AI-FIT: The Theoretical Foundation of Statistical Functional Safety

Over the past decade, artificial intelligence has transformed how we design and operate autonomous driving systems. Early vehicles relied on rule-based software — deterministic code that followed predefined conditions. Then came the era of machine learning, where pattern recognition replaced hand-coded logic.

But in the last years, something entirely new has arrived: transformer-based foundation models. Originally built for language and vision tasks, these large models are now finding their way into the vehicle stack. They are capable of end-to-end learning, scene understanding, reasoning, and even multi-modal fusion — tasks that previously required multiple handcrafted modules.

This rapid transition from modular AI to foundation-model-driven autonomy is rewriting not only how cars perceive and decide, but also how we must think about safety.


The Changing Nature of Safety

Traditional safety engineering — as defined in standards like ISO 26262 or IEC 61508 — is built around deterministic logic. Failures are typically systematic or architectural: a wrong line of code, a missing signal, a hardware short. Every failure can be traced, analyzed, and mitigated through established methods such as FMEDA, FTA, or static code analysis.

AI systems don’t fit that world anymore. Their “logic” doesn’t live in code lines; it lives in model parameters, data distributions, and training dynamics. A model may fail not because a line of code is wrong, but because a rare scenario was missing in the data, or because the model made a statistically uncommon but unsafe inference.

Current safety frameworks such as ISO/IEC TR 5469 focus primarily on AI components based on machine learning or deep learning recognition modules, where safety assurance depends on the quality and labeling of datasets. While these approaches provide valuable structure for perception-oriented AI, they remain insufficient for transformer-based foundation models that exhibit emergent reasoning and autonomous decision-making behaviors. These models go beyond labeled recognition—they generate, interpret, and act upon complex, context-dependent scenarios that cannot be fully validated through conventional dataset-based methodologies.

This creates what we can call a failure model duality:

  • AI modules still have a systematic component (architecture, software interface, hardware execution).
  • But the dominant risk increasingly comes from probabilistic behavior — the inherent uncertainty of model inference.

As models gain autonomy, the proportion of probabilistic failure grows.
And that’s where today’s safety frameworks start to struggle.


Why Validation Has Become the Bottleneck

For autonomous driving, validation is no longer about verifying individual functions. It’s about assessing a complex, data-driven behavior that evolves with every model update. In traditional software, functionality and logic are explicitly defined by code — each condition, rule, and decision is determined by human engineers through deterministic programming. Safety analysis frameworks like ISO 26262 were built upon this paradigm: logic was transparent, traceable, and verifiable line by line.

Modern AI modules—particularly large-scale models—no longer derive their functional logic directly from code. Instead, their behavior emerges from data, model structure, and learned parameters. The designer no longer programs “what to do,” but instead shapes the conditions under which reasoning develops: the size, diversity, and distribution of the data; the architecture of the model; and the objectives encoded in its training.

Each AI inference represents a new probabilistic event, shaped by unseen data and contextual uncertainty. Failures can’t be traced to a single line of code or a single dataset — they emerge from the interaction of parameters, data, and environment. That means validation must shift from deterministic verification to statistical evidence.

This shift from coded logic to learned behavior fundamentally challenges how we validate safety.
In traditional systems, validation meant demonstrating that each function performs as intended under all specified conditions. Test cases were finite, repeatable, and traceable to clear requirements.

In contrast, validation of large-scale AI systems is open-ended. The space of possible inputs—the combinations of sensor signals, traffic situations, weather, and interactions—is effectively infinite. Because model behavior depends on statistical generalization rather than explicit rules, even a small shift in data distribution or unseen scenario can lead to unexpected outcomes.

Moreover, every model update—whether retraining, fine-tuning, or adding new data—changes the internal decision boundaries in ways that are opaque and nonlinear. This means previous validation results may no longer hold. Traditional regression testing can’t fully capture these shifts, as the “logic” being tested has itself evolved.

Simulation and large-scale data evaluation offer partial solutions, but they introduce their own challenges:

  • Simulations must represent the diversity and uncertainty of the real world.
  • Statistical confidence requires massive sample sizes, far beyond what physical testing can provide.
  • Rare but safety-critical edge cases dominate risk, yet are the hardest to generate or validate.

As a result, validation is no longer a binary task (“pass/fail”), but a probabilistic measurement—estimating how often the system will behave safely across billions of inferences in open-world conditions.


From Functional Safety to Statistical Functional Safety

In conventional functional safety, hardware reliability is expressed through the metric FIT (Failures In Time) — the number of random hardware failures expected per billion hours of operation. It provides a probabilistic measure of failure exposure that complements systematic design assurance.

For AI, we adopt and extend this concept rather than replace it. Modern AI components—especially those used in autonomous driving—exhibit a duality in safety characteristics: they behave like software, with systematic risks linked to design, training, and data; but they also act as logic elements, producing probabilistic outputs whose reliability can be assessed statistically over time.

AI-FIT (AI Failures in Time) generalizes the traditional FIT concept to this new domain. It quantifies the expected number of safety-relevant erroneous inferences occurring during a defined operating period.
In other words, instead of counting transistor faults, we measure how often an AI model produces a critical misjudgment—within the same probabilistic reliability framework that underpins ISO 26262 and IEC 61508.

This analogy allows functional safety engineers to evaluate AI components with familiar metrics, bridging the gap between hardware reliability methods and data-driven, probabilistic AI behavior.

That’s the idea behind AI-FIT.


What Is AI-FIT?

AI-FIT (AI Failures in Time) is a reliability metric that extends the classical FIT (Failures In Time) concept from hardware to artificial intelligence systems. It quantifies the expected number of safety-relevant erroneous inferences produced by an AI model during operation.

Unlike hardware FIT, which measures random physical faults over time, AI-FIT captures functional failures that arise statistically across many inference or decision-making events. It reflects the probabilistic nature of AI reasoning, where each output depends on data distributions, learned parameters, and contextual uncertainty.

Modern AI components exhibit a dual safety character: they behave like software, where systematic risks stem from design and data;
yet also as logic elements, whose reliability can be expressed through probabilistic failure rates. AI-FIT bridges these domains—connecting the deterministic framework of ISO 26262 and IEC 61508 with the statistical behavior of large-scale foundation models.

Because inference frequency can be estimated (e.g., 10 Hz), AI-FIT values can also be converted into time-based metrics, ensuring compatibility with existing functional safety architectures while enabling quantitative assessment of AI reliability.


Extending FMEDA for AI

FMEDA — Failure Modes, Effects, and Diagnostic Analysis — is central to ISO 26262.But to analyze AI systems, FMEDA needs to evolve.

In an AI context:

  • Failure Modes might include misclassification, hallucination, or incorrect scene interpretation.
  • Effects could range from reduced perception confidence to unsafe maneuvers.
  • Diagnostic Coverage (DC) becomes the measure of how well we can detect or mitigate these failures using uncertainty estimation, redundancy, or runtime monitors.
  • Evidence comes from large-scale simulation, fleet data, and targeted experiments.

This “AI-aware FMEDA” allows safety engineers to integrate AI modules into existing system-level analyses, using familiar processes but extended for probabilistic behavior.


The Role of generative Simulation (genSim) and Statistical Validation

BecBecause real-world driving can never expose every critical edge case, simulation becomes indispensable. Traditional simulation — often built on game-engine pipelines — focuses on rendering environments, physics, and agent interactions for functional testing. However, such approaches remain limited by pre-defined assets and deterministic scenario scripting, which cannot fully represent the statistical diversity or uncertainty that AI-driven systems must handle.

Generative Simulation (genSIM) represents a paradigm shift. Instead of replaying fixed scenes, it leverages high-quality real-world data distributions and generative modeling to synthesize new, statistically consistent, and logically valid scenarios. These simulations do not merely visualize; they reason — creating data that challenges AI systems under rare, high-risk, or emergent conditions at scale.

Through generative simulation, we can:

  • Estimate AI-FIT statistically across millions of diverse virtual inferences
  • Quantify diagnostic coverage of self-checking or ensemble mechanisms
  • Validate safety concepts under controlled but probabilistically rich conditions
  • Generate traceable evidence for safety audits and certification processes

In short, genSIM transforms simulation from testing to statistical safety validation — turning abstract safety concepts into measurable, data-driven evidence. We will explore this concept further in a dedicated post on Generative Simulation: Data Reasoning for AI Safety.


Building the Broader Framework

AI-FIT is one piece of a larger puzzle — a structured framework for AI functional safety in autonomous driving. We can think of it as a set of layers working together:

  1. Safety Concept Layer – Define the intended function, operational design domain (ODD), and high-level safety goals.
  2. Reasoning Layer – Model the logic of decision-making and formalize safety reasoning for AI behavior.
  3. Self-Validation & Adversarial Layer – Include uncertainty estimation, online self-checks, and robustness testing.
  4. Data & Simulation Layer – Use generative simulation, proving grounds, and fleet data as continuous safety evidence.
  5. Experiments & Best Practices – Establish pilot projects to validate the AI-FIT framework in real systems.

Together, these layers create a path from theory to practice — from defining AI reliability to proving it statistically.


Why AI-FIT Matters

AI-FIT doesn’t replace functional safety — it extends it. It gives engineers and safety assessors a shared quantitative language to describe AI reliability.
It allows existing ISO 26262 processes — like FMEDA, SPFM, and PMHF — to evolve instead of being discarded.

As the automotive industry moves toward L3 and L4 autonomy, and as foundation models become central to perception, reasoning, and planning, the ability to quantify AI safety will define the next decade of automotive innovation.

AI-FIT is a small concept with a big ambition: to make AI safety measurable, traceable, and certifiable.

AI-FIT: Statistical Functional Safety for Foundation Models

← Back

Thank you for your response. ✨

Rate your experience(required)

Leave a Reply