--- schema_version: "1" id: 7472f71423a720a686096bce7ee0486d3ead45e9f846921bdf422a2f07320d05 title: Design Ethical AI Models collection: admin version: current locale: en-us portal_url: https://docs.salesforce.com/admin/current/en-us/data/264-0-0/c360_a_ai_ec_ethical_model_building.md canonical_url: https://help.salesforce.com/s/articleView?id=data.c360_a_ai_ec_ethical_model_building.htm&release=264.0.0&type=5 source_site: help.salesforce.com content_type: documentation content_hash: 263e8a276bd2675a2c0477322a6f5cf6e8958f1542c5dab58f93274137ef3a34 description: Salesforce offers guidance to help you practice ethical use of artificial intelligence (AI) in your business. products: - Data 360 release: "264" --- # Design Ethical AI Models Salesforce offers guidance to help you practice ethical use of artificial intelligence (AI) in your business. ## Ethical AI at Salesforce Ethical AI is an essential part of Salesforce's commitment to the ethical and humane use of technology. For an overview, check out these resources. - [Ethical and Humane Use](https://www.salesforce.com/company/ethical-and-humane-use/) at Salesforce - [Responsible Creation of Artificial Intelligence](https://trailhead.salesforce.com/en/content/learn/modules/responsible-creation-of-artificial-intelligence) Trailhead module ## Bias in Modeling Data Bias in data can lead to unfair or inaccurate predictions that can unintentionally disadvantage certain groups or skew business outcomes. In AI Models (formerly Einstein Studio), building ethical models starts with understanding where bias exists in your data and taking proactive actions to minimize its impact throughout the model lifecycle.
Binary or Multiclass Models
BIAS TYPE SCENARIO Recommended Action
Historical

Past data reflects existing inequalities or human decisions that weren't objective.

For example, a model that predicts customer prioritization can perpetuate service quality imbalances by continually favoring existing customer profiles. Historical support data indicates that certain customers received faster resolutions due to perceived importance or spending power.

Before deployment, audit or test models that use diverse datasets and fairness metrics to check for bias.

To detect and address misclassifications, implement a human-in-the-loop review process for critical decisions.
Variable (Feature)

Input variables that are used to train a model unfairly influence predictions or reveal hidden bias. They’re introduced by choice or are a representation of variables in your data. Bias occurs when certain attributes, such as location, age, or gender, disproportionately influence predictions.

For example, a model that predicts customer satisfaction uses the “preferred communication channel” variable. Some customer segments prefer specific channels (chat, phone, or email) because of age, accessibility, or cultural factors. The model misjudges satisfaction for groups that engage through less-dominant channels.

Before deployment, audit or test models that use diverse datasets and fairness metrics to check for bias.

Remove or modify variables that directly reference sensitive attributes, or replace them with less biased proxies.

Change decision thresholds for underrepresented groups to mitigate unfair outcomes.

To detect and address misclassifications, implement a human-in-the-loop review process for critical decisions.
Labeling

Outcomes in data are inconsistently or unfairly labeled due to inconsistent, inaccurate, or subjective judgment.

For example, a model that predicts churn uses the churn label “no purchase in 90 days”. However, some customers can have long buying cycles, such as seasonal shoppers. The model predicts that these customers “churned” even though their consumer behavior is normal for their segment.

Audit labels to ensure that they accurately reflect the real-world outcome.

Exclude data points where labels are clearly inconsistent or subjective. Ensure that segments that were misrepresented have sufficient examples with correct labels.
Sampling

Training data isn’t diverse enough to represent customers or use cases.

For example, a model that predicts the likelihood to purchase overestimates engagement or satisfaction scores. Because inactive or churned customers are underrepresented, the model hasn’t learned from less active customer segments.

Audit training data to identify which segments of your customer base or use cases are underrepresented. Also, compare model performance across different segments to identify disparities.