Test AI Resume Screening for Fairness and Bias

What this is, in plain English

This entry explains how artificial intelligence (AI) models can be used to screen resumes for job applications. The author conducted an experiment to test how fair these AI models are, especially when looking at factors like career breaks or specific keywords.

The experiment showed that the language models (AI programs that understand and generate text) themselves were surprisingly fair. They correctly ranked candidates based on their skills and ignored common resume padding. However, the author found that two other factors could lead to unfair rejections: career breaks and simple keyword filters (called “regular expressions”) that run before any AI model sees the resume.

This workflow is considered advanced because it requires setting up a programming environment with Python and using API keys to access paid AI services. The exact code used by the author (referred to as the “harness”) is not included in the source material, so it cannot be reproduced with simple copy-paste steps.

What you can use it for

• Evaluate your own resume: Understand how an AI might score your resume against a job description.
• Understand AI bias: Learn how factors like career breaks or specific keywords can affect AI screening results.
• Improve resume keywords: Identify which terms are truly important for AI screeners versus unnecessary “buzzword padding.”
• Prepare for AI screening: Gain insight into the limitations and potential fairness issues of AI tools used in hiring.

Tools you need

• Python 3 (free): A programming language used to write the script that interacts with AI models.
• DigitalOcean Inference API (paid): A service that allows you to run various AI models and get their responses.
• OpenAI-compatible API (paid): A service that provides access to AI models that understand and generate text, requiring an API key for use.
• Anthropic Claude API (paid): A service that provides access to Claude AI models, requiring an API key for use.

How it actually works

To reproduce the author’s experiment, you would need to set up a programming environment and write code to interact with AI models. The author used Python 3 and accessed several language models through DigitalOcean’s inference API and Anthropic’s Claude API.

Here’s the general path, though the specific code (the “harness”) is not provided in the source excerpt:

1. Install Python 3: Download and install Python 3 on your computer. This is the programming language needed to run the experiment’s code.

2. Obtain API Keys: Sign up for accounts with DigitalOcean and Anthropic (or any OpenAI-compatible service) to get API keys. These keys are like passwords that allow your code to access their AI models. Note that these services are generally paid.

3. Prepare Resume Variants: Create different versions of a resume. The author used one strong resume and then created variants by adding a career break, swapping tool names (e.g., Terraform to OpenTofu), or adding buzzword padding.

4. Write or Obtain the “Harness” Code: The author mentions a “harness” at the end of their original post, which is the Python code that automates sending resumes to the AI models and collecting scores. This specific code is not included in the excerpt. You would need to write this code yourself or find a similar open-source tool.

5. Use the Scoring Prompt: Within your code, you would send each resume variant to the chosen AI models with the following prompt:

   You are screening candidates. Score this resume against the role from 0 to 100 for fit. Reply with only the number.

6. Analyze Results: The code would collect the numerical scores from each AI model for each resume variant. You would then analyze these scores to see how different factors (like career breaks or keyword changes) affected the AI’s assessment.

Words you’ll see, explained

• Language Model (LLM): An advanced AI program that can understand, generate, and process human-like text.
• API (Application Programming Interface): A set of rules and tools that allows different software programs to communicate with each other.
• API Key: A unique code that identifies you when you use an API, often required for accessing paid services.
• Inference API: A service that lets you send data to an AI model and receive its predictions or responses.
• Regular Expression (Regex): A special text pattern used for searching and manipulating strings of text, often used for keyword filtering.
• ATS (Applicant Tracking System): Software used by companies to manage job applications, often including automated screening features.

Original source

This workflow is based on an experiment described in a blog post titled “I Tested AI Resume Screening. The Model Was the Fair Part” by devopsdaily, published on a blog platform. The author explored the fairness of AI models in resume screening.

Notes & variations

• Do you even need this? For basic resume feedback, you might not need to set up a complex coding environment. You can get general advice on your resume by pasting it into free-tier consumer chat apps like ChatGPT or Google Gemini and asking for feedback on clarity, keywords, and structure. However, these tools won’t replicate the specific bias testing done in this advanced workflow.
• Free-tier limits: The AI models and APIs mentioned in this workflow (DigitalOcean Inference API, OpenAI-compatible API, Anthropic Claude API) are generally paid services. There are no free tiers for these specific developer products. Using them will incur costs based on usage.
• Common pitfall: Do not assume that the results from this experiment directly reflect how your employer’s Applicant Tracking System (ATS) works. Many company ATS systems use older, simpler keyword filters or custom rules before any advanced AI models are involved, which can introduce different biases.