The Other Alignment Problem
Anthropic is asking how to align AI with human goals. There’s a question underneath that one — how do we stay aligned with our own?
A few days ago, Computerworld ran a piece on a new Anthropic blog post titled “When AI builds itself.” Anthropic’s researchers — Marina Favaro and Jack Clark — laid out three possible futures for AI development. The third is the one keeping them up at night: AI systems capable of recursively improving themselves faster than humans can supervise.
If you’re in technology leadership, the timing is significant. We’ve spent the last three years moving from “AI as productivity tool” to “AI as autonomous agent.” Gartner now predicts that by 2028, 15% of day-to-day work decisions will be made autonomously by agentic AI systems. One-third of enterprise software will have agentic capabilities embedded. These are not hypothetical numbers anymore.
But the line from the Computerworld piece that stayed with me wasn’t from Anthropic. It came from Gartner’s senior principal analyst, Ashish Banerjee:
“Human-in-the-loop is not a strategy if the human cannot keep up with the loop.”
Read that twice.
For the last two years, every governance framework I’ve reviewed — every enterprise AI rollout I’ve consulted on — has leaned on “human-in-the-loop” as the safety mechanism. The human approves. The human reviews. The human signs off. We’ve treated that human as a fixed constant in the equation, a steady reviewer of decisions flowing past them.
But the loop is accelerating. And the human is not.
The alignment problem we’re missing
Most of the conversation around AI alignment focuses on technical alignment: how do we make sure AI systems pursue the goals we want them to pursue? It’s a deep, important question, and the research community is right to take it seriously.
But there’s another alignment problem sitting underneath it that gets less airtime: human alignment. How do we make sure we are still pursuing the goals we said we wanted to pursue, as the world around us speeds up?
That sounds soft. It is not.
Consider what it actually means for a CTO or an enterprise architect right now. The volume of decisions coming at you per hour has tripled in the AI era. AI agents propose code changes. AI systems triage incidents. AI tools surface candidates, draft contracts, recommend pricing, summarize meetings, suggest hires. Each of those is a decision point that previously took human time. Now they take human approval.
If the human-in-the-loop is going to stay in the loop, two things have to be true:
- The human has to have the capacity to keep up with the loop’s pace.
- The human has to have the clarity to evaluate what they’re approving against actual goals — not just rubber-stamp whatever flows past.
The first problem is mechanical. We can solve it with bounded autonomy, embedded guardrails, decision boundaries, and architectural supervision (as Forrester’s Charlie Dai correctly pointed out in the Computerworld piece). That’s solvable engineering.
The second problem is harder. It requires that the human, in the moment of approval, knows what they actually want. Not what the system is asking. What they want.
That is the alignment problem no one is talking about.
Why this is a cloud leader’s problem first
I’ve spent twenty years inside cloud and digital transformation. The pattern I’ve watched, project after project, is this: the technology gets implemented. The processes get adjusted. The org chart gets redrawn. And six months in, the leadership team realizes nobody actually decided what they wanted the transformation to accomplish in the first place. They decided to do the transformation — but not what direction it was supposed to point them in.
AI is going to do that pattern at 10x speed.
When you stand up an AI-augmented operations team, the question isn’t “is the AI doing the right thing?” The question is “what is the right thing — and have we made that clear enough to evaluate the AI’s behavior against it?” If you haven’t done that work, your governance framework is a rubber stamp. Your human-in-the-loop is a fiction.
The cloud leaders I’ve watched succeed in the AI era are the ones who’ve stopped reacting and started anchoring. They’ve made the decision early about what their organization is trying to accomplish, what kinds of decisions deserve human attention, and what kinds of drift they’re refusing to tolerate. They’ve defined the goal. Then they’ve built the governance around it.
The ones who are struggling are the ones who treated AI adoption as inevitable and figured the alignment would work itself out as they went.
It doesn’t.
Four questions to ask this week
If you’re an enterprise leader looking at your AI agent strategy right now, these are the four questions worth pausing on:
North — what actually matters? Have we written down, with specificity, what success looks like for this AI deployment? Not the technical objective — the business and human one.
East — what are we letting in? Which AI-generated outputs are flowing into our decision-making process unfiltered? Which inputs are quietly shaping our team’s thinking before they have a chance to evaluate them?
South — how is the human actually spending their time in this loop? Are they reviewing thoughtfully, or rubber-stamping because the queue is too long?
West — what are we refusing to let AI do, on principle? Where is the line we won’t cross even if the efficiency case is strong?
Those four questions are the spine of a book I just published called The Digital Compass: Finding Direction When Everything Moves Too Fast. It’s a personal framework, but it transfers cleanly to organizational governance — because every governance failure I’ve seen at the enterprise level traces back to a leadership team that hadn’t done the same work at the personal level first.
The deeper question
Anthropic’s researchers ended their blog post by acknowledging that even slowing AI development carries risks — if a slowdown only lets the least cautious actors catch up, everyone is less safe.
I think they’re right. The answer isn’t to stop building.
The answer is to get clearer, faster, about what we’re building toward — and who we want to be on the other side of it.
Speed without direction is just noise.
The leaders, organizations, and communities that thrive in the AI era will be the ones that use technology to amplify human potential without losing sight of what makes us human in the first place. Trust. Empathy. Judgment. Presence. Connection.
Human-in-the-loop is not a strategy if the human cannot keep up with the loop. But it can become one — if the human does the work of knowing what they’re actually in the loop for.
The Digital Compass is available now on Kindle. Paperback ships June 29. Order on Amazon →
Read the original Computerworld article: Anthropic suggests slowing AI research until we can align it with human goals
Read Anthropic’s “When AI builds itself” blog post on the Anthropic Institute site.