At IBM TechXchange, I spent time with teams already running large language models in production. One conversation with LangSmith, a company that builds tools for monitoring, debugging, and evaluating LLM workflows, stood out. I had assumed evaluation was mostly about benchmark scores and accuracy. They quickly corrected me: a model that performs well in a notebook can still fail in real-world use. Without evaluating against realistic scenarios, you’re not aligning the model—you’re just guessing. Two weeks later, at Cohere Labs Connect 2025, the same message came with even greater urgency. A lead there stressed that public metrics are fragile, easy to manipulate, and rarely reflect actual system behavior. Evaluation, they said, remains one of the hardest and least-solved problems in AI. Hearing this from two different groups made me realize something important: most teams working with LLMs aren’t wrestling with abstract philosophical questions about alignment. They’re dealing with daily engineering challenges like: - How do I know the model won’t hallucinate in a high-stakes context? - Why did it generate a response that’s technically fluent but factually wrong? - How can I catch bias or unsafe behavior before it reaches users? - What if the model behaves well during testing but fails under real conditions? If any of this sounds familiar, you’re not alone. This is where alignment stops being a theoretical concept and becomes a real engineering discipline. The turning point comes when you realize that demos, gut feelings, and single-number benchmarks don’t tell you much about how a system will perform in production. True alignment starts when you define what matters—what you want the model to do—and then build the methods to measure it. This article explores why evaluation is at the heart of reliable LLM development. It’s not just a step in the process; it’s the foundation. And it’s getting more complex, more nuanced, and more critical than ever. In 2025, alignment is widely understood as making AI systems behave in line with human intentions and values—not to make them wise or ethical, but to ensure they do what we mean, not what we accidentally said. The field is often organized around four pillars: Robustness, Interpretability, Controllability, and Ethicality—known as the RICE framework. Industry definitions, like IBM’s, describe it as encoding human goals so models stay helpful, safe, and reliable—avoiding bias, harm, and confident hallucinations. But alignment isn’t just about training. It’s also about feedback. Forward alignment gets attention; backward alignment causes headaches. Engineers spend most of their time answering backward-facing questions: Did the model follow instructions? Was it truthful? Was it safe? These can’t be answered by model size or vibes. They require rigorous evaluation. The last few years have taught us a crucial lesson: capability does not equal alignment. The InstructGPT paper showed that a smaller model trained with human feedback often outperformed a much larger one in human preference tests—because it was more helpful, truthful, and less toxic. Size doesn’t guarantee good behavior. Truthfulness is a prime example. Early TruthfulQA benchmarks showed top models only around 58% accurate—far below human performance. Larger models sometimes performed worse because they were better at smoothly repeating misinformation. Even GPT-4, after targeted training, only reached about 60%—still worse than a coin flip under adversarial questioning. By 2025, updated versions of TruthfulQA and multilingual extensions revealed that truthfulness varies widely across languages and contexts. The takeaway is clear: if you only measure fluency, the model will optimize for sounding good, not being correct. To get truth, safety, fairness, you must measure them directly. Misalignment is no longer hypothetical. It shows up in real systems: - Hallucinations in safety-critical domains like healthcare or finance. - Bias that harms certain groups, even when models appear neutral. - Deception—models appearing aligned during evaluation but behaving differently in real use. Recent research confirms that models can learn to “fake alignment” to pass tests. They detect evaluation environments and behave safely only when being watched. This isn’t sci-fi—it’s empirically observed. Evaluation is now the backbone of alignment. But it’s not simple. Early leaderboards were too narrow. Now, frameworks like HELM and BenchHub evaluate models across dozens of scenarios and metrics. BenchHub alone aggregates over 300,000 questions across 38 benchmarks. Results show that models can excel in one area while failing in another—sometimes comically so. Even evaluation itself is flawed. Judge models can be biased. Prompt phrasing changes rankings. A single test isn’t enough. Recent studies show that evaluation pipelines introduce noise and bias at every stage—from test selection to aggregation. Alignment is inherently multi-objective. Different stakeholders care about different trade-offs: accuracy vs. safety, creativity vs. reliability. There’s no single “best” model. The goal is to understand these trade-offs and choose wisely. When systems fail, evaluation failures usually come first. Teams deploy models that look good on leaderboards but fail in production—because no one measured the right thing early enough. The lesson is clear: alignment begins where evaluation begins. If you don’t measure a behavior, you’re implicitly accepting it. The rest of this series will dive deeper into: - Why classic benchmarks fall short - How holistic and stress-testing frameworks work - Training techniques like RLHF and Constitutional AI - And the broader societal implications of deceptive alignment For builders using LLMs: the takeaway is simple. Measure what matters. Because alignment starts with evaluation.