Page 12 - AI Vol 2: Risks of AI
P. 12
obtained on the Internet, a significant percentage customers, and parents and students aware of
of which is copyrighted. A number of high- how the agency is utilizing AI and what protocols
profile lawsuits have been brought challenging and safeguards are in place. Doing so can help
these companies’ use of this copyrighted work. If alleviate skepticism surrounding the agency's use
these lawsuits are successful, it is possible that of AI and mitigate the perception that the agency
these AI services may suddenly be unavailable, is concealing its AI practices, which could lead
causing hardship for agencies that have come to negative consequences. In addition to general
to rely on them. Locally deployed open-source transparency on AI deployments, agencies should
models have the advantage of being immune be especially careful and transparent about
from issues with the viability of the creators of public and student facing AI deployments given
those models, but if the models themselves are the risk of harmful or inappropriate AI outputs.
ruled to contain “copies” of protected work, an For example, if a city deploys an AI chatbot to
agency’s deployment of such models may subject assist users in identifying applicable building
the agency itself to potential liability. and municipal code requirements but does not
prominently warn the users that the AI may
Another crucial issue is public transparency produce inaccurate responses, and the AI model
regarding AI use. As the public becomes more produces an inaccurate output which is relied
aware of and familiar with AI systems, and the upon by the user to his or her detriment, the user
related risks, it is likely there will be increased may have a legal action against the agency.
scrutiny on how agencies are (and are not) utilizing
AI in their provision of services. Agencies should Similarly, use of AI systems by public agency
be proactive in making the public, constituents, officials and employees may result in the creation
of “records” for purposes of the California Public
Records Act. This is very much an untested area
of the law, and whether a particular interaction
with an AI system results in the creation of a
disclosable “record” under the California Public
Records Act may depend on a number of factors
such as how the system is deployed (locally run
open-source model vs commercially available
model which uses user prompts for training)
and how the model was used. Agencies should
still take into consideration the potential need
to disclose both user prompts and AI outputs in
response to Public Records Act requests. This
underscores the central importance of establishing
protocols on AI use, educating staff and officials
on permitted uses and risks and limitations of AI
12 | VOLUME 2 RISKS OF AI | LOZANOSMITH.COM