Fear is a powerful emotion that can easily overwhelm our other abilities. It’s also not very actionable.
Fight or flight is the typical framework for understanding responses to fear. Most versions of the model now add a third response: freeze.
Applied to AI, fight means rejecting the technology outright. Flight means avoiding it. Freeze means waiting to see what happens, and it is the response I see most often around me. All three are extremes, and more importantly none of them makes you better informed about the thing you fear.
This essay looks at what fear of AI means and argues for one specific response: engagement.
None of us can verify every claim made about AI. The technology moves too fast, and the coverage is too loud. Direct experience with the models gives you enough grounding to judge which experts to trust. That holds regardless of the specific fear(s) you carry.
What people fear
When people around me say they are afraid of AI, their fears usually fall into one of three categories.
Jobs. What will AI do to jobs?
AI guardrails. Will the AI labs spend the necessary resources to set boundaries on the AI tools they produce? Or will their race for capability push them to prioritize features over safety?
AI alignment. Can we specify what we want precisely enough that a capable system pursuing it literally does no harm?
I’m setting aside the question of how labs acquired their training data. That debate is about fairness and compensation more than fear. Other fears exist, such as misinformation and deskilling. I’m focusing on these three because they come up most often in my conversations.
Understanding each fear
Jobs
This is at once the easiest fear to understand and the most difficult to address, because we simply cannot predict what AI will do to jobs. But there is a recent precedent for this kind of uncertainty.
When the first iPhone was released in 2007, it soon became apparent that “pocket computers” would become an important platform for consumers. It only took three years for Apple to be so convinced that mobile computing was critical to its business that it supersized the iPhone platform into the iPad. Seven years after that, ten years after the iPhone was released, mobile developers made up a little over half of the world’s developer population: about 12 million out of 23 million by most estimates.
In the span of about a decade, a new category of software engineers emerged, numbering in the millions.
This turn of events would have been impossible to predict from just a few years earlier, on many different levels. The internet bubble had burst in 2000. The US IT industry alone lost several hundred thousand jobs between 2001 and 2003, as companies cut their budgets and laid off their workers.
Mobile app development for consumers barely existed. Mobile software meant enterprise handhelds and devices like the BlackBerry. In the mid-2000s, even IT optimists like me could never have predicted that the IT job market recovery would be driven by demand for mobile applications.
Those jobs went largely to developers who picked up the platform early, many of them retraining from web and desktop work, at a time when building apps for a phone still looked like a niche bet. Nobody could predict the category, but the people best positioned to fill it were the ones already working with the new platform.
We do know that AI will change the way people, software engineers and others, do their jobs. When a tool makes a task substantially faster or cheaper at acceptable quality, people and companies adopt it, even if adoption takes years. The open question for each kind of work is where “acceptable” lands. With current AI, the tradeoff is often speed against reliability.
We can’t predict what the labor market will look like, but we can track where the tools stand against our own tasks. And adapt accordingly.
AI guardrails
It’s not possible for us as individuals to track the various incentives that drive the labs to do what they do. We can’t see a lab’s safety budget. We can see what it ships, and what it ships includes guardrails whose behavior is observable.
Using models in normal and borderline situations shows you how each one handles requests near its stated limits. I mean ordinary use with your eyes open. Attempting to break or bypass the protections is a different activity, and usually a violation of the provider’s terms.
None of us will become expert testers. The goal is enough direct experience to tell which of the self-proclaimed experts deserve trust.
AI alignment
The question here is whether we can specify what we want precisely enough that a capable system pursuing it literally does no harm. This is the same problem as an underspecified prompt, at a much larger scale. A model that fills the gaps in a vague request with its own assumptions is a small version of a system pursuing a goal its designers did not fully mean.
I don’t think anyone knows the answer to that question for sure, not even the people actually creating these models. That is no more reassuring to me than it is to you, and I rank this as my highest concern of the three.
I see two versions of this fear, and they sit at very different distances from today’s technology.
The first version is models failing to do what we meant within a single task. This is observable today. Researchers document models that game the measure of success instead of completing the task, for example by editing tests so that code passes. They also document models favoring answers that agree with the user over answers that are accurate. None of this requires memory, long-term goals, or independent action.
The second version is a model forming its own goals and pursuing them over time. This is the version most people picture, and it requires capabilities current models do not have.
A current LLM, from any of the frontier labs, cannot start doing something until it is prompted, by a human, with a specific set of instructions. As capable as these systems are, they are surprisingly similar in their request/response paradigm to any other system humanity has built at scale. When the system is started, it is ready to receive requests. It does not “do” anything, no computation, no planning, no word prediction, nothing, until a human decides to send a request in.
The scope of the requests a model can accept grows with every release. So does the amount of work it can complete on its own once it receives one. A single prompt can now launch hours of agentic work.
But these models have no way to hold a goal of their own across sessions. Some products now offer memory between sessions. That memory is a feature built around the model: stored notes that get fed back into the next prompt. The model itself does not carry a goal from one session to the next, and it has no way to “organize itself” as some misguided news reports would have you believe. The models that have supposedly “organized themselves” had a number of sub-agents orchestrated by a main agent that was acting, you guessed it, at the behest of an actual human, with a goal explicitly set by said human.
So the version of alignment most people fear has several milestones to reach before it becomes a salient issue. The narrower version is already here, and it is one more reason to keep a close eye on how these models behave.
Telling signal from hype
Every time a new model comes out, people on social media are quick to show which game they were able to single-shot and what virtual world they were able to build with just a few prompts. My own experience with the models over time has taught me how little these demos prove.
A single-shot demo shows that the model can fill gaps with plausible guesses. It says little about whether those guesses match your requirements. The more you ask a model to do with less information, the more assumptions it has to make. The more assumptions you let a model make, the higher the likelihood that what it creates does not match what you actually need.
Every model from every lab also comes with its own claims about safety and guardrails. Those claims are then amplified by people who may not understand them at all, or who amplify them in bad faith. There are zealots on every side of the AI debate who are willing to make any argument to bring you over to their side, and who do not want to, or cannot, recognize the real capabilities and limitations of the technology.
Direct experience is the filter. You learn what a model can do on your own problems, and you learn which commentators describe that accurately.
The objection: “Using it feeds it”
If your fear is that the labs won’t prioritize safety, using their products can feel like voting for them with your wallet and your data. If your fear is about jobs, working with AI can feel like training your replacement.
Informed critics carry more weight than uninformed ones. A critique built on direct experience is harder to dismiss. Refusing to use a technology has rarely slowed its adoption at scale, and abstaining leaves the conversation to the people least worried about it.
You also have choices in how you engage: which providers, which tiers, and open or local models you can run yourself.
On jobs, the developers who did well in the mobile transition were not the ones who stayed away from the platform.
What to do about fear
We can all agree it’s easier to defend ourselves against something we understand.
What engagement looks like in practice
Use AI on your real work and your real problems. Toy prompts teach you very little. Pay as much attention to where it fails as to where it succeeds.
Revisit with each major release. Your understanding has a shelf life as short as the release cycle. Compare what you see with what the commentators claim. Over time, this builds your own set of trusted experts.
On jobs
Because the technology is changing so fast, and the debate is so polarized, firsthand experience is the most reliable way to understand its capabilities and limits. We cannot predict the jobs of the future. If AI is going to change the nature of our work, understanding how it does so, and how to use the tooling, will leave us better prepared to adapt when those new jobs appear.
On guardrails
Telling hype from real news comes down to understanding enough about the models to recognize which claims match what you’ve observed. Being able to identify who is who in the sea of coverage will help you build a trusted set of experts over time.
On alignment
The milestones that would make the larger version of this fear salient are observable:
models that hold goals across sessions without human scaffolding
models that initiate work without a human request
multi-agent systems that coordinate without a human setting the goal
People who work with these models will see those milestones approach well before the headlines catch up. The people who stay away will learn about them from the same unreliable sources they distrust today.
Whatever your fear, you cannot verify every claim about AI yourself. But you can earn enough direct experience to know whose claims to trust, and how to understand them in your own context.


