The five sub-categories of AI tester in 2026.

The category contains tools that read source code and produce unit tests directly. Two paradigms have emerged: reinforcement-learning search (occupied principally by Diffblue Cover for JVM languages) and large-language-model prompting (Qodo Cover, GitHub Copilot, Tabnine, JetBrains AI Assistant).

Diffblue's 2025 vendor benchmark study compared Cover against several LLM-based code assistants on Apache Tika and Spring PetClinic. The published methodology measured mutation score (the proportion of synthetic bugs the generated tests detect) and reported significantly higher scores for Cover than for the LLM assistants tested. The numbers are linked from the unit-test-generation deep-dive page (Diffblue, 2025).

The MuTAP paper (arXiv:2308.16557) studies how LLM prompting can be augmented to improve mutation score, providing a peer-reviewed datapoint on the LLM paradigm.

The category contains tools that ingest plain-English specs, user goals, or recorded sessions, then drive a real browser to validate the application. Output ranges from durable Playwright code (QA Wolf publishes its tests as standard Playwright, QA Wolf docs) to opaque LLM-managed flows where the test logic lives in vendor-managed prompts (testRigor, Momentic).

The lock-in trade-off is the published one: tools that emit Playwright or Selenium can be exported and continue to run if the vendor relationship ends; tools that store tests as proprietary YAML or LLM session blobs do not.

The self-healing category does not replace existing test suites. It augments them: when a primary CSS selector or XPath stops resolving, the runner falls back to alternative identifiers (text content, role, accessibility label, AI-described element fingerprint).

Mabl's documentation describes its auto-healing as based on multi-attribute element fingerprints (Mabl docs). Testim (now part of Tricentis) describes a similar approach (Testim docs). Functionize publishes its self-healing claim as part of its "Adaptive Locators" capability (Functionize). Each vendor's confidence threshold and fallback algorithm differ.

Visual regression is the older sub-category: capture a baseline image of a UI state, capture the current state, diff. Modern tools augment this with AI-tuned thresholds to suppress trivial pixel differences (animation frames, antialiasing).

Applitools publishes its "Visual AI" methodology (Applitools docs). Percy and Chromatic offer pixel-and-DOM diffing at the component level. Meticulous occupies a related but distinct sub-category: it captures real user sessions in production, replays them against a candidate build, and flags unexpected behavioural divergence (Meticulous docs).

The published trade-off across the category is false-positive volume: stricter diffs catch more regressions but also more inconsequential noise. Each vendor documents its tuning model. Meticulous publishes a dedicated page on managing and dismissing false-positive diffs (false-positive diffs).

This category is the oldest of the five and the most varied in shape. testRigor advertises plain-English step ingestion (testRigor docs). Tricentis Tosca with Vision AI offers ticket-driven test design. Functionize publishes a similar flow.

Output is candidate test cases, often gherkin-style, intended to be reviewed and refined by a human test engineer. The category remains useful for teams wanting to widen test coverage of new features quickly; the published failure mode is over-generation of low-value test cases that need pruning.