Granted Patent B2

US 12,481,916 · App. 17/652,754 · Granted Nov 25, 2025

Inventors: Nicholas Tsang (Saratoga, CA); Dana L Price (Surf City, NC); Diane Chalmers (Rochester, MN); Andrew R. Freed (Cary, NC)

Assignee: International Business Machines Corporation

G06N20/00G06F40/205G06F40/40

Abstract

Claims (90)

1 . A computer system for processing challenges to potential artificial intelligence (AI) biases comprising:

one or more processing devices;

one or more memory devices communicatively and operably coupled to the one or more processing devices;

a bias evaluation engine, at least partially resident within the one or more memory devices, configured to:

inject one or more first values associated with one or more respective variables into an AI platform during runtime;

capture and list one or more features of a respective algorithm of the AI platform;

generate, through the one or more features of the respective algorithm of the AI platform, one or more first outcomes, further comprising:

altering the one or more first values which are used by an application machine learning (ML) scoring model to generate one or more revised outcomes;

producing details of the one or more revised outcomes; and

transmitting the details of the one or more revised outcomes via a summary metrics output module;

construct a bias feature-to-AI feature mapping between one or more bias features and the one or more features of the respective algorithm of the AI platform;

use the bias feature-to-AI feature mapping generated through a preparation portion to identify at least a portion of the one or more respective variables in each of one or more bias factor variables that are either directly or indirectly related to the one or more bias factors;

collect feedback with respect to the one or more first outcomes;

parse the feedback through a natural language understanding (NLU) application;

determine, subject to the parsing, one or more bias factors embedded within the AI platform; and

resolve any potential biases to mitigate any inadvertent effects.

2 . The computer system of claim 1 , wherein the bias evaluation engine is further configured to:

inject the one or more first values associated with the one or more respective variables into a machine learning (ML) model.

3 . The computer system of claim 1 , wherein the bias evaluation engine is further configured to:

transmit the one or more first outcomes; and

generate a natural language claim of one or more biases embedded within the one or more outcomes.

4 . The computer system of claim 1 , wherein the bias evaluation engine is further configured to:

identify at least a portion of the one or more respective variables in each of the one or more bias factors.

5 . The computer system of claim 4 , wherein the bias evaluation engine is further configured to:

generate one or more second values for each of the one or more respective variables for each of the one or more bias factors; and

inject the one or more second values associated with the one or more respective variables into the AI platform.

6 . The computer system of claim 5 , wherein the bias evaluation engine is further configured to:

generate, through the AI platform, one or more second outcomes.

7 . The computer system of claim 6 , wherein the bias evaluation engine is further configured to:

produce details of the one or more second outcomes, including one or more of:

summary metrics;

comparisons with the one or more first outcomes; and

natural language explanations; and

transmit the details of the one or more second outcomes.

8 . A computer program product embodied on at least one computer readable storage medium having computer executable instructions for processing challenges to potential artificial intelligence (AI) biases, that when executed cause one or more computing devices to:

inject one or more first values associated with one or more respective variables into an AI platform during runtime;

capture and list one or more features of a respective algorithm of the AI platform;

generate, through the one or more features of the respective algorithm of the AI platform, one or more first outcomes, further comprising:

altering the one or more first values which are used by an application machine learning (ML) scoring model to generate one or more revised outcomes;

producing details of the one or more revised outcomes; and

transmitting the details of the one or more revised outcomes via a summary metrics output module;

construct a bias feature-to-AI feature mapping between one or more bias features and the one or more features of the respective algorithm of the AI platform;

use the bias feature-to-AI feature mapping generated through a preparation portion to identify at least a portion of the one or more respective variables in each of one or more bias factor variables that are either directly or indirectly related to the one or more bias factors;

collect feedback with respect to the one or more first outcomes;

parse the feedback through a natural language understanding (NLU) application;

determine, subject to the parsing, one or more bias factors embedded within the AI platform; and

resolve any potential biases to mitigate any inadvertent effects.

9 . The computer program product of claim 8 , further having computer executable instructions to:

inject the one or more first values associated with the one or more respective variables into a machine learning (ML) model.

10 . The computer program product of claim 8 , further having computer executable instructions to:

transmit the one or more first outcomes; and

generate a natural language claim of one or more biases embedded within the one or more outcomes.

11 . The computer program product of claim 8 , further having computer executable instructions to:

identify at least a portion of the one or more respective variables in each of the one or more bias factors.

12 . The computer program product of claim 11 further having computer executable instructions to:

generate one or more second values for each of the one or more respective variables for each of the one or more bias factors; and

inject the one or more second values associated with the one or more respective variables into the AI platform.

13 . The computer program product of claim 12 further having computer executable instructions to:

generate, through the AI platform, one or more second outcomes.

14 . A computer-implemented method for processing challenges to potential artificial intelligence (AI) biases comprising:

injecting one or more first values associated with one or more respective variables into an AI platform during runtime;

capturing and listing one or more features of a respective algorithm of the AI platform;

generating, through the one or more features of the respective algorithm of the AI platform, one or more first outcomes, further comprising:

altering the one or more first values which are used by an application machine learning (ML) scoring model to generate one or more revised outcomes;

producing details of the one or more revised outcomes; and

transmitting the details of the one or more revised outcomes via a summary metrics output module;

constructing a bias feature-to-AI feature mapping between one or more bias features and the one or more features of the respective algorithm of the AI platform;

using the bias feature-to-AI feature mapping generated through a preparation portion to identify at least a portion of the one or more respective variables in each of one or more bias factor variables that are either directly or indirectly related to the one or more bias factors;

collecting feedback with respect to the one or more first outcomes;

parsing the feedback through a natural language understanding (NLU) application;

determining, subject to the parsing, one or more bias factors embedded within the AI platform; and

resolving any potential biases to mitigate any inadvertent effects.

15 . The computer-implemented method of claim 14 , wherein the injecting one or more first values associated with the one or more respective variables into the AI platform comprises:

injecting the one or more first values associated with the one or more respective variables into a machine learning (ML) model.

16 . The computer-implemented method of claim 14 , wherein the collecting feedback with respect to the one or more first outcomes comprises:

transmitting the one or more first outcomes; and

generating a natural language claim of one or more biases embedded within the one or more outcomes.

17 . The computer-implemented method of claim 14 , wherein the determining one or more bias factors comprises:

identifying at least a portion of the one or more respective variables in each of the one or more bias factors.

18 . The computer-implemented method of claim 17 , further comprising:

generating one or more second values for each of the one or more respective variables for each of the one or more bias factors; and

injecting the one or more second values associated with the one or more respective variables into the AI platform.

19 . The computer-implemented method of claim 18 , further comprising:

generating, through the AI platform, one or more second outcomes.

20 . The computer-implemented method of claim 19 , further comprising:

producing details of the one or more second outcomes, including one or more of:

summary metrics;

comparisons with the one or more first outcomes; and

natural language explanations; and

transmitting the details of the one or more second outcomes.

Assignments (1)

Continuity (1)

References Cited (28)