Polished Work Needs Draft Marks

The dangerous thing about a good-looking draft is that it has manners.

It arrives aligned. The headings are balanced. The chart has chosen a color scheme with the confidence of a consultant who owns exactly one excellent carry-on. The spreadsheet has labels, formulas, frozen rows, and enough formatting to suggest that someone was awake and responsible. The deck looks like it has already survived a meeting.

This is pleasant. It is also a little suspicious.

OpenAI's GPT-5.6 release is a clear marker of where the category is moving. The company describes stronger end-to-end knowledge work: presentations, documents, spreadsheets, frontend interfaces, multi-agent workflows, tool use, and computer use that can inspect and refine the rendered result before handing work back. Microsoft describes Copilot in similar workplace language: agents, notebooks, enterprise search, designed content, Work IQ, and handoffs across the apps where organizations already turn uncertainty into files.

The story is not simply that AI can write more text. It can increasingly return the artifact in the shape the organization recognizes as nearly done.

That changes the social life of review. A rough draft asks to be touched. A messy spreadsheet confesses that assumptions are still wet. A bad first slide gives everyone permission to talk about the argument instead of praising the template. Roughness is not always a virtue, but it is a signal. It tells the room that the work is still available to judgment.

Polish sends a different signal. It whispers that the thing has passed through a process. It has enough finish to make objection feel picky. The reviewer opens it between meetings, sees competent structure, and starts looking for typos instead of asking whether the premise is tired, the denominator is wrong, or the recommendation is mostly a vibe in a nice jacket.

Anyone who has reviewed office work knows this trap. A deck with beautiful hierarchy can hide a weak causal claim. A spreadsheet with tidy formulas can bury a bad input. A memo with calm prose can make an unresolved argument feel settled. We already had this problem before AI, of course. Organizations have always mistaken formatting for thought. AI just lowers the cost of the costume.

Current research around office artifacts makes the caution practical, not aesthetic. A recent PowerPoint agent benchmark treats slide work as a rich computer-use problem and finds that strong frontier agents still make partial progress, requiring rubrics that can penalize unnecessary changes and weak aesthetics rather than scoring only success or failure. Another paper on GPT-assisted spreadsheet modeling finds promise in structured model generation, but also inconsistency and reproducibility problems that leave skilled users responsible for verification.

In other words, the artifact can look more finished than the underlying work deserves.

That gap is where the next interface problem lives. We do not only need AI tools that produce better artifacts. We need artifacts that carry better evidence of their own becoming.

Call them draft marks.

Not watermarks in the moral-panic sense. Not scarlet letters for machine assistance, or a compliance theater where every sentence arrives with a tiny badge of technological shame. Draft marks are review cues. They show where the system inferred, compressed, rearranged, calculated, stylized, filled a gap, skipped an alternative, or made a decision that should remain available to human judgment.

A spreadsheet might mark generated formulas, fragile assumptions, external data pulls, and cells whose values changed because the model interpreted a plain-language request. A deck might show which slides were reorganized for narrative flow, which claims came from source material, and which speaker notes contain unverified inference. A memo might keep the awkward uncertainty visible rather than sanding it into executive calm.

This sounds fussy until you picture the ordinary workplace version. A director asks for a board update at 4:12 p.m. Someone hands an agent a folder of reports, meeting notes, last quarter's deck, and a few anxious instructions about tone. The agent returns something shockingly usable. The human has twenty minutes before a child pickup, a commute, or the bleak little sandwich that counts as dinner. What kind of interface helps that person review well?

"Here is a beautiful draft" is not enough.

The useful artifact would say, quietly and specifically: here are the numbers I carried forward, here are the places where the source was stale, here is the chart whose axis deserves suspicion, here is the customer quote I summarized too cleanly, here are three claims that still need a person who knows the business to blink at them.

That is not anti-AI. It is pro-quality. The whole point of delegating the mechanical parts of work should be to leave more attention for the consequential parts. But if the machine spends its new skill making the consequential parts look already settled, we have traded effort for theater.

There is a health angle here too, because review is not just a process step. It is an attentional state. Finish changes posture. It changes how much disagreement feels socially expensive and whether the reviewer feels invited to alter the work or merely obligated to bless it. When everything arrives with the sheen of completion, the human role can quietly shrink from editor to liability shield.

The better design goal is not to make AI output uglier. Nobody needs a renaissance of badly spaced bullet points for the sake of the soul. The goal is to preserve the inspectable middle inside the finished surface. Let the deck look good, but let its claims keep rough edges. Let the spreadsheet calculate cleanly, but leave handles on the assumptions. Let the memo read well, but do not let fluency launder uncertainty.

This will require product choices that may feel less magical in demos. A "show draft marks" toggle is not as glamorous as a perfect first pass. Source trails, assumption panels, change maps, and review modes that distinguish polish from substance all add texture to the object. They slow the gaze in the places where speed would be convenient and expensive.

That is the point.

Work artifacts do not merely communicate decisions. They help organizations decide what is real. As AI gets better at producing the forms of finished work, the humane question is whether those forms still leave room for doubt, correction, and situated knowledge. A polished surface can be a gift. It can also be a velvet blanket over an unfinished argument.

The next serious office AI interface may not be the one that makes every output look ready. It may be the one that helps people see where readiness is still being negotiated.