Page 75 - iiA

The Snowball Effect

importance. According to the words convey stronger con-
Audit Committee Practices notations than words such as
are trained, they increasingly
“the,” which convey relatively
Report, a research report
little information. As LLMs
from Deloitte’s Center for
Board Effectiveness and the
Center for Audit Quality,
identify the most meaningful
words based on how often
‚ % of directors say AI gov-
ernance has appeared on
those words appear in the
data used to train the model.
their board agenda at least
semi-annually. Another €€%
Traditional machine
say the topic has come up
“as needed.” AI governance learning models, such as
those powering the core
audits tailored for LLMs can functions of Uber ride-hailing
help address concerns over apps, are trained with data
accuracy, bias, privacy, and from known sources. In
cybersecurity and help orga- contrast, the most wide-
nizations remain committed spread open-source LLMs,
to responsible AI. like ChatGPT, are typically
trained on a large variety of
COMPOUNDING information sources, includ-

n , within just five that exacerbate the chal- AI ISSUES ing those collected from the
days of ChatGPT’s release lenges for responsible AI. Machine learning models internet and random users.
to the public, the chatbot Some of the risky behaviors power numerous products, This raises challenges for
had already amassed mil- that LLMs exhibit include such as the Uber apps and ensuring answers are accu-
lion users. In comparison, delivering biased recom- Amazon recommender sys- rate, copyright-free, unbi-
it took Instagram  days mendations, returning erro- tems, providing users with ased, and inoffensive.
and Facebook € ‚ days to neous information based on accurate and helpful infor- LLM models are what
achieve the same number of untrustworthy sources or mation. The models that they eat, so to speak. If an
users after launch. The rapid due to AI hallucinations, and power most AI-driven prod- LLM creates new content
adoption of these so-called making organizations more ucts are trained on informa- that is published online, this
conversational agents that vulnerable to data leaks. tion from known resources, information becomes avail-
are based on large language Responsible AI refers to AI including information that able for training LLMs going
models (LLMs) speaks to that is developed and used may be copyrighted. forward. If published con-
their impressive utility and in a trustworthy and ethi- In simple terms, LLMs tent is inaccurate or biased,
adaptability. Chatbots apply cal manner. In particular, are machine learning mod- it risks exposing LLMs to
human language capabil- the responsible develop- els that predict the next, best a feedback loop of spread-
ities to automate tasks, ment and use of AI systems word in a sentence. Through ing false or prejudiced data.
such as summarizing docu- should generate results that repeating the prediction, Erroneous information
ments and creating content, are robust and free from LLMs generate content that snowballs, creating a vicious
and show great promise bias with clear account- is customized for users on cycle of inaccuracy and bias.
when it comes to increasing ability, transparency, and demand — such as a sum- A distinctive feature of
worker productivity. explainable processes, while mary of customer service LLMs is that they prompt
However, it has also been ensuring stakeholder privacy transcripts or a poem based users for information and
clear from the beginning and information security. on a theme. The prediction incorporate these inputs
that these artificial intelli- To help organizations partly depends on the fre- into their learning process,
gence (AI)-based systems meet these criteria, AI gov- quency of meaningful words creating a risk that sensi-
have innate characteristics ernance is growing in LLMs are given. Meaningful tive information from one

42 INTERNAL AUDIT TODAY 72
April 2025