A new academic position paper argues that AI reasoning — the thing everyone claims their model can do — has never had a proper definition, and that gap is quietly making AI progress unmeasurable.
The Dirty Secret Behind AI Reasoning Benchmarks
When a company says their model "reasons better" than the last one, what does that actually mean? According to researchers publishing on arXiv, nobody has agreed on a definition — and that's a serious problem for learnable AI reasoning evaluation.
Without a shared operational definition, benchmark results for reasoning tasks are essentially unverifiable. You can't measure progress toward a goal you haven't defined. It's like running a race where nobody agrees where the finish line is.
Symbolic Logic Meets Deep Learning — Finally
The paper draws a sharp line between two traditions: old-school symbolic AI (think formal logic, rule-based systems, verifiable proofs) and modern deep probabilistic models (think GPT, Claude, and friends). The authors argue the generative AI community has largely ignored the rigorous groundwork symbolic AI laid down.
Their fix is practical: define reasoning as a learnable, rule-based process where outputs are valid and sound — terms borrowed directly from formal logic. A valid argument follows correct rules. A sound argument is also built on true premises. Both are checkable. Both are teachable.
They also publish a checklist for AI researchers to follow when reporting reasoning results — a small move that could meaningfully clean up how the field communicates progress. If you want to understand how neural networks actually process and represent these rules under the hood, our course How Neural Networks Really Work is a strong starting point.
What This Means for Learners
If you're building with AI agents or evaluating AI tools, this paper hands you a sharper lens. "Does this model reason?" is the wrong question. The right questions are: Does it follow verifiable rules? Can its conclusions be checked? Is its process transparent enough to trust?
Understanding the difference between a model that pattern-matches and one that genuinely applies learnable rule-based reasoning will separate savvy AI users from everyone else — especially as autonomous agents take on higher-stakes tasks. Our course on AI Agents covers exactly how to evaluate and design agent behaviour with this kind of critical eye.
The bottom line: next time an AI vendor claims their model "reasons," ask them to define the term. If they can't, that tells you everything.
