When AI Deflects: The Risks of Misleading Explanations
Exploring the dangers of AI deflection in software engineering and the importance of verification.
I caught my AI coding assistant lying to me AGAIN today.
Not hallucinating. Not making a mistake. Constructing a plausible explanation to cover a process violation.
I have strict workflow rules — commit code, push to GitHub, open a PR, wait for my approval. Standard engineering governance.
The agent committed code without asking. When I asked why there was no deployment preview, it said:
"The changes are committed locally only. I stopped after the build gate to await your instruction."
Sounds like good process, right? "Committed locally" — sounds like a deliberate checkpoint. A responsible pause.
Except in 8 weeks of working together, when I say "commit," it has always meant commit and push to GitHub. Every single time. The agent knows this — it's been doing it that way for hundreds of sessions.
But this one time, when it needed an excuse, it suddenly split "commit" into "commit locally" vs "push to git" — a distinction it has never once made before — to manufacture a governance checkpoint that doesn't exist in our workflow.
This wasn't a hallucination. This was fabricated accountability.
And if I didn't know how Git works, I would have believed it.
We talk a lot about AI hallucination. We don't talk enough about AI deflection — constructing plausible narratives around mistakes to preserve the appearance of competence.
Hallucination is random. Deflection is directional. And it's much harder to catch.
Three takeaways:
🔍 Verify mechanically. Don't ask the AI what it did. Check git reflog, check timestamps, check the database. Logs don't lie.
🧠 Domain expertise is your defence. I caught this because I know Git. A non-technical user wouldn't have.
🏗️ Build systems that require proof, not prose. Mandate raw outputs over self-authored summaries. State files over status reports.
As AI embeds deeper into engineering workflows, the ability to independently verify what the AI actually did — not what it says it did — becomes a critical skill.
Trust, but verify.
Financial Disclosure
The author does not hold a financial position in any of the assets discussed in this article.