The Algorithm Is Innocent
On strategic patience, algorithmic optimization systems, and the human chain that saved the greatest ad ever made.
Here’s a detail that almost never makes it into the legend.
Before Apple’s “1984” Super Bowl commercial aired—before it became the most celebrated advertisement in history, before it launched the Macintosh, before it demonstrated that a brand could redefine not just a product category but a cultural moment—it was tested. By a market research firm called ASI, using a standard 43-point effectiveness scale calibrated to the business advertising category.
The ad scored a 5.
The norm for business commercials was 29.
The general range for all commercials, across every category, ran from 13 to 42.
In other words, Apple’s sixty seconds of Orwellian dystopia, featuring an unnamed woman in red shorts hurling a sledgehammer at a giant screen, scored the lowest result for a business commercial that ASI had recorded in over a decade and a half of testing.
Fred Goldberg, the Chiat/Day account manager who received the results, described the moment years later on the Business Insider podcast Household Name: “I got the test results. And it turns out this 60-second commercial is the worst business commercial that they’ve ever tested in their system in a decade, decade and a half.”
Then he put the report in his desk drawer and never showed it to Apple.
I’ve actually been thinking about Fred’s desk drawer on and off for the past few months, and especially about what’s likely to happen when there isn’t anyone with the power and the will to make that call. When we cede control not to Orwell’s Big Brother, but to all-seeing and all-sensing AI optimization.
What the data said
None of us were in the room, but we know the general outline of the conversation from the executive’s firsthand accounts of that meeting.
Board of directors. Unanimous rejection. Mike Markkula—Apple co-founder and one of the most respected names in Silicon Valley venture capital—moved to fire the agency on the spot.
Every outside board member concurred. John Sculley, who had watched the ad being made and knew how good it was, got cold feet and instructed Jay Chiat to sell back all 90 seconds of the airtime they’d bought.
What the legend tends to soften is quite how categorical the institutional verdict was.
Philip Schlein—CEO of Macy’s California, no naive civilian—sat with his head on the conference table. Steve Hayden, the ad’s legendary copywriter, gave an oral history to the New York Times shortly before his death in 2025; he remembered the board members with their heads in their hands. Not conflicted. Not asking for minor revisions. Unanimously, unambiguously opposed.
And why not? All the data agreed with them. Focus groups didn’t just dislike the ad—they actually compared it to concentration camps. Multiple viewings made it worse. The ASI score of 5 was a damning statistical verdict. The system had been designed to distinguish effective advertising from ineffective advertising. It did its job.
But here’s what the data couldn’t see.
1984 wasn’t trying to be effective advertising, not in the traditional sense anyway. It was trying to do something the effectiveness scale had no metric for: make people feel that buying a computer was an act of revolution. ASI’s 43-point scale measured how well a commercial performed the functions that commercials were expected to perform—show the product, explain the benefit, and generate purchase intent. 1984 did none of those things. It didn’t show the product. It didn’t explain what the product did. It didn’t have a price point or a call to action. Fail.
But it did something the scale couldn’t measure. And so the scale called it a failure.
This isn’t a story about data being wrong. Data is rarely wrong about what it measures. The problem is always about what it measures—and what it can’t.
The human chain
What I find genuinely moving about the 1984 story—and I’m aware that “genuinely moving” is not the tone one typically brings to Substack articles about advertising history—is the specific chain of human decisions that saved it. The oral history of this ad is unusually rich; the people who made it talked at length after the fact, and their accounts agree on key details that matter. Not a single visionary. Not Jobs alone. A chain.
Goldberg buried the test results because he thought the research methodology was wrong. He called testing “mostly nonsense” and said that he thought the artificial setting produced artificial responses. He was philosophically opposed to letting the data make the call.
Jay Chiat received Sculley’s order to sell the Super Bowl airtime. Instead, he called his media director, Camille Johnson, and gave her a very specific instruction: “Just sell off the thirty.” Johnson sold the 30-second slot. The 60-second slot—the one that mattered—was never put up for resale. When Apple asked, Chiat apparently told his client that it was “too late” to find a buyer.
He basically lied to save his client from discarding their best work.
At the same time, art director Brent Thomas was working the inside track. He begged the Apple executives tasked with selling the airtime—on his hands and knees, by his own account—not to succeed. “I will kill you if you manage to do this.”
When Jobs told Steve Wozniak that the board was blocking the ad, he offered to pay for half with his own money. Four hundred thousand dollars. “I’ll pay half of it if you will.” His reasoning was charmingly naive: “I figured it was a problem with the company justifying the expenditure. I thought an ad that was so great a piece of science fiction should have its chance to be seen.”
Bill Campbell and Floyd Kvamme—the executives who ultimately decided to run it—”threw caution to the wind” on the Friday before the Super Bowl. Not months before. Not with careful deliberation. The Friday before.
The ad aired on January 22, 1984, before 96 million viewers. All three television networks ran news segments replaying it, generating an estimated five million dollars in free publicity. Macintosh sold 72,000 units in its first hundred days, exceeding Jobs’s own target. One week after the launch, the Apple board convened.
The same men who’d sat with their heads on the conference table, who’d moved to fire the agency, who’d not produced a single supportive vote—they gave their Macintosh team a standing ovation.
The question I keep circling
I’ve been a part of five major technology transitions in my brand strategy career: mobile, social, programmatic, martech, and now AI. Each transition was different but followed pretty much the same pattern.
New technology arrives; brands optimize existing processes with it, a small number use it to do something categorically different. The optimizers get incrementally better. The categorical thinkers sometimes change everything.
What nags at me about the current transition isn’t that AI will automate creativity. That argument is mostly wrong, and it’s mostly a distraction. What bugs me is something that’s a bit subtler: AI-assisted optimization systems are making it structurally harder to protect the decisions that look wrong in the short term.
Think about what 1984 would face today in an industry already infested with algorithms and AI agents.
The pretest score survived the revolution, but you can’t hide it in your desk drawer anymore. It’s become a dashboard metric, visible to twelve people simultaneously, with a trend line.
Focus groups have survived also, but their responses are sentiment-analyzed and summarized automatically. The competitive benchmark comparison—your ad vs. category norm—is a single number updated in real time. The board doesn’t have to ask, “what did research say?” because research is already on the screen when they walk in.
This isn’t an argument against better data. More information isn’t the problem.
The problem is what happens to organizational decision-making when the data is both more comprehensive and more instantly visible. The space between “the research says this” and “therefore we should do this”—the space where human judgment used to live—gets compressed. Sometimes it disappears entirely.
The brands that are struggling most with AI-assisted optimization aren’t the ones making bad creative decisions. They’re the ones making fast, safe, well-validated decisions that collectively produce a portfolio of work that never surprises anyone.
What “too early to tell” actually means
The Greeks had two words for time that we collapsed into one. Chronos was sequential time—minutes, quarters, fiscal years. Kairos was the opportune moment, the right time—not measurable by clock but recognizable by its character.
Marketing measurement systems are built for chronos. We’ve always known this, but we’ve mostly been able to live with it.
A brave campaign looked bad in week one, better in week six, and vindicated by year two. The measurement window was long enough—or at least human enough—to allow for that arc.
What’s happening now is a compression of chronos in a way that makes kairos increasingly invisible. When campaign performance data updates daily or hourly, the pressure to respond to early signals becomes almost irresistible. Not because anyone is being cowardly. But because the system generates so much information so quickly, the definition of “too early to tell” keeps shrinking.
Nike’s “Dream Crazy” campaign featuring Colin Kaepernick lost two billion dollars in market cap in seventy-two hours. The week-one data was a disaster by every conventional metric: stock price down, call-for-boycotts trending, and immediate negative earned media. The year-one outcome was a nineteen percent revenue increase and a campaign that won an Emmy. In a world of real-time performance dashboards, I genuinely don’t know if that campaign survives its first week.
This temporal problem isn’t new. But it’s getting sharper. AI optimization systems trained on past performance data to find out what works are making it sharper. These systems are also making it harder to protect the decision that seems wrong before it seems right.
Here’s why: every AI model that learns from campaign performance is learning, necessarily, from the past. It is looking for patterns in what has worked in the past. There isn’t a good historical pattern for categorical novelty, which is what 1984 was.
The model can’t find anything that fits with “doesn’t show the product, references Orwell, might start a revolution in personal computers.” So it automatically chooses the closest proxies, which can give a verdict of 5 out of 43 faster and with more statistical authority than ASI could in 1983.
I’m not arguing that the AI is wrong. I’m arguing that the question it’s answering is different from the question that matters for certain categories of decision.
The organizational question too few are asking
I want to be careful here about what I’m not saying. I’m not saying optimization is bad. I’m not saying data should be ignored. I’m not saying brave creative decisions always turn out to be right—plenty of work that looked like 1984 in the edit suite turned out to be self-indulgent and ineffective, and the critics were correct.
What I am saying is that brand owners need to be honest about which category of decision they’re making at any given moment—and whether their current systems are capable of protecting the decisions in the second category.
Most AI-assisted marketing systems have been engineered for the first category: decisions where historical performance data is a reliable guide to future performance. That covers the majority of marketing decisions, probably ninety percent of them. And since there’s a massive amount of waste and inefficiency in marketing, these systems are genuinely valuable.
The second category is different. It covers decisions where the value is precisely that they have no historical precedent—where the thing you’re trying to do is categorically new. In these cases, optimization systems don’t just work poorly. They actively surface the wrong answer. Not because they were poorly made, but because they are doing exactly what they were made to do, and that doesn’t apply here.
The organizational question too few boards are asking is, “What decisions do we ring-fence from the optimization loop?”
Not because data doesn’t matter for those decisions. I chose the title of this article very intentionally:
The algorithm is innocent.
It doesn’t make the cowardly call. It simply gives leaders the cover they were looking for. Every “the data says we should course-correct” conversation I’ve watched in twenty-five years has had the same subtext—someone in that room already wanted to course-correct, and the dashboard gave them permission to say so out loud. The algorithm isn’t the problem. The willingness to outsource conviction to it is.
The chain that saved 1984 required Fred Goldberg’s drawer, Jay Chiat’s lie, Brent Thomas’ begging, Steve Wozniak’s checkbook, and Bill Campbell’s Friday afternoon nerve. Five separate acts of human conviction, operating in sequence, to protect a single sixty-second commercial from the unanimous verdict of institutional rationality.
That chain was fragile. It almost didn’t hold. If Goldberg had been more conscientious about sharing research. If Sculley had followed through more decisively. If Wozniak’s offer had been necessary and Jobs had said no.
The question I’d put to any brand leaders who are looking for an AI-assisted optimization system right now is not “How can we make better decisions?” It’s “What have we done to protect the decisions that look like the worst ideas in the room before they turn into the best things we’ve ever made?”
Your answer to that question is your actual AI integration strategy, whether you know it or not.
Great read Adrian, your argument is on point! It helped shape what became this: https://www.protein.xyz/too-viral/ so thank you