THE IMPACT OF LARGE LANGUAGE MODELS ON THE VALIDITY OF THE WRITTEN LABORATORY REPORT AS AN ASSESSMENT TOOL FOR LEARNING OUTCOMES IN HIGHER EDUCATION

Written laboratory reports remain a principal assessment instrument in distance education, yet the growing capability of large language models (LLMs) undermines the tacit assumption on which the validity of such reports rests: that writing the report requires first solving the task. This article examines whether leading current-generation LLMs, given a minimal prompt, can fully complete typical laboratory assignments of a university Databases course at the level of the highest student grade under the course's unmodified rubric. In a pilot case study with comparative benchmarking, eleven models from five providers (Anthropic, OpenAI, Google, xAI, DeepSeek) were tested in their native chat interfaces on authentic assignments from a live course: closed tasks with deterministic outcomes and an open-ended experiment-design laboratory. All anonymised responses were rated manually by three independent raters. The current generation handles the content confidently: median correctness is 91.6%, four models reach ≥ 98.9%, and over half of all content scores fall within the top rubric band. However, solving the task and producing a submission-ready report proved to be distinct capabilities: every model's formatted MS Word report was rated below its content correctness, and two models returned no usable file at all. The remaining barrier is technical rather than cognitive: an agentic tool with access to the course files generated complete reports with genuine execution output for all 11 assignments of a semester in a single run. We conclude that, for the content component, the unsupervised written report no longer measures the student's skill. The defensive barriers of the written report examined in the study are ordered by their observed resilience: from those already fallen (the creativity of the open-ended task, assignments on the student's "own data") through those holding only against an isolated chat (linked assignments, the complex report template) to a potentially durable class of artefacts external to the model, which rests not on the models' cognitive limits but on their lack of authority over external state, and therefore does not expire with the next model generation.