phd-thesis/ai-llm-use-disclosure.tex
2025-10-29 18:50:06 +01:00

63 lines
5.2 KiB
TeX

\chapter*{Use of Artificial Intelligence in This Thesis}
\adjustmtc
\addcontentsline{toc}{chapter}{Use of Artificial Intelligence in This Thesis}
This thesis has been written during the years of 2020 - 2025. In this time, Artificial Intelligence (AI) technology
including Large Language Models (LLMs) has entered widespread adoption. I have used such LLM systems in the preparation
of this thesis. At the time this thesis was written, LLMs were a powerful and useful technology, but often produced
wrong output. Thus, I used the following list of observations to guide my LLM use during the writing of this thesis.
\begin{enumerate}
\item Passing text through an LLM is an imprecise operation. Especially when large amounts of text are passed
through an LLM, despite clear instructions such as ``only fix spelling errors'', the LLM output might deviate
from the source text. Therefore, the document text should never be passed through the LLM, and the LLM should be
prompted to point out problems, or to produce a list of suggestions for improvements instead.
\item LLMs are really bad at summarizing text that contains novel concepts. LLM summaries of text often converge to
a re-stating of the general consensus on the text's main topic. Where the source text deviates from conventionla
wisdom or makes novel points, an LLM summary will likely mis-represent those conclusions. Additionally, LLMs are
bad at capturing the point of a text. Unless extreme care is taken when prompting, it is easy to lead an LLM to
produce an inaccurate summary of a text that agrees with the prompt, but misses the gist of the text. Therefore,
extreme caution should be applied when using an LLM for summarization, and LLM output should be checked
diligently in such instances.
\item LLMs are bad at generating text from scratch. Especially on topics of academic interest that are novel and
that do not have well-known answers that can be found in the training corpus for these models, in general they
will not produce useful text when prompted. Therefore, LLMs should never be used to generate novel text.
\item LLMs are really bad at giving references. Prompts that ask for academic references on a topic are likely to
produce non-existing ``hallucinated'' references. The existing references an LLM is most likely to dig up
usually occur on the first page of a web search on the topic, or are frequently cited in literature on the
topic. Thus, LLMs should never be directly queried for references. When researching a new concept, a better use
of an LLM is the generation of query strings for search engines like Google Scholar.
\end{enumerate}
Applying these observations, I never copied text from the LLM into this thesis. Where I edited the text of this thesis
using suggestions from LLM output, I critically evaluated the LLM output and carefully considered each edit. Following
are some examples of how I used LLMs in the writing of this thesis.
\paragraph{For checking spelling and grammar,} the LLM was prompted with an instruction to review the text and output a
list of errors. The list was then reviewed and the errors were fixed in the source document by hand. An example prompt
for the LLM in this case might be: ``The attached file contains the LaTeX source code of a chapter of an doctoral thesis
titled `...'. Review the text and list any mistakes in spelling or grammar.''
\paragraph{For improving formulation patterns,} the LLM was prompted with a short excerpt of text of at most two
paragraphs and instructions asking for an improved version of the text. In response to such a prompt, the LLM will often
change the meaning of parts of the text. Thus, I used the output as a reference example, and manually adjusted the
source document applying parts of the LLM response where fitting. An example prompt in this case might be: ``The
following text are two paragraphs from a chapter on `...' in a PhD thesis on `...' . Improve the wording of these
paragraphs to make them easier to read and understand.''.
\paragraph{For improving the structure of the text,} the LLM was prompted with an instruction to review the text and
output a list of recommendations. The list was then reviewed, and changes were made to the source document by hand. An
example prompt in this case might be: ``The attached document contains the LaTeX source code of a chapter of a PhD
thesis on `...' . Critically assess the structure and organization of the chapter and write a list of suggestions for
improvement.''
In accordance with the recommendations of the University and State Library Darmstadt regarding the labelling and
documentation of AI-generated materials dated September 22, 2025\cite{RecommendationsUniversityState2025}, instances
where I used an LLM to edit parts of the text of this thesis as described above have not been explicitly labelled in the
text. The LLM in this use assumes a similar role a human editor might assume reviewing the text.
Besides the use of LLMs as described above, a specialized machine translation tool was used to create the German
translation of the abstract at the beginning of this thesis. This use is marked explicitly.
\chapterbibliography