US-12625909-B2 - Enabling or blocking processing of queries to an artificial intelligence system based on intents of the queries

Abstract

Methods, systems, and non-transitory computer readable storage media are disclosed for controlling access to artificial intelligence systems based on determined intent of queries. The disclosed system utilizes one or more digital content analysis models to determine an intent of one or more queries to an artificial intelligence system. The disclosed system utilizes the one or more digital content analysis models to determine an intended use of the artificial intelligence system. Additionally, the disclosed system determines whether the intent of the one or more queries aligns with the intended use of the artificial intelligence system by generating a similarity score and comparing the similarity score to a similarity threshold. Based on whether the intent aligns with the intended use, the disclosed system executes computing instructions to enable or block the one or more queries from being processed by the artificial intelligence system.

Inventors

• Shane Wiggins

Assignees

• ONETRUST LLC

Dates

Publication Date

20260512

Application Date

20240517

Claims (20)

1. 1 . A computer-implemented method comprising: determining, by one or more hardware processors utilizing one or more digital content analysis models, an intent of one or more queries to an artificial intelligence system based on digital content in the one or more queries; determining, by the one or more hardware processors, an intended use of the artificial intelligence system from documentation associated with the artificial intelligence system; generating, by the one or more hardware processors, a similarity score by comparing the intent of the one or more queries to the intended use of the artificial intelligence system; and in response to a comparison between the similarity score and a similarity threshold, executing, by the one or more hardware processors, computing instructions to: enable processing of the one or more queries by the artificial intelligence system; or block processing of the one or more queries by the artificial intelligence system.
2. 2 . The computer-implemented method of claim 1 , wherein determining the intent of the one or more queries comprises: intercepting the one or more queries from a query device to the artificial intelligence system; determining one or more text strings or one or more digital images from the one or more queries; and determining the intent of the one or more queries by utilizing a text processing neural network or one or more image processing neural networks to perform a semantic analysis on the one or more text strings or the one or more digital images.
3. 3 . The computer-implemented method of claim 2 , wherein enabling processing of the one or more queries by the artificial intelligence system comprises: determining that the similarity score of the one or more queries meets the similarity threshold; and providing the one or more queries to one or more computing devices implementing the artificial intelligence system.
4. 4 . The computer-implemented method of claim 2 , wherein blocking processing of the one or more queries by the artificial intelligence system comprises: determining that the similarity score of the one or more queries does not meet the similarity threshold; and preventing one or more computing devices implementing the artificial intelligence system from receiving the one or more queries.
5. 5 . The computer-implemented method of claim 1 , wherein determining the intended use of the artificial intelligence system comprises: accessing the documentation associated with the artificial intelligence system from a computing system implementing the artificial intelligence system; and determining the intended use of the artificial intelligence system from the documentation utilizing a text processing neural network.
6. 6 . The computer-implemented method of claim 5 , wherein determining the intended use of the artificial intelligence system comprises: accessing, from a third-party computing system, a set of requirements for the artificial intelligence system from a system requirements framework associated with the artificial intelligence system; and determining the intended use of the artificial intelligence system from the set of requirements and the documentation utilizing the text processing neural network.
7. 7 . The computer-implemented method of claim 1 , wherein determining the intent of the one or more queries comprises determining the one or more queries from a first text input previously entered into a chat interface with the artificial intelligence system and a second text input being entered into the chat interface with the artificial intelligence system.
8. 8 . The computer-implemented method of claim 1 , further comprising: determining a ground-truth intent for the one or more queries and a ground-truth intended use of the artificial intelligence system; determining one or more losses based on the intent of the one or more queries, the intended use of the artificial intelligence system, the ground-truth intent, and the ground-truth intended use; and modifying parameters of the one or more digital content analysis models or one or more neural networks that determine the intended use of the artificial intelligence system according to the one or more losses.
9. 9 . The computer-implemented method of claim 8 , wherein determining the one or more losses comprises: determining, utilizing one or more loss functions, a first loss indicating a first difference between the intent of the one or more queries and the ground-truth intent; and determining, utilizing the one or more loss functions, a second loss indicating a second difference between the intended use of the artificial intelligence system and the ground-truth intended use.
10. 10 . A system comprising: a first computing system implementing an artificial intelligence system; and a second computing system comprising one or more hardware processors to: intercept one or more queries from a client device to the artificial intelligence system; determine, utilizing one or more digital content analysis models, an intent of the one or more queries to the artificial intelligence system based on digital content in the one or more queries; determine an intended use of the artificial intelligence system from documentation associated with the artificial intelligence system and accessed from the first computing system; generate a similarity score representing a similarity of the intent of the one or more queries and the intended use of the artificial intelligence system; and in response to a comparison between the similarity score and a similarity threshold, execute computing instructions to: enable processing of the one or more queries by the artificial intelligence system by submitting the one or more queries to the first computing system; or block processing of the one or more queries by the artificial intelligence system by preventing the one or more queries from passing to the first computing system.
11. 11 . The system of claim 10 , wherein the one or more hardware processors are further configured to intercept the one or more queries from the client device to the artificial intelligence system by routing data input by a plurality of client devices comprising the client device via an application interface associated with the artificial intelligence system to the second computing system.
12. 12 . The system of claim 10 , wherein the one or more hardware processors are further configured to determine the intent of the one or more queries by: determining one or more text phrases form the one or more queries; and determining the intent of the one or more queries by determining the intent from the one or more text phrases utilizing a text processing neural network.
13. 13 . The system of claim 10 , wherein the one or more hardware processors are further configured to determine the intended use of the artificial intelligence system by: determining one or more digital documents comprising information associated with one or more of an architecture, training data, hyperparameters, validation data, evaluation data, input data, or output data of one or more machine-learning models of the artificial intelligence system; and determining the intended use of the artificial intelligence system by processing the one or more digital documents utilizing a text processing neural network.
14. 14 . The system of claim 10 , wherein the one or more hardware processors are further configured to determine the intended use of the artificial intelligence system by: determining one or more digital documents comprising data requirements of a system requirements framework associated with the artificial intelligence system; and determining the intended use of the artificial intelligence system by processing the one or more digital documents utilizing a text processing neural network.
15. 15 . The system of claim 10 , wherein the one or more hardware processors are further configured to execute the computing instructions to enable processing of the one or more queries by the artificial intelligence system by: determining that the similarity score meets the similarity threshold indicating that the intent of the one or more queries is within the similarity threshold of the intended use of the artificial intelligence system; and providing the one or more queries to the artificial intelligence system at the first computing system.
16. 16 . The system of claim 10 , wherein the one or more hardware processors are further configured to execute the computing instructions to enable processing of the one or more queries by the artificial intelligence system by: determining that the similarity score does not meet the similarity threshold indicating that the intent of the one or more queries is not within the similarity threshold of the intended use of the artificial intelligence system; and preventing the artificial intelligence system at the first computing system from receiving the one or more queries.
17. 17 . The system of claim 16 , wherein the one or more hardware processors are further configured to provide a notification to the client device indicating that the one or more queries have not been provided to the artificial intelligence system with an indication that the intent of the one or more queries does not align with the intended use of the artificial intelligence system.
18. 18 . A non-transitory computer readable medium comprising instructions that, when executed by computing hardware, cause the computing hardware to: intercept one or more queries from a client device to an artificial intelligence system; determine, utilizing one or more digital content analysis models, an intent of the one or more queries to the artificial intelligence system based on digital content in the one or more queries; determine, utilizing the one or more digital content analysis models, an intended use of the artificial intelligence system from documentation associated with the artificial intelligence system; generate a similarity score representing a similarity of the intent of the one or more queries and the intended use of the artificial intelligence system; and in response to a comparison between the similarity score and a similarity threshold, execute, by the one or more hardware processors, computing instructions to enable or block processing of the one or more queries by the artificial intelligence system.
19. 19 . The non-transitory computer readable medium of claim 18 , wherein the instructions that, when executed by the computing hardware, further cause the computing hardware to: determine that the intent of the one or more queries does not match the intended use of the artificial intelligence system in response determining that the similarity score does not meet the similarity threshold; and execute the computing instructions to block processing of the one or more queries by the artificial intelligence system by preventing the one or more queries from being provided to a computing system implementing the artificial intelligence system.
20. 20 . The non-transitory computer readable medium of claim 18 , wherein the instructions that, when executed by the computing hardware, further cause the computing hardware to: determine that the intent of the one or more queries matches the intended use of the artificial intelligence system in response to determining that the similarity score meets the similarity threshold; and execute the computing instructions to enable processing of the one or more queries by the artificial intelligence system by providing the one or more queries to a computing system implementing the artificial intelligence system.

Description

BACKGROUND Advances in computer processing and data storage technologies have led to significant advances in the use of artificial intelligence for many different purposes. For instance, many entities utilize artificial intelligence to provide interaction interfaces with users or to interface with other computing systems and perform a variety of different tasks. To illustrate, many entities utilize large language models (or other generative neural networks), data analysis neural networks, or other machine-learning models to implement interactive tools for providing support, content generation and editing, data analysis, or other computational tasks. Because machine-learning is such an integral part of so many computing processes, ensuring that the machine-learning models are utilized in a way that they were intended is an important and often challenging aspect of ensuring that the computing processes are accurate and efficient. For example, an entity that has implemented one or more large language models in connection with a chat bot that provides support to users via a chat interface has an interest in ensuring that the chat bot provides accurate and relevant information. Accordingly, bad actors or other users utilizing the chat bot in ways that deviate from the intended use (e.g., by submitting queries unrelated to support issues) can often have an impact on the training and outputs of the chat bot. Some conventional systems rely on built-in guardrails for machine-learning models (e.g., through specialized training) to prevent misuse of the models. Although such guardrails can be useful in preventing certain types of misuse by the models, implementing such guardrails can require very specific training datasets and/or a significant amount of processing time and resources. Additionally, such protections can lead to inconsistency with edge cases for similar (but not exactly the same) scenarios or across domains. Thus, the conventional systems are often inaccurate and inefficient. Furthermore, some conventional systems use content classification to determine specific types of data provided to machine-learning models in connection with managing use of machine-learning models. Although such conventional systems allow for preventing machine-learning models to access/obtain certain data types (e.g., personally identifiable information), these systems are also limited in protecting the use of the models to intended uses. For example, bad actors can circumvent intended uses even using allowed data types or inputs through a series of inputs that circumvent or break controls associated with the machine-learning models. Thus, these conventional systems fail to protect the use of the machine-learning models in a variety of different scenarios. Accordingly, existing systems lack efficiency and accuracy in ensuring proper inputs and outputs of machine-learning models. SUMMARY This disclosure describes various aspects for controlling access to artificial intelligence systems based on determined intent of queries. For example, the disclosed systems utilize one or more digital content analysis models to determine an intent of one or more queries to an artificial intelligence system. Additionally, the disclosed systems utilize the digital content analysis model(s) to determine an intended use of the artificial intelligence system from documentation associated with the artificial intelligence system. The disclosed systems also compares the intent determined for the one or more queries to the intended use determined for the artificial intelligence system and generates a similarity score indicating the similarity of the intent and intended use. Based on a comparison of the similarity score to a similarity threshold indicating that the intent does or does not align with the intended use of the artificial intelligence system, the disclosed systems determine whether to enable or block processing of the one or more queries by the artificial intelligence system. Additionally, the disclosed systems execute computing instructions to enable or block processing of the one or more queries by the artificial intelligence system. The disclosed systems thus utilize content analysis to determine whether queries to an artificial intelligence system align with the intended use of the artificial intelligence system and enable or block access of the one or more queries to the artificial intelligence system. BRIEF DESCRIPTION OF THE DRAWINGS Various aspects will be described and explained with additional specificity and detail through the use of the accompanying drawings. FIG. 1 illustrates an example of a system environment in which an intent analysis system can operate in accordance with some aspects. FIG. 2 illustrates an example of the intent analysis system determining whether to block or enable processing of a query by an artificial intelligence system based on intent of the query in accordance with some aspects. FIG. 3 illustrates an exampl