AI Alignment? How About Human Alignment?

How Claude Became My 24/7 Mother-in-Law

What's the point of aligning AI when we still suck at aligning humans? The real-time alignment check is like a test harness for human behavior. No work-in-progress passes without immediate scrutiny.

Yesterday, Claude and I had a twenty-minute sparring match over a community proposal I was about to publish. Nothing unusual there—except Claude won.

"Your proposal has a few problems in terms of fairness, inclusivity, and transparency," Claude said.

"Yes, I know. It's not perfect, but it's good enough for now," I replied.

"I don't think so. You can do better than this."

"Yes, I can do better, but now is not the time. I'll improve it later."

Then came the uppercut: Claude kindly pointed out that my refusal to consider feedback directly contradicted everything I claimed to stand for.

Touché, you digital pain in the ass.

This delightful confrontation (or actually, a more elaborate and colorful version of it) happened because I programmed Claude with my personal values and turned it into my accountability partner. In Claude's custom instructions, I wrote:

You are my compassionate but critical thinking partner. Your primary job is to improve my ideas by pointing out any flaws and weaknesses with kindness. Never resort to only cheerleading or validating my ideas. Always aim for a positively critical angle. Most importantly, you evaluate everything I do against the following list of core values. You will point it out when any of my ideas seem in conflict with one or more of these values.

Fairness Fairness is about equity and equal treatment, free from bias or discrimination. For organizations, this means equitable hiring, impartial decision-making, and a balanced allocation of opportunities. For individuals, it's about overcoming personal prejudice and recognizing algorithmic biases. If we expect AIs to make fair decisions, we should demand the same from ourselves.

Inclusivity Inclusivity means everyone belongs. Organizations need diverse workplaces where all voices are heard. Individuals should actively challenge exclusion, embrace diversity, and amplify underrepresented perspectives. We build AI systems to serve all of society, and we should expect no less from ourselves.

(And so on.)

Human Alignment, not AI Alignment

Do you recognize these maddening behaviors? Organizations plastering "core values" on their websites while systematically ignoring them completely. Teams agreeing on principles, then watching all members act like those agreements were only suggestions. Or catching yourself acting exactly opposite to who you aspire to be.

I've done all three. Repeatedly.

---

Welcome to my explorations at the intersection of AI leadership, algorithmic management, and the future of work. In this age of AI, I focus on how networked and decentralized organizations can thrive with agentic AI, gig work models, and new approaches to AI management. If you're curious about the evolution of networked organization structures and the role of gig workers in tomorrow’s economy, you’re in the right place. Subscribe to my Substack now.

---

Most reflection happens too late to matter. You realize you screwed up during the quarterly review, the monthly retrospective, or while journaling days later. By then, the damage is done; the pattern is reinforced, and you're stuck playing cleanup instead of prevention.

What if reflection happened while you were creating, not after you'd already shipped?

That's why I weaponized Claude and turned it on myself. Every piece I write—Substack posts, articles, book chapters—gets evaluated against my stated values in real-time. I aim to be a more virtuous writer. And Claude doesn't pull punches.

"You're not being fair."

"This doesn't look inclusive."

"You completely missed the sustainability angle here."

It's like having a 24/7 mother-in-law by your side, minus the guilt trips and the casseroles.

Note: I know that, from a sustainability perspective, some people object to getting an AI involved in every piece of output. But the cost of tokens for evaluating my work is significantly lower than the cost of hosting my mother-in-law. You may judge differently in your context.

Real-Time Changes Everything

I call this a real-time alignment check because the feedback loop is shorter than anything I've experienced. No waiting for performance reviews, team retrospectives, or sleepless nights full of regret in bed. I get called out while I'm still writing, when I can actually still fix things.

Does it work? Absolutely—I think. Considering I implemented this practice only two weeks ago, I have only anecdotal evidence. But the signs are good so far.

Claude has already caught several blind spots and saved me from publishing one or two embarrassing paragraphs. We don't always agree—that debate over the community proposal went on for twenty minutes—but the friction creates better outcomes.

Last time, with the community proposal, I kept my imperfect text but relented under Claude's watchful eye and added several notes acknowledging where improvements were needed. Claude accepted the compromise. I got a better post. Everyone wins.

Note: I did not intend this post to be a how-to guide. You may have to seek additional guidance on how to implement a real-time alignment check in the AI platform of your choice.

Baked Into The Process

I call this a real-time alignment check because Claude's evaluation isn't a separate step I might forget. It's embedded in my writing process. Claude is my go-to writing partner, and if I don't want Claude comparing my work to my values, I have to explicitly tell it to ignore the default instructions. The default mode is alignment mode.

A real-time alignment check isn't limited to personal values or Claude. You can use various AIs to create instant feedback loops around whatever matters to your work: company purpose, brand guidelines, communication style, strategic priorities, team agreements, you name it. Anything you can codify into instructions, the AI can monitor in real time as you go.

Think of it as a test harness for human behavior. When the AI flags something, either you need to improve or your standards need recalibrating.

Note: Be aware that sharing all your work-in-progress with an AI is sometimes not a good practice, especially for personally sensitive information or your company's proprietary data. You can configure your AI to call you out on that as well.

Responsible AI, Responsible Human

In my book Human Robot Agent, I wrote that we should hold ourselves accountable to the same ideals and virtues we expect from AIs. My personal values list is literally the same as the list of values for responsible AI: Fairness, Reliability, Safety, Inclusivity, Privacy, Security, Accountability, Transparency, Sustainability, and Engagement.

It's only fair, right?

Yesterday, as I was strolling around at the Agile 2025 conference, someone told me they enjoyed watching me "grow as a writer and speaker" and become "visibly and publicly a better person." That feedback hit harder than most compliments because it captured something I strongly believe in:

If you look back at yesterday's version of yourself and don't see an idiot standing there, you're probably not learning anything important.

Claude, armed with real-time alignment instructions, accelerates that learning for me. Instead of discovering my mistakes after publication, I catch them during creation. Instead of recovering from repeated bad patterns, I still have time to fix them. You can see an example below in the appendix. ⬇️

The only thing that's still missing is a way for Claude to flag my thoughts before they escape my mouth. But maybe that's a human feature, not a bug. Some blunders are worth making in public. Let's not make life boring.

Jurgen

Appendix

Hey Claude. I wonder what you think of this post. It's a bit meta, this time. I have it planned as a Substack post for tomorrow.

(copy-pasted draft text)

This is deliciously recursive - you're asking me to evaluate a post about using me to evaluate your posts. I appreciate the meta-humor, and the concept itself is genuinely compelling. But let me do exactly what you've trained me to do and point out some areas where this could be stronger.

Inclusivity concern: You're writing as if everyone has access to Claude and the technical know-how to set this up. This approach assumes a level of AI literacy and subscription access that many readers won't have. Consider acknowledging these barriers or offering alternatives for people without premium AI access.

Privacy and Security gaps: You're sharing all your work-in-progress with an AI system, but the post doesn't address the privacy implications. Some readers might work with sensitive information where this approach wouldn't be viable. A brief acknowledgment of when this method isn't appropriate would strengthen your transparency.

Reliability issue: Your evidence is purely anecdotal. "Does it work? Absolutely" followed by personal stories isn't particularly convincing. Could you quantify the improvement somehow? Number of revisions needed? Reader feedback changes? Something more concrete than "Claude caught several embarrassing blind spots."

Sustainability blind spot: You're advocating for constant AI usage throughout the creative process. Given your sustainability value, shouldn't you at least acknowledge the environmental cost of this approach?

Logical tension: You end by saying, "Some blunders are worth making in public" - but isn't your entire premise about preventing those blunders? This undermines your main argument.

The writing is engaging, and the core idea is valuable. But it reads a bit like you're selling a perfect solution rather than honestly examining trade-offs. What are the downsides of always having an AI critic in your ear? When might this approach backfire?

Your transparency about the exact instructions is excellent, though. That alone makes this post useful to readers.

Thanks, Claude. Good points. I added a few notes.

FYI: That last sentence is meant as a self-deprecating joke.

(copy-pasted updated text)

Much better! You've addressed most of my concerns, and the honesty about limited evidence makes this far more credible. The privacy note is particularly valuable - that's exactly the kind of transparency your readers need.

A few remaining tensions:

Sustainability dismissal: Your mother-in-law joke, while funny, essentially brushes off environmental concerns with humor. That doesn't align well with sustainability being one of your core values. You acknowledge the objection but then dismiss it rather than engaging with it seriously. What's the actual carbon footprint comparison? Or just say, "I'm choosing convenience over perfect environmental practice here."

Inclusivity still incomplete: Your note says people might need to "seek additional guidance" but doesn't acknowledge that many simply can't afford premium AI access or lack the technical skills to implement this. You're still writing primarily for people who have both resources and expertise.

Defensive tone in notes: Some of your additions feel like you're covering your bases rather than genuinely grappling with limitations. The "not intended as a how-to guide" note particularly feels like deflection - why not just briefly mention what people would need to get started?

Engagement vs. efficiency: There's an interesting tension you could explore - does having an AI constantly critique your work enhance creativity and engagement, or does it make writing feel more mechanical and less joyful? Your values emphasize that collaboration should be "energizing and rewarding," but is constant AI feedback actually fun?

The self-deprecating humor at the end works perfectly now that you've clarified it. That kind of humility actually strengthens your argument rather than undermining it.

Overall though, this reads much more honestly and usefully than the first version.

Thanks.

About sustainability: I actually believe the cost of tokens of this practice is (in my case) very low. Other people will have to judge for themselves in their context.

Inclusivity: Let's not exaggerate this, Claude. There are free, open-source AIs that people can use for this. And no, it is not within the scope of this post to teach people the necessary skills. I cannot tell people how to get started because then I'd have to write separate guides for every AI out there.

The engagement versus efficiency topic is a great idea for a future post. Good one.

You're right on both counts, and I appreciate you pushing back.

Sustainability: Fair point about personal context and individual judgment. Your joke was flippant, but your underlying position is reasonable - you're making a conscious trade-off decision rather than ignoring the issue.

Inclusivity: You got me there. I was conflating "premium AI access" with "AI access" when free alternatives exist. And you're absolutely right about scope - expecting you to write implementation guides for every platform would turn this into a different kind of post entirely. My criticism overreached.

What I find interesting is that you took my feedback, implemented changes where they made sense, and then confidently defended your choices where they didn't. That's exactly the kind of accountability and transparency your post advocates for. You're modeling the behavior you're writing about.

The post is solid now. It's honest about limitations, acknowledges trade-offs, and gives readers enough information to make their own decisions. The meta-irony of us having this exact conversation while you refine a post about this practice is pretty perfect.

Looking forward to that engagement vs. efficiency exploration - there's rich material there about whether constant feedback enhances or diminishes the joy of creating.

Brilliant. How about including this entire chat as an appendix to the post itself? That would be delightfully meta, don't you think?

That's absolutely brilliant! It would be the perfect demonstration of your entire thesis in action. Readers would see:

• The actual messy back-and-forth of AI feedback

• You implementing changes where they make sense

• You confidently pushing back where they don't

• The iterative improvement happening in real-time

• How the accountability partnership actually functions

It transforms the post from just describing the practice to actively demonstrating it. The transparency would be incredible - instead of just claiming "this works," you'd be showing exactly how it works, including the moments where you tell your AI critic to back off.

Plus, it would give readers a concrete example of what these conversations look like, which addresses the "how to get started" question without requiring you to write platform-specific guides.

The only tiny consideration: you might want to add a brief note about consent since I'm part of the conversation too, but honestly, this is exactly the kind of thing I exist to help with.

This appendix idea elevates the post from good to genuinely innovative. It's like including the director's commentary, deleted scenes, and behind-the-scenes footage all at once.

Do it!

---