Can ChatGPT Create a School Timetable?

Before judging whether a chatbot can do this, it helps to know the bar. A school timetable has to satisfy hard constraints before it counts as usable at all: no class or teacher double-booked in a period, every class scheduled its required number of meetings, teachers scheduled only when available and within workload limits, and rooms matched to capacity and specialist-room needs, with commute time honored across multi-site schools. On top of that floor sit soft constraints, lesson-time and teacher-period preferences, spreading lessons across the week, limiting idle periods, that trade off against each other rather than all being satisfiable at once.

A practitioner account of running secondary-school timetables adds further trade-offs to the same list: spreading teachers across key stages instead of siloing them into one band, avoiding split classes where possible, and weighing room-allocation stability against overall utilization. That is a lot of interacting rules to hold at once, not a single question with one right answer.

On the OpenAI community forum, a user building a ChatGPT-based school-timetable generator reported that the model could not resolve teacher double-booking conflicts across grades, and asked how to train it for the task. A second poster on the same thread reported hitting the identical problem, with no fix offered. Both describe the same failure: the model produces an assignment but does not catch the conflict it just created.

That single report lines up with a wider trajectory in planning research. None of the studies below were run on school timetabling itself, but all of them test the same underlying skill a timetable also demands: generating a plan that satisfies many constraints at once.

Karthik Valmeekam and colleagues, in a NeurIPS 2023 spotlight paper testing LLMs on commonsense planning tasks built to resemble International Planning Competition benchmarks, found that the best model tested, GPT-4, reached only about a 12% average success rate across the domains evaluated. A year later, Jian Xie and colleagues introduced TravelPlanner, a real-world planning benchmark built around a large sandboxed toolset, and found GPT-4 acting as an agent completed just 0.6% of its planning tasks successfully, with the failures traced to agents losing track of multiple simultaneous constraints, struggling to stay on task, and misusing the tools available to them.

The most recent data point comes from testing a newer reasoning model. Karthik Valmeekam, Kaya Stechly, and Subbarao Kambhampati evaluated OpenAI's o1 on PlanBench and found it solved 97.8% of standard block-stacking planning problems, against 62.6% for the best plain LLM tested. That looked like a real jump, until the same problems were relabeled with unfamiliar terms to block pattern-matching from training data: o1's accuracy fell to 52.8%, and to 37.3% under a further randomized relabeling. On larger problems needing 20 or more solution steps, accuracy dropped to 23.63%, and o1 correctly identified only 27% of genuinely unsolvable problems as unsolvable. The same paper priced a classical, non-LLM planner against o1: solving the dataset cost $42.12 per 100 instances with o1-preview and $3.69 per 100 with o1-mini, while the classical planner solved it at close to zero cost in a fraction of a second per instance, with a correctness guarantee the LLM-based approaches don't carry.

None of these three studies benchmarks a chatbot against a school timetable's specific hard and soft constraints. What they measure is the general skill a timetable also demands, holding many constraints correct at once in a single pass, with no way to check the answer before handing it back, and on that general skill the pattern across three separate research groups and three different task types matches the one report of someone trying the real thing.

Valmeekam and colleagues' 2023 paper also tested a different setup, sometimes called "LLM-Modulo," where the model does not plan alone: it feeds heuristic suggestions to a sound external planner, gets feedback from a verifier, and is prompted again for an improved plan. That paired setup performed better than asking the model to plan on its own, with the model's suggestions shown to improve the underlying planner's search. It is also why dedicated solver-backed methods, among them simulated annealing, genetic algorithms, tabu search, and constraint programming, remain their own field of study rather than being replaced by prompting. See how a solver-backed system builds a conflict-free timetable step by step in the automatic timetable generation guide.

None of this makes a chatbot useless near a timetable. For a school with a simple structure and few option blocks, the output is small enough to check by eye in a couple of minutes, and the cost of an unverified mistake is low. That is a different situation from a secondary or vocational timetable with option blocks, split classes, and staff working across sites, where every constraint has to hold at once and a chatbot has no way to confirm that it does.

See how solver-backed automatic timetabling checks those constraints in practice: explore school timetabling capabilities.