Generative AI vs Predictive AI: Text Against a Value

A predictive component returns a value: a score, a class, a probability. A system acts on it, it is checked by comparing many outputs against what actually happened, and it fails by drifting. A generative component returns material. A person acts on it, it is checked one output at a time against the material it was given, and it fails by asserting something nothing supports.

Generative and predictive components get grouped together because both were built by related means, and almost nothing anybody deploying them has to decide follows from that shared origin. Every decision a deployer does have to make follows from something much plainer: what comes out, and who reads it. Every practical difference between the two traces back to the shape of the output rather than to the method that produced it. A difference resting on the output will still hold when the methods have moved on.

What does each of the two actually produce?

A school makes the plainest starting point. Both objects are familiar there. At the end of a term a child comes home with two things on one sheet. There is a mark, say 62 out of 100. And there is a comment: three sentences from a teacher about how the term went. Nobody confuses the two, and nobody would dream of handling them the same way.

The mark is a value. The mark gets used without being read in any meaningful sense: it goes into an average, it crosses a cut-off for the next set, it decides a prize list. A clerk copies it into a column. If the marking got quietly harsher this year, no single mark looks wrong, and the only way anybody would find out is by lining up many years of marks against what those students actually went on to do.

The comment is material. Somebody reads every word of it, and a parent will hold on to one sentence of it for a decade. If the comment says the child won the reading prize and she did not, that is instantly and obviously wrong to exactly one reader, and it never shows up in any average anywhere.

A predictive componentOne that returns a value such as a score, a class or a probability, rather than something written out in words. produces the mark, and a generative componentOne that produces new material, usually text a person can read straight off. produces the comment. That is the whole of the distinction, and the four rows that follow from it are the ones anybody deploying either has to settle. How either was built appears in none of those rows. Grouping them by method therefore puts two very different objects under one heading.

Sumeru Bank Limited, an invented mid-sized Indian bank, runs both inside one retail personal loan intake chain. Component 6, its scoring model, returns a value. Component 8, its drafting assistant, returns paragraphs. Set the two side by side and the differences read off.

Who acts on the output, and how long does that take?

Follow one file through Sumeru Bank Limited's chain and time the two outputs. Component 6 returns its value and the workflow router acts on it immediately. Nobody opens it, nobody reads it, nobody forms a view about it. On the straight-through path the whole chain returns an answer to the applicant in about four minutes. No part of that chain has a person considering the score, so consideration takes up none of the four minutes. In month six, the first steady month, the arrangement determined the outcome of 5,981 of the month's 8,600 decisions, or 69.5 per cent.

Component 8 returns its draft and then the file stops moving. A desk officer opens the note, reads it against the material it was drafted from, works through five numbered checks and signs it. The reading takes nine minutes, measured. The desk does it 3,010 times a month, once for every file routed to a person. Multiplied out, the desk spends 27,090 minutes, or 451.5 hours, on reading in a month. The predictive component consumes zero minutes of human attention and the generative one consumes 451.5 hours a month, and that gap is not a detail of staffing, it is the whole governance difference between them.

Ismail Sheikh, who runs the exception desk, describes the difference the way any operations head would. The score is something his desk lives with. His people's names go on the bottom of the draft, so his desk is answerable for the draft sentence by sentence.

What does each one need supplied before it can work at all?

Think about a wedding. Two jobs land on the same table. The first is working out how many people will actually turn up, so the caterer can be told a number. The second is writing the card that goes to one particular aunt who has been unwell. Both are hard, and they are hard in completely different ways.

The headcount needs the past: how many were invited to the last four weddings in the same circle, and how many walked in. Nothing about this particular aunt helps at all. The card needs the present: her name spelled right, what she is recovering from, whether she came last time. Four previous weddings supply nothing that can go in that card.

A predictive component needs past cases with outcomes attached, and a generative component needs the material for this one output, in front of it, at the moment it runs. Those two requirements do not substitute for each other in either direction, and each one is a real project before anything can be deployed.

Sumeru Bank Limited fitted component 6 on a past window holding 3,00,000 applications. Only 2,40,000 of them, or 80.0 per cent, were accepted and therefore had an outcome anybody could observe. The other 60,000 were declined, and a declined application never shows whether it would have gone bad. The bank then had to choose what bad even meant, and it chose 90 days past due observed over 12 months, both of which are its own choices and are not a standard anybody set. On those definitions 8,160 of the 2,40,000 carried a bad label, or 3.4 per cent. A past that has already finished happening is the supply side of a value.

Component 8 needs nothing of the sort and could not use it. The one thing component 8 needs is the file in front of it now. The material supplied with one exception note averages about 4,800 words, and the longest five per cent run past 12,000. Alongside it sits a store of 11,400 policies, procedures and product terms, cut into 47,000 passages, from which 6 passages are pulled for each question the drafting step has to answer. The store, the cutting and the pulling exist to solve a problem the predictive side does not have: assembling, freshly and for every single output, the material that output is allowed to rest on.

How is each one checked, and by whom?

Checking sounds like one activity with two settings. Checking is in fact two activities, and they share no equipment. A population checkChecking a component by setting many of its outputs against what actually happened afterwards, rather than looking at any one of them. gathers a lot of outputs and sets them against what happened afterwards. Nobody reads any single output while doing it. An output checkChecking one output on its own, against the material that output was given, without waiting for any later event. takes one output and reads it against the material that output was handed. Nobody waits for anything to happen while doing it.

Component 6 gets the first. Revathi Balan, as the named accountable person for it, does not review a score; there is nothing in a score to review. She reviews a distribution: what share of files were accepted this month against last, how the accepted files behaved once they were on the book, whether the shape of the output has moved. The population check has one awkward property built into it. Nothing can start until outcomes exist, and on this bank's own definition an outcome takes twelve months to arrive.

Component 8 gets the second, and it happens 3,010 times a month at the desk rather than once a quarter in a review. The nine minutes is five numbered checks: every figure traced back to the supplied material, every assertion traced to a passage, nothing asserted that the material does not contain, the required format and statements present, and the decline-to-answer case handled properly where it arises. Asking the reader to notice something that is not there makes the third of those five the slow one, and nothing in any document is harder to notice.

Notice what settles which check a component needs. Not how it was built, not what it is called on the register, not which decade it came from. One question about the output, asked before anybody opens a policy: does a person read this output before it takes effect?

How does each of the two fail?

Two words name the two failures, and holding them apart is most of the work. The predictive side fails by driftA learned component's behaviour moving over time although nothing inside the component itself has changed.. The generative side fails by assertion failureAn output stating something the material it was given does not support, said in the same confident tone as everything else in the output..

Drift means the component did not change and its behaviour did. Nobody edited it, nobody refitted it, nobody approved anything. The change sat upstream of the component, in the data arriving. At Sumeru Bank Limited an income field changed format on one channel in month 8, and the reading step downstream of that change began handing the scoring model a slightly different picture of the same applicants. The component behaved exactly as it always had on what it was given. The input was something else.

Assertion failure means the output states something and nothing in the file supports it. Not a typing error, not a confusion of tone: a sentence that reads exactly like the true sentences around it and is not anchored to anything. Measured across 200 exception notes in month six, 23 drafts contained such a statement, or 11.5 per cent, about one draft in nine. Twenty-one of the 23 were caught by the person doing the nine minutes, and 2 were not, and those 2 reached a customer.

What does each failure look like from outside?

Set side by side on a desk, the two do not look like the same category of problem. Not the same size, not the same shape, not the same kind of evidence.

The drift at Sumeru Bank Limited looked like a line bending. Six weeks passed between the field changing format and monitoring flagging it. About 12,900 files were decided in that window, 430 a working day across 30 working days, and 176 of them moved out of acceptance and into the referral bandThe range of scores the bank routes to a person rather than deciding automatically either way.. The month's approval rate fell from 57.0 per cent to 55.6, and monthly referrals rose from 391 to 508, a rise of 117 that matches the fall in accepts exactly. Every one of those files got a correct-looking decision. Nothing in any individual file was wrong on its face.

The assertion failure looked like one sentence. A paragraph in a letter to one applicant, saying something the file never contained, in the same steady tone as the sentences on either side of it. No aggregate anywhere in the bank moved by so much as a decimal when that sentence went out, and that is precisely what makes it invisible to the machinery built for the other failure.

Would either one's check have found the other one's failure?

Neither, in either direction, and it is worth being precise about why rather than treating it as a slogan.

The population check cannot see an assertion failure because the assertion failure does not disturb any population. One wrong sentence in one letter changes no rate, no distribution, no total. The population check could run on every letter the bank sent for a year, and because it never opens a letter, the sentence would sit there untouched.

The output check cannot see drift because drift is not visible in one output. The desk officer reading a note about a file in the referral band sees a note that correctly describes a file in the referral band. She has no way of knowing that the file would have been accepted six weeks earlier. A firm running both kinds of component therefore needs two controls, and a stricter version of one of them is not a substitute for the other.

Where does one deployed chain put each of them?

Position in the chain matters more than either name. Component 6 sits at the decision itself, and in month six it determined the outcome of 5,981 files: 4,902 accepted automatically, 688 declined automatically, and 391 more that landed in the referral band and went to a person. Component 8 sits after the decision, writing about something already settled, and it determined the outcome of none of the month's 8,600 files.

The component that decides nothing is the one whose actual words a customer reads, and the component that decides 5,981 outcomes produces nothing a customer will ever see. Anybody ranking the two by importance is going to rank them the wrong way round at least once, depending on which unit they happen to be holding at the time. So do not rank them. State each one's reach in its own unit and let the two numbers sit beside each other without a comparison operator between them.

What happens when one policy heading covers both?

Sumeru Bank Limited's first policy on this subject did the natural thing. Because both components had derived their behaviour from data, the policy put them under one heading and required quarterly performance monitoring of every component in scopeCaught by the wording of a policy, so its approvals, records and reviews apply to that component.. One heading, one control, applied evenly. Read on paper it looks careful.

On component 6 that control is exactly the right one, and it is the control that eventually surfaced the drift. Quarterly is the correct cadence for a component whose behaviour moves slowly across a population, and comparing a quarter's outputs against outcomes is precisely the check drift is visible to.

On component 8 the same control produces a quarterly report about a component whose failures happen one note at a time, in a nine minute window, at a desk. Three months of notes is 9,030 outputs. The report sampled 30 of them, one output in 301. At the 11.5 per cent already measured on 200 notes, a quarter of 9,030 drafts carries in the order of 1,038 with an unsupported statement in them, and a sample of 30 would be expected to contain three or four. The 1,038 is arithmetic on a measured rate rather than a count anybody made.

What do the two genuinely have in common?

The two do share one property, and it is worth being fair about. The shared property is a useful one, and it is not the one people usually name.

Neither carries a written procedure a reviewer can read. Sumeru Bank Limited has four components in the same chain that do: the income corroboration rule is 34 lines, and Neelima Rao read all 34 of them end to end in 25 minutes. There is nothing equivalent to open for either component 6 or component 8. Whatever governs their behaviour was derived rather than typed, so a reviewer arriving with an afternoon and a highlighter has nothing to highlight.

The absence of a written procedure has a precise consequence, and the precise wording matters because a loose wording circulates. Repeatability does survive: put the same input in and the same output comes back, for as long as the fitted numbers stay where they are. Repeatability survives in the form that can be tested rather than the form that can be read, and it ends at the next refit; readability does not survive at all. So both components are reviewed by looking at data and behaviour instead of at instructions, and that shared property is what makes it so tempting to write one policy heading over the two. The similarity is genuine, and it has no bearing on anything the checking has to do.

Why is there nothing to slide between the two?

What does this change about the way checking is done?