Governance defines required AI behavior while quality engineering proves it; a gap exists between policy and code, widening as software moves faster than verification. A robust framework uses measurable acceptance criteria, aligns with NIST and EU rules, and requires continuous testing with clear ownership and verifiability to prevent unsafe releases.
Translated title. Original: Может ли ИИ писать автотесты за QA-инженера? Проверили на реальных задачах / Хабр
A SpaceWeb QA engineer reports on using a local language model to draft test cases, generate API test data, write automation scripts, and add security checks. The examples show where AI can save time, while QA engineers still review and adapt its output.
A QA community poll and discussion reflect job insecurity among testers, though the poll is not representative of the industry. The writer also weighs AI-based visual testing against its accuracy, upkeep, and token costs, arguing that employers should cover tools they require.
AI can speed software development, but businesses need evidence that intelligent systems act safely before granting them more autonomy. The author argues that quality assurance should assess business outcomes, test difficult cases and monitor AI after deployment, making it a board-level concern.
AI agents automate data validation, scan large datasets for gaps and duplicates, and spot inconsistencies. Continuous cross-system checks improve data accuracy, reducing errors in reporting, analytics, and business decisions.
Translated title. Original: Fachgesellschaften für Qualitätssicherung bei KI-Anwendungen mit medizinischen Inhalten – News – Deutsches Ärzteblatt
The article discusses the need for quality assurance and digital, interoperable medical guidelines to ensure AI-based health information is accurate. It notes an industry push, including the AWMF's Dissolve-E project, toward a structured digital knowledge base and highlights calls for transparency and accountability in AI health communication.
An open-source toolkit provides end-to-end evaluation for coding agents, including a CLI, MCP tools, and a workflow to test, compare, and improve Skills; it supports cross-platform installation and exports multi-format reports to detect regressions.
LongTail-SWE is a multilingual software issue-resolution benchmark with task environments, prompts, and patch-scoring tools. Its text describes how tasks are judged and repaired, reports improved scores on repaired tests, and notes limits on contamination checks and plans for a larger training release.
A tester creates a bug-filled fake online store to measure how AI QA agents discover UX issues under different prompts, using two AI models and automated browser testing. Results show that more detailed prompts don’t always improve outcomes, and only Level 3 revealed keyboard accessibility problems.
The paper presents Thinking-in-Bets, a training-free method where agents pre-write a private claim checklist before user feedback to preserve verified work when challenged; this approach improves accuracy and confidence across multiple models and tasks, reducing false reversals and unnecessary edits, with gains most evident when uncertainty triggers claim recording.