Page 10 - AI Vol 2: Risks of AI
P. 10
offering thorough training to employees is essential to
prevent accidental misuse and any liability associated AS AI BECOMES THE NEW
with that use. INFRASTRUCTURE, FLOWING
INVISIBLY THROUGH OUR
It is also important for agencies to be informed about DAILY LIVES LIKE THE WATER
the variety of AI systems to select the best system for IN OUR FAUCETS, WE MUST
the agency’s needs. For example, open-source models UNDERSTAND ITS SHORT- AND
LONG-TERM EFFECTS AND
deployed on an agency’s own servers may provide KNOW THAT IT IS SAFE FOR
certain data privacy benefits and be more appropriate to ALL TO USE.
the extent the agency expects these systems to work with
sensitive information, though currently available open- - KATE CRAWFORD
source models perform at a lower level than the major
commercial models. It depends on the service, however,
most paid AI services allow for users to opt out of having
their prompts used for training purposes. Monitoring
and controlling employee use of AI is possible when
the employee is using employer provided computer
equipment or accounts; otherwise it would prove difficult
and an employer would need to rely on any employee
policy that is in place. If the model will be deployed in
a public or student facing way, it is important to consider
what protections are available to ensure the model does
not produce inappropriate, offensive, or harmful outputs.
Does the agency have mechanisms in place to alert them
of any inappropriate use of the AI or to notify them if
the model generates a prohibited output? Is the agency
able to test the models for potential biases and are there
mechanisms to correct such biases? These are just some
of the considerations the agency should be engaging in
when evaluating AI deployments.
After models are deployed, it is critical that the agency
continually evaluate the model’s performance for
accuracy and effectiveness. The agency may find that
their LLM deployment excels in certain tasks, saving
the agency staff time and resources, while it provides
poor or inaccurate results in other tasks. The agency
may find that certain prompting strategies produce better
10 | VOLUME 2 RISKS OF AI | LOZANOSMITH.COM