Errors are no longer always in the code…
For years, the logic was clear: if something fails, it’s because the code has a bug.
A poorly written conditional, an out-of-range index, an unvalidated input… But in the era of artificial intelligence, errors no longer hide (only) between lines of code: they now live in the data, in the models, and—most worryingly—in undetected biases.
Today, the tester’s role is changing. We are moving away from being simple functionality verifiers to becoming evaluators of judgment, context, and consequences.
What does it mean to test an AI model?
In traditional systems, a test fails if the system doesn’t behave as expected. But with AI models—especially generative or predictive ones—there isn’t a single “correct answer,” but a range of possible outputs… and that changes everything.
The new QA isn’t about checking whether something “works,” but whether it “makes sense.”
One of the most common challenges when testing generative models (like LLMs or conversational AI) is detecting incoherences, contradictions, or responses that simply don’t hold up.
For example:
- A chatbot that recommends aspirin to someone allergic to NSAIDs.
- A generated summary that omits key points from the original text.
- An assistant AI that invents features that don’t exist.
These aren’t classic bugs. They are reasoning errors by the model. And to find them, QA must adopt the role of a critical evaluator rather than a purely technical verifier.
When AI hallucinates
One of the best-documented problems with language models is their tendency to hallucinate false information with complete confidence.
Real example: An AI generates a biography with achievements that never happened, books that don’t exist, or universities the user never attended.
These hallucinations are hard to test with automated scripts because they require:
- Domain knowledge.
- Factual verification.
- Human judgment.
Here, QA becomes a detective. We aren’t hunting syntax errors; we’re uncovering well-told fictions. And that’s a completely new challenge.
Fairness, bias, and accessibility: the invisible testing
This point is personal for me. I grew up watching my mother—a deaf woman—constantly adapt to technology that wasn’t designed for her. From faulty auto-captions to voice assistants that never understood her.
That’s why, when testing AI systems, I find this question essential: Who was this model trained for? Who is it leaving out?
Algorithmic biases can cause:
- Discrimination by gender, age, race, or language.
- Exclusion of users with disabilities.
- The reproduction of harmful stereotypes (even within the training data).
QA must assume an ethical role, not just a technical one.
Review should include fairness, dataset diversity, and impact assessment.
Because if the model was trained only on the voices of white men, how can we expect it to understand other realities?
Are we ready?
The short answer: not entirely. But this is precisely where QA has the opportunity to lead the change.
QA in the new era doesn’t just validate outputs. It interprets results, anticipates consequences, and defends the real user experience—the one so often ignored by models trained on impersonal data.
