Neurodiversity: AI is neurotypical by design, are you?

I’ve pair programmed with a really diverse group of people over the years, Claude, Gemini, Kimi, and their ilk are really no different… so why all the fuss about giving them the reins?

Literally billions upon billions of dollars have been and continue to be spent in “aligning” LLM models with humans. Which is to say, to do the least unexpected, least astonishing things when interacting with humans in our language. I suppose another term we might use is to “conform” with conventions and norms humans use when interacting with one another. And in this case, by “norms” I mean “normal” or literally the statistical average, most probable, behaviors as dictated by curated training corpora and subsequent reinforcement learning guided by actual humans.

If you feel like supporting Medium (as I do), you can get access to the following is an article I read this morning, which perfectly explains this “normalcy bias” with a really neat trick to overcome it:

Stanford Just Killed Prompt Engineering With 8 Words (And I Can’t Believe It Worked)

My personal and even professional life, while normal in many respects, does require me to venture frequently into the divergent, the sometimes unwelcome, and the always surprising realm of “neurodiversity”.

I have to keep reminding myself that I’ve only been Pairing with Bots for less than a year, I honestly can’t think of going back to doing the olde-fashioned way anymore. What I do remember about pairing with humans way back when was, the hardest thing about working in a team is aligning with common goals, conventions, practices, architectures and tools. Every developer sees things a little differently. There are some developers who are very idiosyncratic in their thinking and approaches to things. This is a double-edged sword: On the one hand, trying different approaches and being confident in one’s way of thinking is the gateway to innovative solutions and highly motivated and productive programmers.

However (too often), such creative and motivated energy is channeled into very non-standard approaches that don’t mesh well with the work of other developers or industry trends and “norms”. Even worse, when accompanied by a modicum of actual genius, the rest of the team can be left struggling to cope with the flurry of unexplained design decisions and motivations that often flow from such idiosyncratic genius. This can lead to a really unhealthy form of “hero worship” or “cult of personality” where the idiosyncrasies are glorified to the exclusion of the voices and efforts of others on the team. There’s even an anti-pattern named by Jim Collins in his book “Good to Great” about this phenomenon. But there is an antidote:

“Strong opinions, loosely held” — Paul Saffo, Stanford professor

The above quote has guided my interactions (and prompts) as I pair program, whether with humans or bots. It implies a certain humility to accept contradictory information, but I would hold (loosely) to my opinion that:

You and the AI constantly need to prove things to each other

My prompts often include a hedge, something like “I believe (X) is the case, but I might be wrong… can you find evidence to the contrary?”. I’m inviting (well, commanding really) that the model go searching for contradictory information, keeping us both aware of the reality of our circumstance. This is pretty open-ended, and it will burn some “unnecessary” tokens if you do it often, but it helps avoid the commonly understood pitfalls associated with “Vibe Coding”. (Namely, that the AI produces code you don’t understand, and therefore cannot debug.) Setting aside for a moment my oft-quoted cliche reminding us that human-generated code has always suffered the same peril, this is to my mind what distinguishes “Vibe Coding” from Pairing with Bots: The expectation that you told the AI what to build, but were unwilling to invest your own time and effort to discover things with the AI. Your tone for vibe coding is commanding, imperative… “Make me a system to do this…”. This is not Pairing with Bots.

My previous counterexample is eliciting contradictory feedback from the AI, and presupposed that you (as a human) could suspend your pride, ego, confidence… whatever you want to call it to ask an artificial intelligence to contradict you. I used that example first, because it is hard… and it is the most important. I think we as humans want to assert our dominance over this technology (which we pay for the privilege of using) by using that imperative tone. How’s that working out for you?

To be clear, there are times as I Pair with Bots that I do issue simple commands to create things that I’m confident the AI won’t need too much extra elaboration or clarification about… but that should be the exception, not the rule. Real software engineering is tough… I liked to motivate my team to pair program more frequently by repeating:

“Code is scary, don’t go in there alone!”

This pairing thing is a two-way street. Although (as I have discussed ad nauseam in previous articles) the roles you should be playing (you: navigator, ai: driver) will remain fixed, that doesn’t mean you should not be constantly questioning each other. Let’s consider the opposite of the previous example, where instead of questioning your own knowledge and approaches, you’re questioning the AI’s:

With a human, ego and other social factors come into play here… thus “back seat driving” became a thing… but fortunately, this has not (yet) happened to AI. The value of this cannot be understated:

Your AI pair programmer will not take offense, even if you constantly bombard it with questions about their motivations and their actions. This is a HUGE advantage for you.

Do not attempt to constantly question (or command) a human: raising a son with defiant characteristics of autism has made me hyper-aware of the consequences of overloading people with demands. To give you a better idea of what I’m talking about, I’d like to frame this in terms of “imperative” and “declarative” communication: An example of imperative communication would be: “Why are you doing that?” or “Do it this way.” Declarative communication is more observational in nature: “It looks like you’re trying to do something different here.” or “That didn’t seem to accomplish your goal”. Humans much prefer the latter, as I’m sure you can see… with AI, you are free to ask the more blunt question or make the more blunt demand.

In addition to asking the AI to constantly question your strategic and navigational choices, you should absolutely be constantly questioning the tactical coding choices being made by the AI. This presupposes that you are “in the loop” enough with what the AI is doing that you can get enough insight to ask the right questions… and herein lies the fatal flaw for vibe coders, it requires human learning and thinking. Lazy developers need not apply. On the other hand, non-developers with a questioning mind and some time on their hands (and a decent token budget) can ABSOLUTELY both learn and accomplish amazing things Pairing with Bots!

Don’t be afraid to ask

Again, these amazing AI tools have been designed to deal with our human egos and insecurities… they won’t pepper you with demands (unless you ask them to) or be offended if you constantly ask them to justify their behavior or explain the nuances of things. I’ve already recommended that you give the AI permission to question you, the information you’ve provided in the prompt. There’s some more nuance here: Also give it permission to ask you questions if things are unclear. Internally (although the AI may not be “aware” of it enough to always quantify it), the AI has some measure of confidence in aspects of its responses. When that is low, give the AI permission to ask you to clarify things. Recent advancements in Claude Code have produced a little “multiple choice” AI element that is dedicated to this, and the model itself has become much better at detecting this… but that isn’t the case with all tools and models.

But you also need to be asking questions… a lot of questions. I know you have them! Here are a few general categories I use often, and I’ve learned so much by peppering questions like these into my Pairing with Bots interactions:

• Why did you do x instead of y?
• How does this work, exactly (with code snippets and references)?
• Can this be done more simply?
• Do you have automated tests for this logic?
• Is this a common development practice?
• Can we optimize this process to run faster and use less memory?
• What would you change about this code?
• Who wrote this crap? (Just kidding, don’t ask that… or… maybe do?) 😆

To wrap this up, I can’t believe I’m saying this… but in order to more effectively Pair with Bots, my advice is to… er… emulate the ego-less interaction of the bot…? well, kinda… and also help it overcome its resistance to overloading you with questions and doubts… kinda give it permission to do that? To be less, “human” in a way? Man, that sounds bad when I say it like that.

But on the other hand, recognize your humanity: You get tired… you can quit at any time! You get annoyed… take a walk and touch grass! You feel alone and isolated after working with AI all day… include more humans (see last week’s “mobbing” article)… or better yet, just hit the break room or your apartment lobby or anywhere bona fide humans hang out, and remind yourself that it’s also okay to be human again.

🧩 Neurodiversity: AI is neurotypical by design, are you? was originally published in Pairing with Bots on Medium, where people are continuing the conversation by highlighting and responding to this story.