nuilab.org ↗

Natural User Interaction Lab · Colorado State University

How we use AI

Our current approach to the use of AI in the NUILab, and how we disclose it.

Why we disclose

In science, disclosure is extremely important. Space limits sometimes force us to choose what goes into a paper. But disclosing how we used AI is no different from reporting non-significant results or listing the tools we used. Yet how much is disclosed depends on the venue, the space limits and the task at hand. This page gives the overall approach to AI use in the NUILab.

What should be disclosed keeps changing. ACM’s early policy asked authors to disclose AI-generated content such as text, figures and code. Its current Policy on Authorship and FAQ also ask about AI used in research methodology and prototyping. We expect these expectations to keep shifting, and this page will change with them.

LLMs have real problems. We also believe smaller models and neuromorphic computing are more efficient for certain problems, and the lab works on both. Notwithstanding these critical challenges, LLMs are a useful tool, and that can’t be denied.

Tools we use

We use a variety of tools:

  • Frontier models, under terms that exclude training on our data: Anthropic’s Claude, OpenAI’s GPT models (through Codex) and Google’s Gemini.
  • Open-weight LLMs run locally: Qwen3 (including Qwen3-Coder), gpt-oss, GLM-4.7, Llama 3.3, Gemma, Phi-4 and DeepSeek-R1.
  • Vision and language models run locally: Qwen2.5-VL and Llama 3.2 Vision.
  • Retrieval run locally: BGE-M3 embeddings and reranking, plus GROBID for parsing papers, in our own retrieval (RAG) system.
  • Speech run locally: Whisper and Parakeet for transcription, and Kokoro for text-to-speech.
  • Images and 3D run locally: SDXL and FLUX for images, Hunyuan3D and TRELLIS for 3D objects, and Gaussian splatting for scenes.
  • Perception run locally: YOLO, MediaPipe and RTMPose for detection, pose and gesture.
  • Neuromorphic: spiking neural networks.
  • Additional capacity: when a job needs more than we have locally, we can rent GPUs, or call APIs from US-based vendors and through CSU’s Microsoft Azure agreements.

Research

We use AI to verify statistics, to learn about new topics, to bring published papers into our local retrieval (RAG) system, and to help build prototypes or build them outright, among other uses.

Writing

We use AI to revise drafts, to draft messages, to revise documents and to review our work, among other uses. Some of our websites and online content are also built or written with LLM assistance. We always follow the rules of the venue.

Authorship

The authors are responsible for any LLM-assisted work. AI use is disclosed according to the venue’s requirements.

Students

No one in the lab is required to use AI. Students in the lab follow these guidelines.

Updates

This page may be updated infrequently. Contact us for more information.