I have now sat through multiple vendor demonstrations built around the same supposedly inevitable future that AI will generate so much code, so quickly, that human beings will no longer be able to test it.
To meet this new challenge, we are told we need to build ANOTHER AI system generating tests at exactly the same speed. One machine sprays code into the organization, another sprays test scripts after it, and somewhere in the middle sits a “human in the loop” clicking “approve” like Homer and his bird.
This is the state of AI in testing today.
One vendor slide I saw correctly identified the problem: coding agents lack context, cannot properly observe what happened in the browser and frequently produce brittle tests tied tightly to implementation. “Agents ship output. Output isn’t proof,” it concluded.
Exactly.
The proposed solution, naturally, was more agents.
If the volume and speed of generated code is genuinely beyond human comprehension, then the accompanying test output will be beyond meaningful human review. You cannot claim that people are incapable of keeping up with the code but perfectly capable of evaluating thousands of machine-generated checks, understanding their assumptions, checking coverage, and deciding whether the resulting evidence is trustworthy.
The entire agentic testing confidence game is being played on the extraordinary inversion of responsibility that the less we understand something, the more we are expected to trust it.
Meanwhile back in reality, as previously reported the AI Safety Index gives the best-performing frontier AI company a C+. Three companies failed outright. Existential safety remains the weakest area with reviewers warning that safety rhetoric does not match their actual behavior and benchmarks are likely understating real-world harm.
So naturally, the software testing industrial complex response is to reduce scrutiny and accelerate deployment.
We should be testing AI-generated code more, not less. We need deeper investigation, stronger independent evaluation, better operational monitoring and people with enough time and authority to challenge what the code extruding machines produce.
Testing is not a manufacturing race against coding. Its purpose is not to generate test-shaped material at a matching velocity. Its purpose is to discover risks that matter before reality does it for us.
Because eventually, someone is going to get hurt by this recklessness and when they do, I guarantee you there will be no shortage of automatically generated evidence proving that everything passed.
Discover more from Quality Remarks
Subscribe to get the latest posts sent to your email.