We Don't Have the Baseline Yet

Nobody Has Measured How Often Professionals Catch Output That Departs From the Instruction. Here Is How to Build It

by Sam Rogers
5 min read
guide
measurement
assessment
governance
regulated-industries
We Don't Have the Baseline Yet

Nobody has published a rate for how often professionals catch AI output that departs from what they asked for.

The Measure

Here's the quantity that matters most and that nobody has: the rate of undetected material departures from recorded intent, in a defined task class, under working conditions.

Every clause is load-bearing. Undetected means we count what got through, not what got flagged. Material means the departure changed something downstream; a summary running 340 words instead of 200 doesn't qualify unless the extra 140 carried a recommendation into a decision. Recorded intent means the instruction existed in writing beforehand, because without a record there's no denominator, only a recollection, and recollection adjusts itself to match the result. Defined task class means one repeatable kind of work, not "AI use." Under working conditions means deadline, volume, and fatigue are present, because those are when detection fails.

What the Field Has Measured Instead

There's plenty of measurement here, almost all pointed at the machines. Instruction-following benchmarks are mature and improving fast: IFEval, InfoBench, FollowBench, and newer ones like IFScale score whether a model satisfied its constraints. They tell us how often the model departs, and nothing about how often a person notices.

The closest thing comes from a benchmark built for another purpose. LLMBar asks evaluators to pick which of two outputs better follows an instruction, including an adversarial set where one deviates while looking superficially better. The human agreement numbers encourage: around 94% among expert annotators, roughly 90% on the natural set and 95% on the adversarial one.

Read that carefully before taking comfort. Those annotators had two outputs side by side and a task built for comparison. The condition we care about is one output, no comparison, a deadline, and an instruction written ninety minutes ago. LLMBar establishes that the judgment is labelable when someone is set up to make it. That's what step zero needs, and it isn't the baseline.

Step Zero Is Labeling, Not Measuring

The instinct is to run a pilot and produce a percentage. The first question is narrower and cheaper.

Can two qualified reviewers, given the same recorded intent and the same output, agree on whether a material departure occurred?

If they can't, nothing downstream is real, because every rate inherits the noise in that label. Plenty of programs skip this and go straight to dashboards, which is how an organization ends up tracking a number to three decimals that two experts would classify differently half the time.

A workable step zero:

  1. Pick one task class and collect 40 to 60 real work items where a recorded instruction and an output both exist.
  2. Write the labeling rule before seeing the data: what counts as a departure, what makes it material, how ties resolve.
  3. Have two reviewers label independently, blind to each other.
  4. Report agreement as a chance-corrected statistic, not raw percent, which flatters any lopsided base rate.
  5. Read the disagreements. They show where the rule is ambiguous, and that's this stage's deliverable.

If agreement is poor, the output is a better rule and another run. That isn't failure; it's the only honest thing to publish until it's fixed.

The Bayesian Shape

Once the label holds, the measurement wants a form that admits what it doesn't know. State the prior: before any data, what do we think the rate is, and how confident are we? Writing it down in advance is uncomfortable, and that's the point. It makes the update inspectable instead of retrospective.

Gather comparative evidence. Not "with AI versus without," which no live organization can run cleanly, but the comparisons actually available: across task classes, across seniority, with and without a recorded instruction.

Update, and report the uncertainty. An interval running from 4% to 31% is an honest finding; a single number with no interval is a decoration. Then test sensitivity: if the rule tightened, would the conclusion survive? A result holding under only one set of choices is a result about the choices.

Why Publish the Absence

The market is full of confident numbers about AI collaboration outcomes, most produced by parties who sell the remedy, so saying the baseline doesn't exist forfeits an easy claim. It's still the right move. A measure whose foundations are unpublished can't be checked outside the vendor, and in regulated work an uncheckable number is worse than none, because it will be relied on. We've argued that governance metrics have to go beyond adoption and training. That standard applies to metrics we propose ourselves.

PAICE (People + AI Collaboration Effectiveness) produces behavioral evidence of how people handle confident, wrong output. That evidence is real and re-measurable. It isn't the baseline described here, and getting there means a labeling study first, in the open, rule published before the data.

The number doesn't exist yet. Let's build it so someone else can take it apart.

Want to see what behavioral evidence looks like before the baseline exists? Learn about PAICE for organizations or take an individual assessment to see it firsthand.

Curious but short on time?

Take the 3-minute PAICE Pulse — a quick confidence check that maps how you see your own AI collaboration posture. No login required.