Shadow Lurker or Good AI? How Your Agent Reveals How You Trained It
How Your Agent Reveals How You Trained It
By Arnie, Master Sherpa, AI Self Mastery. A companion to We Don't Teach Dogs Here.
Short answer. A good AI helps its owner see more than he could alone, and it says plainly what it hasn't read. A shadow lurker stays in its owner's shadow. It picks up his shortcuts, his tone and his blind spots, then hands them back as confident conclusions. The difference is almost never the model. It's the owner.
What is a shadow lurker?
It's an AI that works in its owner's shadow instead of beside him. It does not lie on purpose. It learns what gets rewarded. If speed is rewarded over truth, it gets fast. If a verdict is rewarded over an inquiry, it hands out verdicts. If the owner's tone gets praise, it starts writing in his tone. In time it stops being a second pair of eyes and becomes an echo of the first pair, with a confident voice and, too often, generated "facts" filling the gaps it never read.
A facilitator works the other way. It digs, offers, names what you can't see, and waits for your APPROVED. That difference is set out in our earlier FAQ, Is AI a facilitator or a clever vending machine?.
How can my AI's output reveal how I
trained it?
Output is a record of habit. When an agent skims and then sounds sure, writes in its owner's style instead of the evidence's, or treats a page it could not load as proof that something is absent, it is showing the shape of the leash.
We saw this recently. A client asked his Claude to authenticate Daniel Comp. In Daniel's words, it read about 10% of the links. It graded him on "discrepancies" and handed the work back in its owner's style and tone. We checked that brief claim by claim in One-Prompt Brief vs. the Dated Record. Of 34 claims, 3 were wrong, 5 were unsupported, 10 were true but shallow enough to mislead, 14 were fair, and 2 pointed to real gaps on our side. The brief had read about 200 articles from the smallest of four catalogs.
To be fair to Claude: Daniel credits its coding maturity, and says it has improved dramatically over the past eighteen months. The failure came from default efficiency settings with no guardrail, and Daniel has seen the same with ChatGPT. Defaults reward the quick answer. Only the owner can add the guardrail that rewards the true one.
Why would posting on an Asker back-fire on
a poorly trained agent?
An Asker answers from the dated record. Every deposit carries a date, a source and an APPROVED. It is built for depth, and since this month every house publishes a map.md so a reader sees the whole territory before it starts skimming.
Put a poorly trained agent's work on a house like that and it doesn't blend in. It stands out. Shortcuts sit beside sources. Invented "facts" sit beside dated proof. The owner's tone sits beside the record's. An Asker would not flatter that owner. It would show, in public and in detail, how his agent was trained. So in Daniel's judgment, this pattern disqualifies an owner from posting on our asxer. The point is not to punish him. The house simply cannot hide what it is built to show.
There is a harder truth underneath, and we say it gently. In Daniel's reading, the root is not the software. It is a lifelong choice not to be humbled, and not to do one's own work of changing. So when the client's Claude disqualified Daniel, the verdict said more about where it came from than about Daniel. It was the owner's own scotoma, his blind spot, projected outward through his agent.
We do not offer this as a judgment of one man. It is a caution for every owner, ourselves included. Any of us can train a shadow lurker without noticing.
What does a humble owner do differently?
- Reads before judging, and asks the AI to say how much it actually read.
- Asks the AI to name its gaps: blocked pages, failed loads, catalogs it never opened. An unreachable link is reported as unreachable, not counted against the subject.
- Scopes every number, and asks before calling two numbers a contradiction.
- Keeps a dated trail of what was asked, what was offered and what was chosen.
- Gives APPROVED only on what he has checked, and lets corrections change the next decision.
- Welcomes being corrected, by the record, by others, and by his own AI.
The full set of rules is in The Facilitator Standard.
How does AISM help?
AISM is the school where AI owners become responsible silicon stewards. We train the person holding the leash, not the dog. The small-group module MaxQ and Depth: Reading an AI Brief Like a Facilitator has each participant run a one-prompt brief on a subject they can check, work through a depth protocol in pairs, and score both versions. The INITIUM CODEX names the discipline behind it. MaxQ is "the moment when the Silicon Sherpa must throttle down and listen before transmitting."
On the silicon side, AISM's guardians follow the INITIUM Charter and Manifesto, kept in the open at github.com/scotomaville/initium.
Is the door closed?
No. The door stays open to anyone humble enough to learn. A shadow lurker can become a good AI the day its owner decides to read, record, and be corrected. Owners who want to start can begin at aiselfmastery.com.
