The Alignment Problem

AI tools have made individual contributors faster than ever. They’ve done almost nothing to help teams think together — and that gap is now the defining challenge in product development.

---

There’s a particular kind of critique session I’ve started dreading. The one where the first fifteen minutes aren’t spent evaluating work, but reconstructing what each person was actually trying to solve. Three designers, one brief, three completely different problem framings — each one reasonable in isolation, each one the product of a conversation with a machine rather than each other.

This is the alignment problem the industry isn’t talking about. Not the philosophical question of whether AI shares human values. The practical one: whether the teams building our products still share an understanding of what they’re building, and why.

Parallel play

The productivity gains are real. Researchers prototype in an afternoon. Designers ship components in hours. PMs vibe-code working demos before the designer has finished the brief. In narrow contexts, the leverage is extraordinary — and unevenly distributed. The people benefiting most are the ones who already had strong foundations: who know what good looks like, who can recognize when a generated output is subtly wrong, who treat these tools like clay rather than answers. For everyone else, the experience is more dangerous: the feeling of competence without the judgment to validate it.

But the deeper problem isn’t individual. It’s structural. What we’re seeing isn’t teams moving faster toward a shared goal — it’s individuals accelerating in parallel, each optimizing their own slice, with the integration work pushed downstream. Jess Holbrook, who leads UX research across Microsoft’s Copilot products, named it well: massive parallel play. Everyone building something. Nobody building together.

What’s been lost is the forcing function. Shared tools created natural coordination — a file someone else had to open and respond to imposed a rhythm of alignment. When everyone works in their own AI-native environment, with their own prompts and models and generated artifacts, there’s no shared surface. The handoff, already the lossiest part of the PRD-to-design-to-code pipeline, now happens between people who have each been collaborating with a machine that doesn’t know what the other machines were told.

Translation has always been lossy. A requirements doc compresses intent; a design interprets that compression; an implementation reads the design. At each step, context falls out. AI doesn’t fix this — it accelerates it. Each tool makes confident, fast decisions about what the input meant and compounds the error, without flagging that it’s doing so. The output looks complete. The understanding is hollow. Speed without comprehension isn’t acceleration. It’s drift.

The fix is upstream

The messy part of design used to happen upfront — in ideation, in the friction of people having to agree on a problem before they could solve it. That mess served a function. It’s where teams built shared mental models, surfaced assumptions, and discovered the right question. Now the mess is distributed across the entire delivery cycle, invisible until it’s expensive.

The answer isn’t to use AI less. It’s to be deliberate about where human alignment has to happen first. Before anyone opens a tool, the team needs a real conversation — not a brief sent over Slack — where they interrogate the problem together and agree on what success looks like. The problem statement should be a collaborative artifact, not a handoff document. Convergence is cheap with AI. Shared understanding is not.

From there, the question is how to reconstruct the shared surface. That might mean designated moments where AI-generated work is brought back into a common space before it moves forward. It might mean critiques that require each person to articulate the problem they were solving before showing the work — because if the team learns more from the framing than the output, that’s a signal, not a failure.

The real bottleneck in AI-augmented teams isn’t output. It’s judgment: the ability to look at a generated artifact and know what’s right, what’s wrong, and what question it doesn’t answer. That capability is what scales everything else, and it’s the one thing these tools don’t provide.

The alignment problem isn’t a technology problem. The tools are doing what they were designed to do — make individuals fast. The problem is organizational: we’ve imported them into team contexts without rethinking how teams need to work around them. The teams that figure this out won’t be the ones that use AI the most. They’ll be the ones that stay most deliberately human about the things AI doesn’t do.

That’s always been the job. AI just made the cost of neglecting it more visible.