-term strategy for organizations working in functional safety should be clear:

Embrace AI for Development Automation

• Invest in AI-powered tools that assist with testing, documentation, and compliance.
• Ensure outputs are traceable, explainable, and auditable.
• Use AI to augment—not replace—human judgment.

Proceed Cautiously With Embedded AI Deployment

• Limit AI use to non-safety-critical functions or tightly bounded applications.
• Explore hybrid approaches, such as rule-based systems with AI augmentation.
• Participate in standards development and pilot programs to shape future regulations.

Final Thoughts

AI is not a silver bullet, but it is a powerful tool. In the realm of functional safety, its greatest value today lies in enhancing the development process, not yet in replacing deterministic logic within embedded systems, though improvements in that domain are ongoing too.

By focusing on tooling automation, organizations can unlock productivity gains and accelerate innovation—without compromising the rigorous safety standards that protect lives.

The future of AI in embedded safety systems will come, but it must be built on a foundation of trust, transparency, and proven efficacy. Until then, let’s use AI where it’s strongest: helping humans build safer systems, faster.