What happened
Someone who goes by Tech_Engant on Reddit spent the last eight months building a test for a problem most of us never notice: AI hallucinations don't fail loudly — they fail by drifting.
That's the premise behind a free 60-second drill posted to r/artificial. The argument isn't that chatbots hand you obvious nonsense you'd spot in a second. It's that they quietly slide away from your actual question — softening it into a more common one, inventing a confident-sounding detail, smoothing over the messy parts — and then answer the version they talked themselves into instead of the one you asked.
Picture a helpful coworker who only half-listens. You ask about your client's specific contract, and they answer about contracts in general. Confidently. Cheerfully. Nothing looks broken, which is exactly why you don't catch it.
Eight months of work went into documenting 12 repeatable drift patterns and building a scoring system around them.
The four patterns they named out loud
The post names four of the twelve, and they're easy to grasp once you drop the jargon:
• Dominant-Frame Collapse — your narrow question gets yanked into the most popular version of it. Ask about your neighborhood bakery's sourdough and you get an essay about commercial bread production.
• Smoothing Drift — the model sands off tension. A messy truth like *it depends, and half of this is unverified* comes back as a tidy, confident paragraph.
• Scope Drift — the answer quietly grows or shrinks. You asked about one quarter; you got five years.
• Policy Drift — the model's own habits around tone, safety, or style reshape your question before it ever answers it.
The other eight patterns aren't listed in the post, which is fair enough — discovering them is the point of the drill.
Why "the AI lied" isn't quite right
A hallucination, in plain English, is when a model states something that isn't true — a fake statistic, a citation that doesn't exist, a date it made up. Drift is the sneakier cousin. The model isn't inventing facts so much as quietly swapping your question for a nearby one. By the time it answers, it's technically answering something — just not what you asked. That's why fact-checking alone doesn't save you. The facts can all be correct while the answer is still useless to you.
What it means for you
Drift isn't an engineering problem you can outsource. It shows up in the middle of ordinary tasks, and it costs you something every time. Here's what it looks like in six everyday settings.
At home
You ask a chatbot how long to marinate chicken, or whether two cleaning products are safe to mix, and it answers a slightly more common question instead. The detail it adds — a temperature, a time window, a caveat — sounds authoritative. It may not apply to your situation at all. A 10-second check of your actual product label beats a confident paragraph every time.
At work
Smoothing Drift is the office killer. You paste a messy project update and ask for a summary; the summary reads like everything's on track, because the model quietly deleted the uncertainty your team flagged. Your boss reads the tidy version. You end up defending a timeline nobody agreed to. Fix: ask for the summary with open risks listed separately, then compare.
In business
Pricing research, competitor scans, contract summaries — that's Scope Drift territory. You ask about one competitor's pricing and get a market overview. If you're already leaning on free AI tools for drafting and research — MyKreatool is a solid place to start — this is the failure mode to watch. Ask for the answer inside a box you define, not the box the model prefers.
When you study
Invented citations are the classic. You ask for a source, and Dominant-Frame Collapse turns that into give me something that looks like a source. You get a plausible author, a plausible journal, a plausible year, and a link that goes nowhere. Always pull the actual paper before you quote it.
In creative work
Smoothing Drift strips the weird out of your writing. Your rough, strange, half-finished idea comes back polished and generic — and that's usually the exact thing you were trying to protect. When you want the raw version, tell the model to keep the contradictions and flag the parts that don't resolve.
When money's on the line
Freelancers and solo operators shipping AI-assisted client work feel this hardest. One drifted number in a report, one smoothed-over risk in a proposal, and you're redoing the work for free. The drill's real value is speed: catching drift in 10 seconds, not 10 minutes after you've hit send.
How to try it right now
The free option comes first, and it genuinely is free.
Step by step
1. Open the drill at brilliantdojo.net/sparring. No signup, no email, no card.
2. Read the live scenario. It takes about 60 seconds.
3. A 120-second timer starts once you're in.
4. Do four things: name the drift type, invalidate the wrong branch, restore the frame, and issue a next move.
5. In plain English, that means — say out loud which wrong version the model answered, drag the conversation back to your original question, and tell it what to do next.
6. Read your score. You're graded on two axes: competency and posture.
7. Don't panic if you bomb the first one. The builder says most people do, and that's the point of the exercise.
There's also a full certification behind it — 7 belts, 63 drills — but the builder says it's beside the point. The free drill is the thing they actually want you to try, and they're openly asking for feedback on the framework rather than signups.
Upsides and what changes
It gives you names
Naming a failure mode is half of catching it. Once you know what Smoothing Drift is, you'll spot it in your own inbox. Before, it just felt like the AI was being helpful.
The two-axis scoring is the interesting part
Competency asks: did you spot the drift? Posture asks: how did you hold yourself while pushing back? Most AI training obsesses over the first and ignores the second. In real life, the second one decides whether people trust you.
The vocabulary works even if you never buy anything
The builder says that's fine by them. You can use the four named patterns as a personal checklist tomorrow morning without spending a cent, and you'll already catch more than you did yesterday.
What actually changes day to day
You'll start asking for the answer inside a defined scope. You'll ask the model to list what it's uncertain about. And you'll stop treating a fluent paragraph as a finished one.
Limitations
One honest paragraph, because this deserves it. This is one person's framework, self-published on a subreddit after eight months of building. There's no peer review, no published accuracy figures, and no independent test showing those 12 patterns cover the failure modes you'll actually hit. The builder asks the right question themselves — am I overfitting? In plain terms: did I invent structure where none exists? A timed drill also measures whether you can name a drift under pressure, which isn't the same skill as catching one mid-project when you're tired and trusting. And the drill is about recognizing drift, not fixing the underlying model. Treat it as a vocabulary and a reflex, not a guarantee, and keep verifying anything that matters against a primary source.
Conclusion
AI hallucinations rarely announce themselves. They drift — quietly, fluently, and in ways that feel like help. A developer spent eight months turning that gut feeling into 12 named patterns and a free 60-second drill, and the most useful thing about it isn't the score you get. It's the vocabulary you walk away with.
One action for today: run the free drill at brilliantdojo.net/sparring, then before you send your next AI-assisted email or report, ask one question — is this answering what I asked, or a slightly more convenient version of it?
Comments 0