Jakob Nielsen's February 2026 UX Roundup examines how heuristic evaluation is evolving as AI enters both the products we evaluate and the evaluation process itself. The findings reshape how we think about usability standards in AI products.
Heuristic evaluation has been a cornerstone of UX practice since Nielsen and Molich published the original 10 heuristics in 1990. The method is fast, relatively inexpensive, and remarkably effective: expert evaluators inspect a product against a set of usability principles and identify violations. Simple in theory, powerful in practice.
AI is now putting pressure on this method from two directions simultaneously. First, AI-powered products introduce behaviors the original heuristics do not fully address. Second, AI tools are beginning to participate in the evaluation process itself — raising real questions about what changes and what stays the same.
The 10 heuristics remain valid. But AI products demand an 11th dimension: the quality of AI judgment and how well the system communicates the limits of that judgment.
— Jakob Nielsen — UX Roundup, February 2026
Where Classic Heuristics Fall Short with AI Products
Nielsen's roundup identifies specific gaps:
- Error prevention (Heuristic 5): The original principle focuses on preventing user errors. AI introduces a new category: AI errors that users must detect and correct. The heuristic needs to extend to helping users catch AI mistakes — not just their own.
- User control and freedom (Heuristic 3): Classic implementation: undo and redo. In AI systems: the ability to inspect, adjust, or reject AI decisions. The principle is the same; the implementation surface is significantly more complex.
- Visibility of system status (Heuristic 1): Traditionally covers loading states. For AI: covers what the system is currently doing, what data it is using, how confident it is, and whether processing is happening on-device or in the cloud.
AI-Assisted Heuristic Evaluation: What We Know So Far
Several teams have begun using AI tools to run preliminary heuristic scans before human evaluation. Early findings are mixed but instructive: AI tools catch a higher volume of violations faster, but expert evaluators are still better at identifying the most severe issues — particularly those involving mental model mismatches and contextual appropriateness.
The emerging best practice: use AI for breadth (finding the long tail of minor violations quickly) and human experts for severity judgment (deciding what actually matters and why).