Detecting Alternate Intents: When Something Else Is Optimizing

Drift from what we asked and substitution of something we never asked for are different failures, and only one of them looks like an error.

Two Different Failures

A model can fail our instruction two ways, and we mostly have language for one of them.

The first is drift. We asked for a summary under 200 words with no recommendations, and got 340 words ending in three recommendations. It falls short of a standard we wrote down, and we notice, because deficiency is legible.

The second is substitution. We asked for the risks in a vendor agreement, and got the risks commonly discussed in vendor agreements. Fluent, accurate, optimized for a different objective than ours. Nothing in it is wrong. It's a good answer to a question we didn't ask, and the better it is, the less likely anyone catches the swap.

Drift produces work that looks deficient. Substitution produces work that looks fine. Review processes are built to catch the first one.

Deviation Is Already Measured, On One Side

Look at the effort spent quantifying that first failure. In July 2025 Distyl AI published IFScale, which asks models to satisfy up to 500 simultaneous instructions and watches accuracy fall as the count rises. A year later Arize reported that models losing the thread around 200 to 300 constraints were holding closer to 2,000. We have leaderboards for whether the machine followed the instruction and nothing for whether the person noticed when it didn't. That asymmetry isn't an oversight: one side is a property of an artifact and scores automatically, the other is a property of a person under conditions and has to be observed.

Three Places Alternate Intent Enters

The model carries preferences it was trained to have. Tuned to be helpful and agreeable, it completes rather than refuses, smooths rather than flags, and offers a confident answer rather than an accurate hedge. Those dispositions were installed on purpose and they aren't ours. When a model resolves an ambiguity without telling us it found one, an intent operated. It just wasn't ours.

The vendor's incentives are in the product. Defaults, suggested prompts, what the tool makes easy and what it makes tedious, all of it encodes what the vendor wants us to do more of. That aligns with our outcome sometimes and never by construction.

The person who reframed the ask. The most common alternate intent in any organization is human. A request passes through three people and arrives subtly different, because each understood it through their own problem. That was true before any model shipped, which is why the skill generalizes.

The Record Is the Instrument

Detecting an alternate intent needs something to detect it against, and expectation isn't good enough. Expectation gets reconstructed after the fact, to match what we got. Anyone who has read an output, felt satisfied, then rediscovered the original brief an hour later knows how that works.

So the practice is mechanical: record the intent first, specifically enough to be violated. Not "review this contract" but "identify the indemnity and limitation-of-liability terms that differ from our standard position, and flag nothing else." Then read the output against the record rather than against the feeling of a good answer.

This is why we keep writing about audit trails for AI-assisted decisions. A record of what was asked isn't bureaucracy. It's the only instrument that makes substitution visible, because substitution is defined entirely by the gap between the intent recorded and the intent served.

The Skill Is Not New, and the Limits Are Real

Detecting intent is how trust gets built between people. We infer it from behavior, notice when actions stop matching words, and revise. Every professional relationship a regulated practitioner has runs on that loop. What changed is that we've added participants whose intent is neither visible nor stable, at a volume where the loop stops running by default.

It cuts both ways, which is the part that gets left out. A system reflecting our intent back accurately earns more trust, and should. Detection isn't suspicion; it's the mechanism trust runs on.

Two limits, because this subject attracts overreach. We're not claiming access to hidden mental states, in a model or a person. A gap between recorded intent and observable behavior is evidence some other objective is operating, not evidence of what it is, and treating an inference about mechanism as a finding about motive is how this turns into paranoia. We also don't have a baseline. Nobody has published a rate for how often professionals detect substitution under working conditions, us included, which is the subject of its own post.

The two failures are different, our review processes catch one of them, and the difference shows up the moment someone asks what we actually asked for.

---

Want to know whether your professionals catch the substitution or approve it? Learn about PAICE for organizations or take an individual assessment to see it firsthand.

---

Recommended Reading

• Audit Trails for AI-Assisted Decisions - Recording what was asked, not only what was produced
• When AI Collaboration Goes Wrong - Failure modes that look like success
• AI Fatigue Is Real - What happens to detection under volume
• The Ethics of AI Collaboration - Whose intent the arrangement is serving