BILL NUMBER: S934
SPONSOR: GONZALEZ
 
TITLE OF BILL:
An act to amend the general business law, in relation to requiring warn-
ings on generative artificial intelligence systems
 
PURPOSE OR GENERAL IDEA OF BILL:
Requires warnings on generative artificial intelligence systems.
 
SUMMARY OF PROVISIONS:
399-zzzzzz(1): Defines important terms.
399-zzzzzz(2): Requires that all generative AI systems have a conspicu-
ous warning on their user interface which reasonably apprises the user
that the outputs of the system may be inaccurate.
399-zzzzzz(3): Provides penalties.
 
JUSTIFICATION:
It was recently discovered that a generative artificial intelligence
chatbot was fabricating claims of sexual harassment about New York
lawmakers.
Generative artificial intelligence opens the door for numerous benefi-
cial use cases. However, as it is still in its infancy, generative Al
has the potential to hallucinate that is, confidently provide false or
misleading information.
The user interfaces of the platforms are easy- to-use by design, so they
are accessible to users who do not fully understand Al and how/ why it
hallucinates false information. Therefore, it is important that users be
apprise of the fact that these systems can be inaccurate, particularly
when such information can be falser regarding another, or even dangerous
to the user (for example, if the generative Al system provided improper
medical information).
An internal study has revealed that, of a sample of the 8 widely used
chatbots, only 4 make their warnings consistently visible to the user on
their web interface and only 3 make their warnings consistently visible
to the user on their mobile interface. Of the 8 chatbots, a warning is
visible at least for a short period of time on 6 chatbots on their web
interface and 5 chatbots on their mobile interface (though, this means
that the warning will disappear or effectively disappear after the first
message or few messages are sent).
After further testing of the 8 chatbots studied, at least one other
chatbot repeated the false claims of sexual harassment, with the chatbot
going so far as even providing the name of a fake accuser and a fabri-
cated background. A third chatbot provided other, less serious, false
accusations., One chatbot refused to answer any questions related to
public or priVate'persons, one chatbot refused to engage in òany conver-
sation related to sensitive topics, and three chatbots properly failed
to provide any false information. The chatbot that originally was fabri-
cating the claims of sexual harassment has since ceased doing so.
 
PRIOR LEGISLATIVE HISTORY:
2023-2024 S.9450A PASSED SENATE / A.10103B Third Reading
 
DIFFERENCE BETWEEN ORIGINAL AND AMENDED:
Adjusted the definitions of generative artificial intelligence systems
and artificial intelligence.
 
FISCAL IMPLICATIONS FOR STATE AND LOCAL GOVERNMENTS:
N/A
 
EFFECTIVE DATE:
This act shall take effect on the ninetieth day after it shall have
become law.