Large language model powered social integrity system
A pseudo account is set up to mimic a potential victim's demographic messaging behavior while seeming completely legitimate to an anti-social (AS) user. This can entail mimicking the messaging behavior of minors and presenting that behavior where AS users look for it. The AS account establishes communication with the pseudo account. The AS account may be flagged as suspicious, and the system can wait for the AS user to make an initial move on a legitimate account and report it before any undesirable communication takes place. The pseudo account can be controlled by a large language model (LLM) such as a generative pre-trained transformer (GPTT).
1 . A system comprising:
at least one non-transitory computer medium comprising instructions executable by at least one processor assembly to:
generate a pseudo account, the pseudo account accessing at least one large language model (LLM) to appear as a peer user account;
identify a first account associated with a user, wherein an input of the user via an input device requests establishing communication with the pseudo account;
communicate, using the pseudo account, with the first account, wherein the communication includes the LLM generating a multi-turn natural language dialog, and wherein at least a portion of the multi-turn natural language dialog is presented to the first account as originating from the peer user account;
determine, by processing data of the communication using an analysis model that is separate from the LLM, that the first account exhibits a first characteristic;
determine, based on the first characteristic, that the first account is to be further monitored without restricting the first account; and
generate an advisory warning based on the first account establishing a communication with a second account.
2 . The system of claim 1 , wherein an account to which the advisory warning that the first account exhibits the first characteristic comprises at least a user account.
3 . The system of claim 1 , wherein an account to which the advisory warning that the first account exhibits the first characteristic comprises at least a social network administrator account.
4 . The system of claim 1 , wherein the LLM comprises at least one generative pre-trained transformer (GPTT).
5 . The system of claim 1 , wherein the instructions are executable to present the advisory warning to an account before any communication is established by the first account with the second account.
6 . The system of claim 1 , wherein the LLM is trained to mimic a behavior of a demographic associated with potential victims to the first account.
7 . The system of claim 6 , wherein the LLM is trained to mimic a messaging behavior of a demographic associated with potential victims to the first account.
8 . The system of claim 1 , wherein the advisory warning comprises:
a warning that the first account is potentially anti-social;
an option for the second account to block further communications from the first account; or
an option for the second account to continue receiving communications from the first account.
9 . A method comprising:
generating a pseudo account, the pseudo account accessing at least one large language model (LLM) to appear as a peer user account;
identifying a first account associated with a user, wherein an input of the user via an input device requests establishing communication with the pseudo account;
communicating, using the pseudo account, with the first account, wherein the communication includes the LLM generating a multi-turn natural-language dialog, and wherein at least a portion of the multi-turn natural-language dialog is presented to the first account as originating from the peer user account;
determining, by processing data of the communication using an analysis model that is separate from the LLM, that the first account exhibits a first characteristic;
determining, based on the first characteristic, that the first account is to be further monitored without restricting the first account; and
generating an advisory warning based on the first account establishing a communication with a second account.
10 . The method of claim 9 , wherein the first characteristic comprises predatory behavior.
11 . The method of claim 9 , wherein the large language model (LLM) is trained to mimic a behavior of a demographic associated with potential victims to the first account.
12 . The method of claim 11 , wherein the LLM comprises at least one generative pre-trained transformer (GPTT).
13 . The method of claim 9 , comprising:
responsive to at least one of a predetermined number of user complaints regarding the first account and plural indications the first account exhibits the first characteristic in respective plural communication sessions between the pseudo account and the first account, suspending the first account.
14 . An apparatus, comprising:
at least one processor assembly configured to:
generate a pseudo account, the pseudo account accessing at least one large language model (LLM) to appear as a peer user account;
identify a first account associated with a user, wherein an input of the user via an input device requests establishing communication with the pseudo account;
communicate, using the pseudo account, with the first account, wherein the communication includes the LLM generating a multi-turn natural-language dialog, and wherein at least a portion of the multi-turn natural-language dialog is presented to the first account as originating from the peer user account;
determine, by processing data of the communication using an analysis model that is separate from the LLM, that the first account exhibits a first characteristic;
determine, based on the first characteristic, that the first account is to be further monitored without restricting the first account; and
generate an advisory warning based on the first account establishing a communication with a second account.
15 . The apparatus of claim 14 , wherein the processor is configured to present the advisory warning on a computer hosting a user account with which the first account attempts to establish communication with.
16 . The apparatus of claim 14 , wherein the processor is configured to present the advisory warning on a computer hosting a parent account associated with a user account with which the first account attempts to establish communication with.
17 . The apparatus of claim 14 , wherein the processor is configured to present the advisory warning on a computer hosting a social network administrator account.
18 . The apparatus of claim 14 , wherein the processor is configured to provide to a computer other than a computer hosting the first account at least one selector selectable to block the first account.
19 . The apparatus of claim 14 , wherein the processor is configured to provide to the second account an indication that the first account may be problematic.
20 . The apparatus of claim 19 , wherein the processor is configured to provide to the second account, along with the indication that the first account may be problematic, an option to contact a social network administrator.