The year TNC went robot

TNC26 is not the first year that we have had submissions about Artificial Intelligence (AI), but it is the year where we have seen a significant rise in the number of submissions referencing and talking about AI and Large Language Models (LLMs). It is, however, the first year where it was really obvious that a lot of TNC submissions were written by, or with the help of, AI – but is this a problem? As the Programme Coordinator for TNC, I want to make sure that we choose presentations that are well thought through and well crafted, but I also recognise my privilege as a native English speaker and the levelling effect that AI can have in writing proposals. I am also a bit of an AI luddite, so reached out to two people within GÉANT to help me think about how to tackle this issue – Daniela Brauner, the GÉANT AI coordinator and Rupert Holland, one of our Technical Authors and a specialist in how to write effectively for our community.   

Recognising the AI tells

One of the biggest issues we saw in submissions is how similar a lot of the writing was: nearly identical titles, same bulleted lists with highlighted sections, and familiar phrases and wording. For the TNC Programme Committee (PC), this was an immediate turn-off and made it difficult to differentiate one submission from another. I can also see the challenge for the individual submitters: the patterns only become obvious when you see the proposals all side by side. Is this something you recognise from your work with AI, and what can we do to improve the submission process?  

Daniela: You are referring to a structural effect of how LLMs operate. They are trained on existing text, so they naturally reproduce statistically common phrasing and familiar structures. When users prompt them for similar topics and in similar ways – i.e. when submitting for the same TNC26 Call for Proposals – the outputs will converge in structure and even have similar takeaways. These systems are not designed to think or generate original ideas in the same way humans would. They recombine patterns they have learned, which is why so many submissions may end up sounding the same. We should stimulate more original ideas for real problems. One way of doing this is including “originality” as one of the criteria for acceptance and by encouraging people to think out of the box rather than producing generic, template-driven proposals. For example, asking “What existing issue in the research and education community does your proposal aim to address?” could help spot human originality. For non-native speakers, AI can be a genuinely useful tool to help structure and correct language, but I agree that it is key to self-regulate and use it to refine your own original ideas, rather than to generate those ideas for you.  

Rupert: As you note, Nicole, the distinct patterns of AI-generated output become easy to spot with familiarity. From the obligatory em-dashes (—) and overuse of bullet lists, to particular cliches like “delve into”, once you know what to look for, you notice it everywhere. LLMs are a tool, ultimately, and like all tools, they can be used for good or bad. I suggest treating AI output only as a starting point – it solves the problem of staring at a blank page, not knowing where to begin, and gives you a basic structure to work with. But if you produce your own text and run it through AI for a final polish, the resulting text will be stripped of your own creative fingerprint and, as Daniela described, flattened into the most statistically likely string of words, in a process people in the AI space have named ‘semantic ablation’. For anyone in the GÉANT community in need of assistance when developing any kind of written material, feel free to reach out to the Technical Authors team. It’s what we’re here for. 

Bias in AI

Bias in AI processes is a concern we face when considering the ethical use of AI. As part of my work on the programme, I was curious as to whether Copilot would be able to make fair decisions and effectively replace the human members of the PC. I fed the titles of the Lightning Talks into Copilot and gave it access to a list of already accepted talks and asked it to choose a fair and balanced selection of 32 talks based on the conference theme and topics. Copilot chose 24 talks on AI and only 6 others. Apart from not being able to count, the PC review made it very clear that this selection was not at all representative of the 120 submissions we received for Lightning Talks this year. What challenges does this bias present in our day-to-day use of these tools?  

D: A LLM can assist with sorting, summarising and translating, but it cannot think or make genuinely fair, value-based judgments, so it should not be used as a decision-maker in contexts where balance and representation matter. It can be used to filter, summarise, identify missing points from a given list, but under human oversight is key. The LLMs tend to amplify dominant patterns in the data, as described in this scenario, the over-prioritisation of trendy topics like AI. Also, the lack of true semantic understanding in LLM responses reinforces the need to use AI as a support tool for analysis, not as a replacement for human judgment, especially in processes involving values, diversity, and collective responsibility. You can train a model to assist on judgement, but you will need a really good and balanced training dataset to try to minimise any strong bias.  

Detecting AI

Working out that a piece of writing is AI generated is still very subjective, and one of the only tools we have to analyse talks is – to use AI. There are many tools available that claim to detect AI generated text, but these generate a lot of false positives. Copilot can be very insistent on the challenges of detecting AI, and anyone who has studied in recent years knows how challenging automated systems such as Turnitin can be in effectively recognising generated or plagiarised text. The PC does however have concerns about how we can be sure that a machine written proposal can translate into an effective talk on stage. Is this something we should be worrying about and what measures should we take?  

R: As touched on above, LLMs do not ‘know’ or ‘understand’ anything; they generate output based on probabilities learned during their training. This could lead to proposals that look good on paper, but translate into rather formulaic talks with flat structure, and don’t feel like they’re building towards anything. AI may also use overly logical transitions, moving from one point to the next mechanically, instead of allowing the speaker to carry the audience on a journey with them emotionally. The most compelling TNC talks have rhythm, tension, momentum, a turning point, a resolution. To gauge how proposals may translate on stage, the PC could consider asking potential speakers to make a short-form video, 30-60 seconds long, giving a quick informal summary of their proposal to the camera. 

Finally, what advice would you give to the TNC PC and TNC submitters for TNC27? 

R: The PC could consider focusing on the narrative shape of proposals, even adding a question to the submission form asking submitters to describe the journey they will take their audience on, alongside the idea of a short-form video summarising their talk.  But like it or not, LLMs are not going anywhere. For those considering submitting a proposal for TNC27, using AI to produce a basic outline is a great way to get inspiration flowing, instead of staring at a blank page. Our community is full of experts committed to their fields of work, with so many exciting and groundbreaking stories to share – so be careful that your voice, your personal experiences and your authenticity aren’t drowned out by AI. As you build your proposal, think about what you want your audience to feel, not just to learn. 

D: We should design the process to spot human originality, creativity, and real cases, rather than trying to detect AI. We can adapt the submission requirements and evaluation process to value real experiences and creative discussion. Submissions should make genuine insight, experience, and opinion easier to see. The PC can focus on recognising positive experiences in the community, based on evidence.