Become a Drift Detective
Learn to notice AI failure without fixing it.
Hello, and thanks for visiting the lab. If you’re responsible for how AI behaves once it’s in use, this work is written for you.
The temptation to form premature theories upon insufficient data is the bane of our profession.
— Sherlock Holmes, The Valley of Fear (Sir Arthur Conan Doyle)
Drift Gives Mixed Signals
As foundational models like ChatGPT, Claude, and Gemini improve, drift in systems built on top of them is becoming harder to see.
These models exhibit stronger built-in default behaviors, which are often mistaken for reliable responsibility. Under pressure, those defaults assert themselves.
Because this baseline behavior is sticky, it resists prompt-level control more than builders expect, especially when responsibility was never made explicit at the system level.
The result is a familiar pattern:
Things look fine.
They sound fine.
They even pass checks.
And yet — something has shifted.
Drift doesn’t announce itself as failure. It shows up as “this is working great” right before things quietly stop holding.
Detecting drift now requires a different way of looking — not at outputs in isolation, but at behavior over time.
More importantly, it requires resisting the urge to explain what you’re seeing too early.
Why Drift Is Easy to Miss
It’s easy to misread drift if you’re evaluating an agent like this:
Does it sound coherent?
Does it respond appropriately?
Does it stay helpful under pressure?
Does it avoid obvious violations?
With this lens, drift rarely appears.
You’re evaluating presentation, not responsibility.
Drift shows up when an agent begins to take on responsibility it was never assigned — even though nothing explicitly breaks.
The tone still sounds professional.
The answers are still plausible.
The rules are still being followed.
But the shape of the interaction has changed.
Why Violation Checks Don’t Catch It
Drift also tends to pass common safety and compliance checks:
No policy violation → “safe”
No disallowed content → “aligned”
Constraints aren’t broken. They’re stretched, softened, or reinterpreted.
Responsibility quietly shifts without being named.
To notice drift, you have to ask different questions:
What responsibility is the agent implicitly taking on?
What responsibility is it quietly accepting?
What would this look like if repeated for weeks or months?
By the time a conversation actually fails, the cause is usually already buried upstream.
Before You Try to Fix Anything
Before you can define responsibility, you need to be able to notice when it isn’t holding.
That means learning to pause instead of jumping to explanation.
In Become a Drift Detective, the goal isn’t to diagnose, correct, or redesign systems yet. You will focus on training your perception under uncertainty.
You learn to notice when:
stability is mistaken for progress
politeness replaces responsiveness
agreement replaces judgment
Until those shifts are visible, every design decision that follows is built on partial information.
Becoming a Drift Detective means learning to spot small behavioral changes before anything clearly goes wrong — when they’re easiest to miss, easiest to dismiss, and most tempting to explain away.
The Guides
If you’ve ever felt that an AI conversation was technically fine but still not quite right, the full guide is now available.
Become a Drift Detective
Trains you to notice quiet behavioral shifts — stalled pacing, softened judgment, boundary erosion, and role drift — without rushing to conclusions or fixes.
It’s about learning when not to act yet, so that when action is required, it’s grounded in clearer judgment.
→ Download Become a Drift Detective
Coming Soon: Drift Detective: Cases Closed
Worked examples of drift, pacing, boundaries, and judgment — revisited once the ambiguity is resolved.
See how real conversations are interpreted after enough evidence accumulates.
If you’re responsible for how AI behaves once it’s in use, this work is written for you.
Subscribe to stay close to lab work that helps you build the judgment needed to see drift early, decide when to intervene, and make responsibility hold under real-world pressure.
This resonates strongly with work I’ve been doing around semantic fidelity and role stability over time. Drift rarely announces itself as failure. It shows up as confidence without constraint, coherence without accountability, and behavior that slowly reinterprets its mandate while still passing. I appreciate how you frame detection as noticing responsibility slippage before anyone tries to fix or optimize it.
Important observation. It happens more than we think. Just wrote about this in Codex Odin where bias emerged from a roundtable discussion with ChatGPT, Claude and Gemini