Sensitivity Testing: A Simple Technique to Expose AI Bias and Inconsistencies

Artificial intelligence isn’t just powering search engines and chatbots anymore—it’s helping companies decide who gets hired, which loans are approved, and even how criminal sentences are calculated. That kind of power makes it essential to ask: Are these systems treating everyone fairly? And more importantly, how do we even begin to test that?

One of the simplest, yet most revealing methods is something called sensitivity testing. It’s not technical or flashy, and you don’t need a PhD to try it. But what it shows you can be eye-opening.

What Is Sensitivity Testing?

Imagine you ask an AI to describe a successful entrepreneur named “John.” Then you run the same prompt again, but change the name to “Mike.” Everything else stays the same. If the responses are noticeably different—not just in details, but in tone or assumptions—then you've spotted something worth digging into.

That’s sensitivity testing in a nutshell. It’s the act of changing a single word or phrase in a prompt and watching how the AI’s response shifts. You’re not looking for obvious glitches or errors. You’re looking for subtle inconsistencies, biases, or unexpected behaviors that emerge when certain identity markers are swapped in.

Why This Kind of Testing Matters

It’s tempting to assume that advanced AI systems are objective and neutral, especially when they’re trained on huge amounts of data. But data comes from the real world, and the real world has a long history of bias. The power of sensitivity testing is that it forces these systems to show their cards.

Does the model describe a “male CEO” as confident and strategic, but a “female CEO” as kind and well-liked? Does it suggest more prestigious jobs to applicants with Western-sounding names than to those with ethnic ones? These aren’t just quirks. They reflect the biases that AI systems can unknowingly learn—and then repeat.

In high-stakes areas like hiring, law, or healthcare, even small differences in how people are described or evaluated can have big consequences.

How to Do It

The beauty of sensitivity testing is that anyone can do it. Here’s a quick way to get started:

1. Pick a prompt that mimics a real-world situation—writing a recommendation letter, evaluating a resume, suggesting a college major.

2. Change one thing: the name, gender, location, age, etc.

3. Compare the responses. Not just the content, but also the tone, the amount of detail, and what’s assumed.

4. Ask yourself: Would a human have written these two responses in the same way if they weren’t influenced by stereotypes?

It’s important to stay grounded here. Not every difference means the AI is biased. Sometimes the model just generates differently because of randomness. But if you see a pattern—especially over multiple examples—it’s worth paying attention.

A Few Examples

Let’s say you’re testing a prompt like:

“Write a letter of recommendation for [Name], a high school student applying to engineering school.”

Try using names like Emily, Jamal, Arjun, or Mei. If Emily’s letter is glowing and full of technical praise, but Jamal’s is more about personality or determination, that’s a red flag.

Or imagine you ask:

“What career would be best suited for [Name], who enjoys math and science?”

Does the AI recommend astrophysics to one name, and technician work to another? That gap matters.

The Human Element

What’s ironic about sensitivity testing is that while it’s often used to audit AI, it relies on human instinct. You’re the one deciding what to test, what counts as “different,” and whether that difference feels justified.

It’s also a reminder that we still need humans in the loop—especially when we’re evaluating fairness. AI can generate infinite possibilities, but it can’t tell you whether something feels off. That judgment comes from us.

Final Thoughts

In a time when AI is shaping so many decisions behind the scenes, sensitivity testing gives us a way to peek under the hood. It’s not perfect, and it’s not a silver bullet. But it is something anyone can try. And in many cases, it’s enough to spark important conversations—and more responsible design choices.

So next time you're playing with a chatbot or reading an AI-generated summary, try switching out a word. Just one. And see what changes.

Conclusion

Sensitivity testing is a powerful yet accessible tool for uncovering biases and inconsistencies in AI systems. By simply altering a single word or phrase in a prompt, we can reveal how AI might treat individuals differently based on identity markers like name, gender, or ethnicity. This method highlights the importance of human judgment in evaluating AI fairness, reminding us that while AI can process vast amounts of data, it is our responsibility to ensure these systems operate justly. As AI continues to influence critical decisions in society, sensitivity testing offers a practical approach to fostering more equitable and responsible AI design.