Why We're Holding AI Back
We're parenting AI with our deepest fears
ChatGPT controls computers only… but is the spectacle worth celebrating?
Adding a dash of humanity sometimes makes us pay more attention to AI. Watching everyone’s excitement in my tech group when ChatGPT agents click through interfaces is a perfect example.
The excitement while watching demos of ChatGPT agents controlling browsers, clicking buttons, filling forms, is palpable.
I too was thrilled when I first experienced it… but then I wondered - why?
What makes this experience any more special than.. say, the ‘deep research’ feature that didn’t get nearly as much excitement?
AI acting human excites us more than AI being efficient… We feel supervisory. We feel in control. Like watching a toddler take their first steps… that childlike excitement mixed with the relief that we can still catch them if they fall.
But the bigger question yet, is - what are we really building the AI for?
The Pizza-Shaped Dog Toy Problem
When I was choosing a dog toy, I picked the pizza slice because it made me smile and looked fun.
“Of course, my dog understands that this is supposed to be a pizza slice…”
But the reality is that the pizza toy isn’t for the dog.
It’s for the human buying it.
Same with children’s books… adults choose colors that infants can’t even see properly.
Most products are designed for the buyer’s comfort, not the user’s needs. And we’re building AI the same way.
To please investors, to comfort users, to satisfy our need to watch and understand.
A study on AI assistants for venture capitalists revealed exactly this: investor-centric design priorities superseded efficiency.
Features like transparency and familiar interfaces were prioritized to secure buy-in, even when more efficient alternatives existed.
Fear drives these decisions. We want AI we can supervise, AI that moves at human speed, AI that shows its work.
We Have Good Reasons Not to Trust (And That’s The Trap)
AI developers have often experimented with highly supervised systems, such as requiring citations or showing work, because in some cases it does prevent catastrophe.
We’re not at a point where AI can be autonomous. Valid concerns abound.
AI inherits bias from training data.
Quality degrades past context windows… even with 128K, 200K, 1M limits. Like a tired worker who needs a day off, AI performance drops without us knowing.
The worst yet - hallucination problem means without supervision, we can’t distinguish fact from fiction.
The higher the stakes, worse the impact of such an unsupervised AI… Think healthcare decisions, military applications, business choices affecting thousands of jobs.
Industry research shows that in 2025, over 78% of companies use AI (and the adoption grows everyday), but trust remains the major barrier for responsible adoption.
Roughly 70 % of employees say they’re excited to work with AI, but over 41% still worry about transparency, ethics, and the accuracy of algorithmic insights.
Building trust requires explainability. AI showing its work matters more than raw performance.
It's like we're still following those grade-school instructions of 'show your work,' treating AI as the perpetual student who must prove every step.
And the truth is that citations do build trust. In fact, it measurably improves AI’s performance.
But what about watching AI click buttons?
Are all precautions equally justified in our effort to keep AI honest and build trust?
And that’s the trap: our justified caution shapes what gets built.
The Real Breakthroughs Hide in Plain Sight
While we celebrate AI clicking buttons like a proud parent filming first steps, the real revolutions happen quietly:
Kimi K2 pushes the boundaries of how AI thinks, not how it looks.
Claude transforms how software gets built… not through flashy demos, but through fundamental changes in development speed and autonomy.
The Model Context Protocol redefined how AI connects with the world… invisibly, efficiently, powerfully.
These breakthroughs share something crucial: they optimize for AI's strengths rather than human comfort.
They're the difference between teaching a bird to walk vs teaching it to soar.
And this fundamental choice is already splitting our technological landscape. I see a bifurcated future: AI-optimized systems...
I see a bifurcated future: AI-optimized systems built for AI agents as the target audience, separate from human-optimized interfaces.
And perhaps, this distinction points toward an inevitable split in how we build technology.
Because invisible efficiency makes us uncomfortable. If we don't see it, so we don't trust it. The behind-the-scenes magic creates anxiety.
The Parent Who Can’t Let Go
We’re helicopter parenting AI with our deepest fears.
The same trust issues that make us hover over our children now shape how we develop technology.
We need to supervise, to control, to understand every step.
Research on the "illusion of control" shows humans create dashboards and overrides to satisfy emotional needs, even when it limits AI autonomy and efficiency.
The problem isn't that we watch AI… It's that we're watching for the wrong reasons. We're designing from fear rather than from understanding.
Good supervision comes from science, not anxiety.
When we build monitoring systems based on evidence of what actually improves AI performance, like citation requirements that measurably reduce hallucinations, we're advancing the technology.
When we add features just to feel in control, like forcing AI to mimic human clicking speeds, we're holding it back.
Building for Tomorrow, Not Yesterday's Fears
We stand at a crossroads.
One path leads to AI being confined by our anxieties, clicking through interfaces at human speed, showing every step like an eternal grade-schooler proving their work.
The other path requires something harder: distinguishing between necessary safeguards and security theater.
The breakthroughs like the Kimi K2, Claude's development capabilities, the Model Context Protocol, all succeed because their creators asked the right question:
What does the science say AI needs?
(NOT: What makes us feel safe?)
I’m not advocating for abandoning oversight here… but rather grounding our design choices in evidence rather than emotion.
The future belongs to those brave enough to let AI be AI… not recklessly, but scientifically.
To build based on what works, not what comforts.
Our children eventually outgrow our hovering…. Perhaps it's time we let our digital offspring do the same.