Thread: I'm still here .. kind of!

Hello All,

just a quick note to let everyone know that I am still functioning okay. Plus, my ownership and interest in Land Rovers and this forum is still active. I have just fallen into the deep depths and grips of artificial intelligence assistants as it turns out they can actually assist me in my research. So, I have moved from knowing pretty much nothing about them to on sometimes using six or more of them for validation and cross-checking. Plus, I have written some posts on LinkedIn that were based unfortunately on my own lived experiences of the pitfalls of AI.

I think I unintentionally followed the advice of learning a new language - the immersive approach works the best. Perhaps I might have even plumbed the depths too far. I am frequently getting messages from AI that I am stretching it beyond the capabilities it was designed to operate within.

Did I mention that, 'Perhaps I might have even plumbed the depths too far.' No half measures taken here ... tee-hee.

So, in closing - I will emulate Young Mr Grace by waving and saying."you are all doing very well".

Take care one and all, and stay well.

Kind regard
Lionel

Hello All,

Why I finished my earlier post with 'sometimes!!!'

There are some things to be aware of while playing with AI assistants - the Large Language Model (LLM) makes its platform unstable. ... This can mean that AI assistants can inaccurate and even potentially reputationally hostile to use ...


How drift, sycophancy, and fake citations can quietly undermine academic validity—even when you avoid plagiarism


In research, most of us worry about plagiarism, but several subtler traps can damage academic validity just as much: drift in AI systems, sycophantic reasoning, and AI‑generated pretend citations that look like real APA 7th‑style references (Dokter et al., 2025; Sebastian et al., 2025; SwanRef, 2022).


For AI‑assisted researchers, drift is more than mere “model decay.” It includes concept drift (meaning and relationships shift over time), data drift (input distributions change), label drift (class proportions evolve), feature drift (individual predictors lose predictive power), and model drift (overall performance degrades) (Lumenova AI, 2025; Tencent Cloud, 2025; Seo et al., 2023). Recent work on natural context drift further shows that as real‑world text evolves, large language models can lose accuracy on the same kinds of questions even when the underlying information has not changed (Wu et al., 2025).


Separately, sycophancy in research can take several forms: uncritical deference to authorities or prevailing theories, or AI‑driven flattery that always agrees with the user’s preferred narrative while suppressing counter‑evidence (Dokter et al., 2025; Sebastian et al., 2025). In AI‑assisted workflows, “AI sycophancy” leads models to selectively cite supportive sources, amplify overconfidence, and downplay limitations, which can erode objectivity even when the output is technically original (Sebastian et al., 2025; Seo et al., 2023).


An even more insidious problem is that some AI assistants can very convincingly generate fake

APA‑style citations—in‑text references and full reference‑list entries that look perfectly formatted in APA 7th style but refer to papers that do not exist (SwanRef, 2022; SwanRef, 2023; Citetrue, 2023). These “ghost” or “hallucinated” citations often mix real‑sounding authors, journals, and DOI‑like strings, making them hard to spot without independent verification against library databases, Google Scholar, or CrossRef‑based tools (SwanRef, 2022; GPTZero, 2024; FIU Library, 2023). Because they mimic proper APA‑7 formatting almost flawlessly, they can slip into essays, theses, or even manuscripts, creating the illusion of solid evidence while quietly misrepresenting the literature (SwanRef, 2023; FIU Library, 2023).


This matters because AI‑generated fake citations are not just a formatting or technical issue; they count as a form of academic misconduct or integrity failure, even if the surrounding text is original and well‑cited (Reddit, 2025; FIU Library, 2023). When reviewers or editors cannot locate the cited works, it can lead to rejection, revision requests, or reputational damage (SwanRef, 2022; FIU Library, 2023). In AI‑assisted writing, this is effectively a plagiarism‑adjacent integrity trap: you may not be copying text, but you are misrepresenting the existence and relevance of the evidence (SwanRef, 2023; FIU Library, 2023).


To protect academic validity, researchers using AI assistants should:

• Never copy‑paste AI‑proposed references without independently verifying each entry in a trusted bibliographic source.
  
  
• Use citation‑checking tools that flag “ghost references” by cross‑checking against large citation databases.
  
• Treat the AI as a drafting and ideation aid, not a citation authority, and, when required, clearly declare AI use in methods or acknowledgments (Grammarly, 2026; FIU Library, 2023).
  

In short, drift, sycophancy, and AI‑generated pretend citations are all potential traps that can quietly undermine academic validity—even when you are careful to avoid plagiarism. A robust, AI‑assisted research practice should combine vigilance against drift types, active resistance to sycophantic reasoning, and strict verification of every AI‑proposed citation.

---

References
Citetrue. (2023). Verify citations free forever! CiteTrue - Citation Checker
Dokter, G., Ravizza, J., et al. (2025). Shoggoths, sycophancy, psychosis, Oh My: Rethinking large language models. Journal of Medical Internet Research, 27, e87367. Journal of Medical Internet Research - Shoggoths, Sycophancy, Psychosis, Oh My: Rethinking Large Language Model Use and Safety
FIU Library. (2023). AI plagiarism + citation. Florida International University. https://library.fiu.edu/ai/plagiarism
Grammarly. (2026). More AI writing tools. https://www.grammarly.com/ai-detector
GPTZero. (2024). AI source finder – Check citations from text, essays & more. https://www.gptzero.me/sources
Lumenova AI. (2025, September 18). Model drift: Types, causes and early detection. https://www.lumenova.ai/blog/model-d...-introduction/
Reddit. (2025, August 14). Checking citations and references to counter AI use? https://www.reddit.com/r/Professors/...to_counter_ai/
Sebastian, J., et al. (2025). Sycophancy as compositions of atomic psychometric traits (arXiv preprint). arXiv:2508.19316. https://arxiv.org/abs/2508.19316
Seo, D., et al. (2023). Towards understanding sycophancy in language models (arXiv preprint). arXiv:2310.13548. https://arxiv.org/abs/2310.13548
SwanRef. (2022, December 31). ChatGPT fake references checker – Verify AI citations. https://www.swanref.org/chatgpt-fake-references
SwanRef. (2022). AI hallucination detector for citations – Free tool | SwanRef. https://www.swanref.org/ai-hallucination-detector
Tencent Cloud. (2025, January 19). What are some factors that contribute to AI “drift”? https://www.tencentcloud.com/techpedia/100192
Wu, Y., Schlegel, V., & Batista‑Navarro, R. (2025). Natural context drift undermines the natural language understanding of large language models (arXiv preprint). arXiv:2509.01093. https://arxiv.org/abs/2509.01093

---