JOINT CIRCULAR
STATE OF NEW JERSEY
NO.: 23-OIT-007
ORIGINATING AGENCIES:
PAGE 1 OF 5
- OFFICE OF INFORMATION TECHNOLOGY (OIT)
- OFFICE OF HOMELAND SECURITY & PREPAREDNESS
- NJ CYBERSECURITY & COMMUNICATIONS
INTEGRATION CELL (NJCCIC)
EFFECTIVE
DATE: IMMEDIATE
EXPIRATION SUPERSEDES: N/A
DATE: INDEFINITE
SUBJECT: STATE OF NEW JERSEY INTERIM GUIDANCE ON RESPONSIBLE USE OF GENERATIVE AI
ATTENTION: DIRECTORS OF ADMINISTRATION, CHIEF INFORMATION OFFICERS, AGENCY IT MANAGERS, AND AGENCY
CHIEFS OF STAFF
APPLIES TO:
FULL TIME, PART TIME AND STAFF
AUGMENTATION IT/TECHNOLOGY
STAFF
AND BUSINESS STAFF SUPPORTING
APPLICATIONS AND WEB-PRESENCE
FOR STATE AGENCIES
FOR QUESTIONS: SEND QUESTIONS TO: CTO@TECH.NJ.GOV
I.
PURPOSE
A.
Generative AI refers to a new set of technologies that utilize machine learning techniques to
generate content in response to user inputs. The content produced can be textual (e.g. ChatGPT,
Bedrock, or Bard), visual (e.g. DALL-E, Canva or MidJourney), spoken (e.g. ElevenLabsM4T,
Polly/LEX) or even musical (e.g. MusicLM or Amadeus Code).
B.
These new tools have the potential to be extraordinarily useful to public servants in your work, but
they also present risks. Therefore, these guidelines serve as an interim resource for employees of
the State of New Jersey to encourage responsible use of these emerging technologies.
C.
Generative AI is a tool, not actual intelligence. We remain responsible for the outcomes and uses
of these tools. Technology enables our work but does not excuse our judgment.
D.
As we explore the responsible use of artificial intelligence, we embrace the following principles to
guide our efforts. These shared values of empowerment, inclusion, transparency, innovation, and
risk management will steer our experimentation and decision-making when employing these
rapidly evolving technologies.
E.
With this Circular, we aim to build trust, spur progress, and ensure AI serves the public good.
II.
POLICY
A.
GENERAL
1.
Empowerment. The use of AI should support our workforce to deliver enhanced services and
products efficiently, safely, and equitably to all residents. We rely on the judgment of our
professionals to ensure we realize the benefits of these tools.
2.
Inclusion and Respect. The use and development of AI should uplift communities,
connecting them effectively with resources to thrive, especially those historically marginalized.
As public stewards, we will use tools respectfully to reflect values of equity and social justice.
3.
Transparency and Accountability. We acknowledge the limits of foresight. But transparency
builds trust and enables collective learning. When AI is used, we must disclose that
responsibly and share our workflow freely with other public servants and with the public.
4.
Innovation and Risk Management. We embrace responsible experimentation that maintains
control and respects privacy and security while developing uses that drive efficiency, dialogue,
and better service. We understand risks may not be fully apparent initially and commit to
proactive risk assessment.
B.
USE GUIDELINES
1.
Ask - Early and Often. Generative AI creates content based on your inputs. That is why it is
good to try out different questions (also known as “prompts”) to see how it responds. You can
specify the tone, style, and length of a text response and the attributes and qualities of an
image. The more you experiment with different ways of steering the tools, the faster you will
learn how to instruct them to yield the best results.
2.
Fact Check. Verify all AI-generated content, especially for public use. Generative AI can swiftly
produce clear prose, but the information may be inaccurate or outdated. Research claims to
ensure accuracy. Watch for: incorrect facts, events, links, or references; and biased information
potentially harmful to vulnerable groups like racial, ethnic, and gender minorities, people with
disabilities, etc. We must actively mitigate risks from AI while benefiting from its capabilities. As
public stewards, we have an obligation to use these tools responsibly. Signoff from Agency
Director level or Asst Commissioner is required before production use.
3.
Disclose. Label content created with generative AI to that effect. Transparency is crucial in AI-
generated content, even when AI is used sparingly. Disclosing AI involvement fosters trust and
aids in error identification. Detail the AI model, prompts, and methods employed. This
documentation aids comprehension and safe usage by colleagues and stakeholders.
Sample AI-generated content disclosures:
"This content was generated with the aid of ChatGPT and subsequently revised by Bob
Smith."
"This text was summarized for clarity using Google Bard."
4.
Sensitive Information. When prompting the AI or using AI models, do not disclose sensitive
or private information. We aim to enable responsible AI use while safeguarding sensitive
information.
a)
Do not share personally identifiable information (PII) about residents, colleagues, or
yourself. Do not share confidential or sensitive content.
b)
Do not use AI tools to transcribe or summarize meetings where sensitive topics are
discussed.
c)
Do not share any information that you wouldn’t share publicly.
III.
PROCEDURES
A.
HOW TO’S
1.
Drafting. Generative AI like ChatGPT can help draft memos, letters, and job descriptions.
When creating prompts, specify any format preferences, keywords, technical terms, or
phrases to include or avoid. Remember, you can tell the software how long you would like the
response to be (e.g. the word count), the style (e.g. professional or informal), the language
(e.g. please respond in Spanish).
Example: Memos: In government we often have to write short documents that present an
argument why a policy should be adopted, or a decision should be made. For instance, try
the prompt in ChatGPT, Bard, and other generative text tools:
Write a 250-word memo to a state agency head about the potential
benefits of the use of generative AI in state government. Please
support the memo with evidence. The tone should be professional, and
the text should be formatted as a memo.
Suggested:
Try to be specific in the prompt. If you give more context, the answer becomes more
relevant.
Try different prompts to see how the responses improve.
Edit and review the content carefully. Seek & use AI Quality Assurance best practices.
Fact check.
Do Not:
Do not rely on generative AI to provide accurate answers.
Do not use generative AI to create communication regarding sensitive topics. For
instance, a renowned institution was criticized for using generative AI to write a press
release regarding a shooting.
Do not use content without carefully editing.
2.
Communicating with the Public Generative AI can help you write clearly and in plain,
accessible language. You can use the prompt to indicate the reading level or audience for a
text.
Example: Use Anthropic’s Claude to upload the text of a website or policy and ask the AI to
help you simplify the language and make it more accessible, concise and easy-to-understand.
Help me to rewrite the text of this page in simple English accessible
to a ten-year old. While you may streamline the language for clarity,
please do not eliminate any directions.
Suggested:
Specify in the prompt if you have a specific audience in mind.
Try different prompts or verbiage of the same sentence to find what works best.
You can pass the output of the text by using a readability app that can identity challenging
sentences, as well as the reading level for the text.
Review the text to ensure that the language is inclusive and respectful. The models might
use language or patterns that appear regularly, but that might exclude some people. For
instance, a model might suggest: “Dear Sir/Ma'am” does not include non-binary people
and could be replaced with “Dear Colleague” or “Dear Neighbor”.
Do Not:
Do not include confidential information in a prompt.
3.
Translation. Generative AI can help you translate content into other languages. Facebook
Reseearch claims it’s free, open source Seamless M4T tool can translate text or speech from 100
languages into 35 languages.
Example: Use M4T, ChatGPT or Google Bard to translate instructions for applying for a
government benefit into the most popular non-English languages.
Suggested:
Ask a native speaker to review.
Try different languages to test the quality of the translation.
Do Not:
Do not include confidential information in the prompt.
Do not use content generated in a language you do not understand before consulting someone
with proficiency in the language. You still need to check for accuracy, bias, etc.
Note that text generated in other languages might be confusing to people who speak different
regional dialects. Do not assume that some text will be easily understood by all speakers. Be
specific in your prompt and ask for regional idioms.
4.
Summarization. Generative AI can summarize longer text or speech into the desired length.
You can use it to summarize lengthy documents for clarity or brevity when required. Many
citizen engagement platforms now make it possible to summarize and organize citizen
comments.
Please write a 100-word summary of this report, highlighting the
most important recommendations. Use bullet points.
Suggested:
Remember that summaries and transcripts are subject to OPRA
If you plan on making a decision based on the summary, you should read the entire
document(s) to make sure you did not miss or mischaracterized the original document.
Be aware that the resulting summary might have biases as it will tend to present language
that is more frequent in the data used to train the model. You can use changes to the
prompt to enhance the results by suggesting that the result incorporates perspectives
from marginalized groups. Even better, you can engage with some individuals in these
communities to better understand their perspectives on the text generated.
Do Not:
Do not include confidential information in the prompt: make sure you have deleted
confidential information from your notes or other inputs.
Do use automated transcription in meetings where confidential or personally identifiable
information is discussed.
5.
Data Analysis. Utilizing a code interpreter with generative AI empowers individuals with
limited coding experience to engage in data analysis tasks. This is particularly beneficial for
those who are less technically inclined, including interns and student workers, enabling them
to effectively contribute to technical projects. For instance, OpenAI’s Code Interpreter can be
employed to interact with datasets through inquiries. Feel free to experiment with various
programming languages and libraries. However, ensure that you grasp the code's logic and
thoroughly go through the relevant component documentation before application. Generative
AI serves as an initial aid, but it might be necessary to adjust the generated code for optimal
results.
Example: “Hello Code Interpreter, I have collected a bunch of survey
forms in PDF format. These forms contain valuable information, but
it's a pain to go through them manually. I am looking to extract the
participant comments and their locations from these PDFs and create
a spreadsheet for easier review. Could you help me generate some
Python code using libraries like 'PyPDF2' and 'Pandas' to extract
this data and organize it into a neat CSV file? It would be a huge
time-saver! Thanks!"
Suggested:
Safely explore new languages and libraries.
Understand the functioning of the code before its practical application.
Rely on Generative AI for a head start.
Do Not:
Avoid inserting sensitive information in prompts. Following development best practices,
abstain from including confidential data like passwords, private keys, or proprietary details
either in your code or the prompts.
Refrain from utilizing code in a production setting without a comprehensive understanding
of its operations.
Ensure you are well-acquainted with novel libraries and dependencies. Familiarize
yourself with potential vulnerabilities and security aspects related to the chosen language
or library.
6.
Generating Images, Audio and Video. Generative AI enables the creation of images, audio,
and videos based on prompts, offering support in crafting captivating and insightful
communication resources. Tools like Mid-Journey and DALL-E generate images. CANVA is
a popular tool for automating the creation of slides. The Music plugin available from ChatGPT,
Amadeus Code and Amper Music are examples of music generators.
Write me a jingle or song to remind residents to enroll in
GetCoveredNJ.
Generate a training video illustrating how to schedule bulky item
pickups, leveraging a provided script.
Suggested:
Recognize the potency of visual, audio, and video communication in conveying messages.
Use generative AI to harness these tools, extending your artistic abilities.
Employ generative AI to draft mock-ups that facilitate effective communication with
graphic designers, videographers, and other creative professionals.
Collaborate with your department's public information officer when using or publishing
images, audio, or videos. Their expertise ensures adherence to accessibility and branding
best practices.
Engage with Cabinet members or counsel; Respectfully seeking their insights helps
identify potentially hurtful, discriminatory, or misconstrued content.
Do Not:
Include confidential information in prompts; Rather, ensure any sensitive data is removed
from notes or inputs. This includes faces, voices, identifications, and license plates of
individuals who haven't granted consent.
Assume that the outputs of generative AI are respectful and non-offensive, especially to
vulnerable residents like diverse ethnic, racial, and gender groups; Test and validate.
Ensure all content aligns with the OIT and Governor’s Office Web Presence Guidelines to
maintain visual consistency and compliance.
Christopher Rein, CTO Michael Geraghty, CISO
Office of Information Technology Office of Homeland Security & Preparedness
