The First Move

What we call human oversight is a signature on a draft nobody on the team chose.

The model writes. A person reads, fixes what's off, and approves. Nothing goes out without a human name attached.

I understand why that feels safe. The person has the final say. They can reject every word.

But look at what reaches them. The angle is already picked.

The audience has been assumed. The argument has a shape before anyone starts checking it.

Most of the time, the draft is fine. That's what worries me.

I think the safest-looking AI setup in marketing is the one most likely to cost you your judgment. The report can show more work, delivered faster, with a person checking everything. It cannot show how much thinking that person stopped doing.

A signature records who accepted the result. It tells you little about who decided what the work was for.

What protects your judgment isn't whether a person reviews the work. It's who made the first move.

That first move supplies the context.

Who is this for? What matters to them? What would make the answer useful?

When I answer before opening the tool, I give myself something to test. The draft can improve my frame or expose its faults. Either way, I have a view of my own.

When I let the tool answer first, every edit starts from its choices. I can change a great deal and still circle the same starting point.

I think of that as the orbit. The work moves, but the draft keeps deciding what it moves around.

The difference grows through use. Framing gives judgment practice. Checking someone else's frame may leave that work untouched, even when the checking is careful.

The same tools can support either order. The difference need not appear in the finished work today. It may appear in what the person can do without them later.

The same tools

In June 2005, two amateurs from central New Hampshire won a freestyle chess tournament on Playchess.com.

Steven Cramton was 28, with a USCF rating of 1685. Zackary Stephen was 24, rated 1381. They used three ordinary PCs, with Fritz and Shredder as their main engines.

Their opponents could use computers too. The final put them against grandmaster Vladimir Dobrov's team. Cramton and Stephen won, 2.5 to 1.5.

Their own account describes work that went well beyond accepting a machine's best move.

They chose candidate moves from their experience or agreement among the engines. Then they split the lines between them. When one found a strong continuation, both examined it and compared engines.

They knew how the programs judged different positions. That knowledge helped them decide where to look more closely.

Garry Kasparov described the tournament in his 2010 essay in The New York Review of Books. He wrote, "Weak human + machine + better process was superior to a strong computer alone."

The people were doing more than approving moves. They were directing the search.

But winning a tournament tells us little about what a process leaves in the person over time.

Following the route

London has required the Knowledge since 1865. Transport for London says mastering it usually takes three to four years. Students learn 320 routes, along with the places around them.

In 2000, Eleanor Maguire and her colleagues found larger posterior hippocampi in London cab drivers than in controls. This region helps with spatial memory.

Its size also correlated with years spent driving a cab. That is a correlation, not proof that driving caused the difference.

In 2017, Amir-Homayoun Javadi, Hugo Spiers and colleagues studied 24 volunteers watching filmed routes through Soho.

When volunteers worked out the route, hippocampal activity tracked the possible paths as they entered new streets. That pattern did not appear when they were told which way to go.

The people following instructions still watched the streets.

In 2020, Louisa Dahmani and Véronique Bohbot tested 50 regular drivers. More lifetime GPS use was linked to worse spatial memory without GPS.

They retested 13 people three years later. Greater GPS use between tests was associated with a steeper decline. The follow-up group was small.

None of these researchers watched anyone use AI. They studied different people, with different methods. They did not test marketing judgment.

What interests me is the gap between arriving and learning the way. A similar gap opens when I review a draft without first deciding where it should go.

Does changing the order help when the tool actually is AI?

In 2021, Zana Buçinca, Maja Barbara Malaya and Krzysztof Gajos tested that question at Harvard, with 199 people.

Participants saw meal photos. They had to replace a high-carb ingredient with a lower-carb choice that kept a similar flavor. An AI offered help, sometimes with a deliberately wrong suggestion.

Some participants saw the suggestion right away. Others had to make their own decision first, then could revise it. Another group waited 30 seconds before seeing the answer.

The designs that made people pause or decide first reduced how often they followed wrong suggestions. They did not remove the problem. People still accepted bad advice.

The benefit also varied. People who enjoyed hard thinking gained more from having to do it.

The less comfortable part was what people preferred. They liked the easier designs more and found the forcing designs more complex.

I can see why. Moving first adds work before the answer arrives. Reviewing an answer lets me begin with something that already looks useful.

That ease makes the review-first setup attractive. Preference alone would not persuade me that it protects the person using it.

What practice leaves behind

That concern starts with my own work. When I let a tool generate first and only edit afterward, I notice less practice in the judgment I rely on.

I think of that judgment as a well. Expertise fills it over years. I cannot inspect its level by reading the finished draft.

Destin Sandlin is an engineer who runs the YouTube channel SmarterEveryDay.

His welder friend Barney built a bike with reversed steering. Turn the handlebars left and the wheel goes right.

Knowing the rule did not let Sandlin ride it. Other people who tried could not ride it either.

He practiced for about five minutes a day. After eight months, it clicked. One day he could not ride it; the next day he could.

Later, in Amsterdam, he tried a normal bicycle and could not ride it, though he had ridden one since he was six. It took about 20 minutes for the old skill to return. His son learned the backwards bike in about two weeks.

This was not a study of lost professional judgment. Sandlin had practiced a conflicting skill, rather than simply stopped riding.

I don't want a clean draft to become my evidence that I can still build one.

My campaign reports have already taught me how much a good result can leave unexplained.

I spent a year optimizing campaigns on data I didn't own.

The results looked great until I tried to explain them.

I had built the MarTech stack before the data architecture. The information that could explain the results did not live anywhere I controlled.

Google knew what people searched. Meta knew what people liked. Those signals did not tell me which leads closed or why good customers stayed.

That knowledge lived after the click, in the CRM and sales conversations. It belonged in the context before the next campaign.

Meta's Advantage+ can generate ad variations. Google's Performance Max can move budget across channels. Both can be useful without either supplying everything I need to decide what matters.

AI can repeat an old mistake faster: teams that plan from inside the company answer their own questions, not the customer's.

A fluent draft gives me no reason to assume the customer was understood. My own results gave me no reason to assume I understood them.

I want the first context to include what happened to real customers. Otherwise, I keep improving a report whose missing information never gets asked for.

That missing information also changes how I read an AI pilot.

The pilot works. The demo is clean, the budget arrives, and the team agrees to proceed.

Then fewer people open it each week. The output gets worse a little at a time. Small workarounds become the old way of doing things again.

The first explanation is the tool. Wrong vendor, wrong model, wrong timing.

Those can be real problems. But another loop is harder to spot.

The tool supplies the frame. People check it instead of making their own.

With less practice, their framing skill weakens. The next task then gives them more reason to ask the tool first.

A better model could make that loop easier to live with. More answers would arrive ready to approve.

The cost surfaces when the task needs something the model cannot supply. Until then, each successful draft could make the next delegation seem safer.

A report of completed work would not tell me whether anyone still knew how to begin it.

The postmortem blames the vendor. It was never the vendor.

Where the good results stop

AI does not become unreliable at one clear boundary.

In 2023, Fabrizio Dell'Acqua and colleagues studied 758 BCG consultants using GPT-4. On tasks within its capabilities, AI users completed 12.2% more tasks and worked 25.1% faster. Their results were rated more than 40% higher in quality.

On one task selected to fall outside those capabilities, AI users were 19 percentage points less likely to be right.

That is the jagged frontier: strong performance beside failure, without a simple line between them.

The study showed an immediate risk with skilled people present, without tracking fading skills.

Mata v. Avianca showed another kind of failure in 2023. Lawyers Peter LoDuca and Steven Schwartz submitted material citing six cases ChatGPT had invented.

The court imposed a single $5,000 sanction on the two lawyers and their firm together. Schwartz said, "I just never thought it could be made up."

Their names were on the filing.

That does not prove the tool had worn down their expertise.

My expectation is that review-first teams will become more vulnerable at this frontier over time. Good results will keep rewarding the handoff while independent framing gets less use.

I would be wrong if those teams retained as much unaided skill as teams that framed first, including on unfamiliar work a year later.

I would want that tested, not inferred from a sense of control. I trust a record more than how sure I feel, including when the confidence is mine.

When order matters less

There are good reasons to let the tool go first.

When an answer can be checked independently, order may matter less. Code can compile. Math can be checked.

A result need not rest entirely on whether it looks right to a reviewer.

Compilation does not settle whether the right problem was solved. But a useful check reduces what a person must judge unaided. I would not insist on doing every calculation before using a calculator.

There is also direct evidence against assuming that human-first means better immediate results.

A 2026 trial by Everett and colleagues tested 70 clinicians. They worked through up to six clinical cases within an hour.

AI-first users averaged 85% accuracy. Clinician-first users averaged 82%. Those using conventional resources averaged 75%.

The difference between the two AI orders was not statistically significant.

I cannot use that result to say deciding first always improves accuracy. In that sitting, both orders helped, and neither clearly beat the other.

It leaves my concern about retained skill unanswered. The trial measured one session, not what clinicians could do without help a year later. None of these studies followed that question through.

I haven't watched this break on a marketing team. I don't have a client's before-and-after to offer, and I'm not going to invent one.

Moving first also costs minutes every time. Those minutes add up. Sometimes my frame is worse than the one the model would have supplied.

I pay that cost anyway. Making my own call gives me something worth keeping, even when I change it afterward.

Longer-term evidence could change my mind. Today's output alone cannot answer a question about what I will retain.

Owning the loop

Moving first need not mean staying involved in every step.

In March 2026, Andrej Karpathy released autoresearch. Its setup separates what the agent can change from what it must accept.

The human-written prepare.py sets the data preparation and evaluation. The agent does not change it. The agent rewrites train.py, runs an experiment, checks the result, then keeps or discards the change.

Training runs have a fixed five-minute budget. The metric, val_bpb, measures validation bits per byte. Lower is better.

Karpathy reported 126 experiments in one overnight run. A person did not need to approve each one. The rules for judging them already existed.

I built LoopKit around the same separation. The brand brief is the frozen file. The ad copy can change every round.

I define the score through relevance, clarity, urgency and brand fit. That score still contains my judgment. Writing it down makes its choices visible, including choices I might get wrong.

This kind of loop lets me attempt work I could not carry out by hand. I think of that as the new instrument. Its reach rests on the first decisions remaining mine to test.

Before I open the tool, I ask: Who is this for? What does it have to do for them? What would a good version look like?

A few rough sentences are enough to start. Then I can compare what comes back with something I chose.

Sometimes the draft will be better. I can learn from that difference because I have a first answer to compare.

A cab driver who knows London can still use a phone. Following it does not mean they have lost the city.

I want the same freedom in my work: help I can use without needing it to tell me where I am.

My name belongs on the result. I want my judgment there before the draft.

I send my questions about who moves first by email.