Healenium, a 201-star open-source testing tool on GitHub, claims in its README to "leverage machine learning" so scripts can heal their own bugs. A Chinese engineer disproved that claim by decompiling the JAR: in the core weighting formula, id+50, class+40, tag+100—all hardcoded constants baked into the source, not a single line of machine learning code. We care about this case because it demonstrates the textbook playbook for spotting "AI washing" when enterprises buy software.

What this is

The most common reason Web UI automated tests fail isn't business-logic bugs—it's "can't find the button selector." When the frontend id or class changes between versions, tests break. Self-Healing Locators—automatically finding replacement selectors when tests fail—has been the engineering response over the past few years.

The most well-known open-source project in this space is Healenium (201 GitHub stars, last commit on 2026-09-07). A long-form Juejin article decompiled its JAR, read the core source, and verified the README's "leverage machine learning" claim. The conclusion is unambiguous: the core mechanism is not machine learning—it's a weighted scoring formula. id+50, class+40, attribute+30, tag+100—the weights are hardcoded constants in the source. No training data, no model inference, none of the basic machine-learning components.

The engineer offered an example: rename the login button from login-btn to signin-btn—the string edit distance is small, so it's judged the same element. But rename it to submit-auth (clearly the same button semantically) and the score drops sharply. Heuristic scoring measures surface string similarity, not functional semantic similarity—that's the ceiling.

Incidentally: Healenium added a real AI-calling endpoint in 2026 (POST /selectors/xpath), but it's locked behind a paywall and unavailable in the open-source edition.

Industry view

The technical community's reaction has two sides.

Supporters argue Healenium's heuristic scoring isn't a scam—it's fast, deterministic, explainable, and has zero external dependencies. It works for simple renames and structure tweaks. A mature project solving 80% of common cases through engineering alone has value, and its 2026 maintenance activity shows it remains the default tool.

Critics are sharper. The Juejin author argues "leverage machine learning" is fine as marketing but misleading as a technical claim—decision-makers are led to overestimate capability ceilings. One test engineer put it more bluntly: "Our team used Healenium for two years and recently realized it's not fundamentally different from the if-else statements we wrote ourselves."

A third path exists. MarketSquare/robotframework-selfhealing-agents is an open-source experiment in the Robot Framework ecosystem, taking a different approach—relocating elements by "what they're doing" rather than "where they sit in the DOM"—but its maturity is far below Healenium's.

Impact on regular people

For enterprise IT: When evaluating testing platforms, "AI self-healing" can't be taken at face value as a procurement criterion. Check whether vendor docs contain weighting formulas, training data, and model versioning. A black box promising high self-healing rates is most likely Healenium-style heuristics wearing a new label.

For individual careers: Test engineer roles are safe. True AI self-healing isn't yet at the point of replacing human maintenance, and the "heuristics vs. semantic understanding" distinction is itself a newly scarce skill—the people who can articulate tool boundaries clearly when introducing AI tools are the most valuable.

For the consumer market: This is yet another sample of "AI washing." SaaS vendors love packaging engineering tweaks as an AI revolution—and this isn't limited to testing. From code assistants to customer-service bots, expect more "decompiled and it's not AI" exposures over the coming year. That exposure game is itself a traffic business of its own.