I have been involved with INLG since 1989. Between 1989 and 2026, I have seen numerous scientific advances and our technology has improved beyond recognition. However, one thing which I suspect is worse in 2026 is NLP research culture. In 1989 all of us were motivated by science. In 2026, many researchers see papers primarily as CV enhancers, and are not interested in high-quality science and careful experiments; indeed some people cheat by getting LLMs to write “AI slop” papers. The NLP community is also less open to new ideas from new people. This is dangerous, and I hope the INLG community can take a lead in improving research culture!

Games are a quite special feature of human behaviour. They involve invented and voluntarily adopted rules and constraints, shared conventions, and - hopefully! - enjoyment of the activity for its own sake. Wordplay brings all this into language, leveraging many central properties of language and human communication that have historically been difficult for machines to deal with. Think of ambiguity, implicitness, and humour, for instance. This makes language games a great way to study the abilities of language models. We will play different games together, each tied to experiments we have run, where we see where and how models succeed and fail and what we can learn from this. I will then change perspective and look at (language) games not as evaluation tools, but as teaching tools in the context of AI literacy. Bringing together the playful and creative aspects of language (games), and the evaluation of language models, I will end with some broader considerations on language and models that I have been mulling over extensively (and probably excessively) in the last few months.

NLP researchers have increasingly acknowledged that humans often diverge in their interpretations of various NLP tasks, and that such variation should be captured if robust language understanding is to be achieved. In this talk, I will focus on analyzing label variation in the Natural Language Inference (NLI) task, in which, assuming the truth of a premise, one judges whether a hypothesis text is true, false, or undetermined. For instance, if one says: “My friend often travels with a heavy suitcase”, can it be inferred that “My friend often travels with a light suitcase”? I will examine the sources of NLI label variation through explanations provided by humans and LLMs, arguing that, in the presence of variation, labels without explanations are not meaningful and that we need better measures to capture similarities of explanations.

Training and evaluation for LLMs has become increasingly complex. Everything in daily life is now an open-domain generative task, and each task has many models acting as the generator, the simulated user, and the proxy evaluator, which together form a complex system. The majority of research presumes that the only important part of this system is the generator model, and optimise only that. I will discuss research, predominantly in the safety and security domains, but also for agentic models overall, showing why is is dangerous to overlook the other parts of the system. I will call for us as a field to have the courage to incorporate the complexity of the systems we build and experiment on.