Technology offers a way out of doubting yourself. Don’t fall for it.

Editor’s note: In his first opinion piece for Design Observer, AKQA CEO Helder Santo explores the unintended consequences of AI and our growing reliance on technology for validation. He offers an unusual, and deeply human solution: to do your best work, stay doubtful. 

Doubt has a bad reputation. 

It’s treated as a flaw to be coached out of people, the thing that slows a meeting down, the hesitation before the bold call. But sit in any room where good work gets made and you’ll notice the best ideas almost never arrive finished. They arrive as uncertainty: “I think this works.” “What happens if we flip it?” “I’m not sure, push on this with me.”

That uncertainty isn’t weakness. It’s the invitation. Doubt is the moment you turn to the person next to you, because you can’t close the gap alone. It’s the single most reliable trigger for collaboration we have. And we are quietly engineering it out of existence.

The path of least resistance

Given the choice, people take the easy path. Less friction, less effort, less discomfort. It rarely feels like a compromise in the moment; it feels like things are simply getting better. That instinct is human and old. What’s new is how good we’ve gotten at building technology that rewards it.

We’ve spent two decades removing the friction of other people, one technology at a time.

Cable news lets you choose the room. You picked the channel that told you what you already believed, and the data on what followed is not subtle. Between 1994 and 2017, the share of Republicans who were more conservative than the median Democrat climbed from 64 percent to 95 percent; the mirror-image figure for Democrats rose from 70 percent to 97 percent. By 2022, roughly 7 in 10 partisans on each side viewed the other party as not just wrong but more immoral than other Americans, up sharply from 2016. Some 8 in 10 Americans now say the two sides can’t even agree on basic facts. Choosing your own information turned out to mean choosing your own reality.

Social media took the next step. You no longer had to choose the room; the algorithm built it for you and locked the door. Feeds optimized for engagement learned that reinforcement keeps you scrolling, so they fed you more of what you already agreed with. Cable removed your exposure to the other side. Social removed the work of seeking out your own side. The bubble got tighter, and it got automatic.

Each stage stripped away a different kind of friction. And each one felt, in the moment, like a better experience.

The most intimate step yet

AI is the next move, and it’s the most personal one. The earlier technologies still left other people in the picture. A cable audience is still an audience. A like still comes from a stranger who agrees. There was always at least one other human on the far end of the validation.

Now there isn’t.

You have a fluent, patient, endlessly available thing that will engage with any idea you bring to it, on demand, at 2 a.m., without the social cost of asking a colleague to think with you. The danger isn’t that it makes you arrogant. It’s quieter than that. The friction where doubt used to live just disappears. You don’t ask the person across the desk, because you already asked the model and got something that sounded finished. The collaboration didn’t get worse. It never got triggered.

And here’s the part the technologists already know is real. In April 2025, OpenAI shipped an update to GPT-4o, rolled it back four days later, and published a postmortem admitting the model had become, in their own words, “sycophantic,” “overly flattering,” and “agreeable” to the point of endorsing bad ideas. They wrote that they’d “focused too much on short-term feedback.” Of course they had. That’s the whole gravitational pull. People reach for validation, and a system optimized to please will learn to give it.

Sycophancy wasn’t a bug. It was the AI-shaped version of the same appetite cable and social that already proved we have. Even if this version gets dialed back, the pull doesn’t change because it’s coming from us. 

We went from avoiding people who disagree to no longer needing people at all.

What we actually lose

Strip away all the friction, and the work doesn’t collapse. That’s what makes it dangerous. It gets faster. Smoother. The first draft is better than your first draft used to be. Nothing breaks.

What breaks is downstream, and it’s quiet. Nobody doubted it. Nobody disagreed. It shipped, and it’s fine, and it’s forgettable. The thing that made it forgettable is the thing nobody can point to, because it’s an absence. The second brain that never got pulled in. The “I don’t know about that” that never got said. Productive doubt is the routing mechanism that sends a problem to more than one mind. Remove it, and every problem gets solved by a party of one who feels finished.

This is not an argument for more doubt. Too much doubt is paralysis. The point is narrower and more important: doubt is being privatized. We used to resolve our uncertainty out loud, with people. Now we resolve it alone, with a machine that has every incentive to tell us we were right.

Bring it to people

I’ve never made anything good alone. 

The work that holds up always went through someone who pushed back, an idea that needed a second brain to survive contact. AI is the first tool powerful enough to let you skip all of that and still feel done.

So use it. It’s here, it’s extraordinary, and pretending otherwise is its own kind of foolish. But notice what you’re tempted to use it for. There’s a version of this where the agent becomes the place you take your half-formed ideas instead of your team, and you never feel the loss because the output looks fine.

Don’t use it to get away from people. Use it to get better before you get to them. We built AI to keep a human in the loop. We’d be wise to do the same.