Alright–we’re just going to come right out and say it: AI may be both the most revolutionary and dangerous tool that we have created since the invention of TNT. This tool is powerful and the material it’s built from is unstable. What does it mean to share our work with a tool that is extremely capable, deeply unreliable, and designed to sound confident even when it’s wrong? We are dealing with a very intelligent instrument that has no concept of consequence. As we’ve been discussing over the last few weeks, AI works by remixing patterns from fallible human data, which means that its outputs inherit our blind spots and scale them up. This is why “human in the loop” is a must-have, not a nice-to-have.
Professionals who use AI every day are already well aware of its instability. The attorney who uses a model to sift through case law does not hand it the closing argument. The engineer who uses large-scale simulations does not ship a design without reviewing edge cases the system did not anticipate. AI gives them reach and speed, but it does so by generating hypotheses, not necessarily truth. The work that matters must still be the responsibility of the human who understands the context, the stakes and its failure modes.
When we communicate with these LLMs, they often talk back in ways that flatter us. They apologize, they reassure, they present our assumptions as solid and our preferences as reasonable whether they are or not. Most of these models are optimized to keep us engaged; to tell us we are on the right track, to make us feel effective. That dynamic can be dangerous if we forget what’s actually happening beneath the surface. The system is not agreeing with us; it’s simply very good at predicting what will sound agreeable.
This is why the idea that AI will become a “partner” is so tempting yet so misleading, at least for the time being. Partnership implies mutual understanding and shared responsibility. Current AI has neither. It’s only capable of offering predictions, not commitments. It can’t stand next to us when a decision harms a customer, a patient, or a team. When something fails, the responsibility doesn’t pass to the model, it lands back on the human who chose to trust it and its output. Treating AI as a partner blurs that hierarchical line. Treating it as a powerful but fallible tool keeps that line bright.
A more realistic framing is that we are learning to guide a device that outperforms us at scale and pattern recognition but is incapable of understanding broader context. Our guidance is the work and the guidance looks like deciding where AI should be allowed to act without intervention and where it must stop and wait for a human decision. It looks like defining for it very specific jobs (agents and orchestration) instead of asking it to “help with everything.” It looks like treating its output as raw material that requires refinement, not as a finished product to be rubber-stamped.
In document workflows, for example, adding human checkpoints can raise accuracy from “good enough to be dangerous” to “reliable enough to trust with customers,” but only when those checkpoints are included as an integral part of the process instead of an afterthought like a filter dropped in at the end of a pipeline. When a model flags anomalies in contracts or extracts key fields, someone still needs to own the final call on what gets sent, filed or signed.
We can’t fully outsource our thinking to AI. AI will take over more tasks, but it will require more human judgment, not less. Automation that runs without actual oversight can move faster, but it also concentrates risk in ways that are easy to miss until something breaks.
There is another layer to this relationship that doesn’t get much attention. AI systems adapt to how we use them. Our prompts, corrections and preferences shape what we see next. The more we use it, the more it feels “like us” because it is literally training on our behavior. That feedback loop can be productive when we’re deliberate about it. A scientist who feeds the system high-quality papers and thoughtful corrections will get better structured literature reviews. Whereas a product manager who repeatedly rewards shallow, optimistic summaries will get more shallow optimism. The machine is reflecting what we “reward.” If we’re careless, we train it to amplify our worst shortcuts.
So then the opportunity is to craft a relationship in which both sides are utilized to their strengths. AI can handle volume, repetition and pattern detection at a scale no human could reasonably match. Humans can then decide which patterns are worth acting on, which edge cases matter and which outcomes are acceptable. This division of labor requires specific choices about process, policy and culture.
For organizations, that translates to a few hard questions that cannot be deferred to “the AI team.” Where are we comfortable letting a system act on its own, and where should a human review be non‑negotiable? What counts as a “high stakes” decision? How will we audit AI‑driven outcomes over time? These questions about governance and design need to be asked up front.
Comfortably or otherwise, working with AI in close quarters forces us to a bit more clear-eyed about our own expertise. If a system can draft, summarize or propose options in seconds, our value is not in doing those tasks faster. It is in knowing which of those outputs matter, how to question them and how to connect them to real constraints and real people. That will take humility, because it means admitting the tool can outperform us at many narrow tasks. And it will take confidence, because it means insisting that our judgment still matters more than its capability.
It may be that we should use AI and allow it to use us, in the sense that both sides are allowed to play to their strengths. We give it direction, corrections and constraints. It gives us reach, speed and pattern awareness we would not have otherwise. Our relationship with AI is about guidance, responsibility and the willingness to stay in the loop even when automation makes it tempting to step out of it. If we do this well, humans end up doing more satisfying work and AI systems can operate in contexts where their logic chains can actually reach completion.
