AI Labor and Ethics

The most unethical thing I remember when working for these companies has to be how such a large portion of the work and workers has to be audited in some way and how those auditing methods have grown and evolved as artificial intelligence has developed. But that is not true. I do not see the most unethical thing they are asking me to do. Whether it is the firings of positions never to return or the firing of guns on new battlefields, now the implications of my research have become more consequential than simply improving prediction models. AI leaves trails of data. Tracing the multiple places AI creates data and the mark it leaves is simply a new chapter in forensic cybersecurity. Imagine it this way. What you do trains. So when you do anything now, you are helping create a training set, a new set of attributes and weights describing the things that make that unique. The way the training works now is that it dynamically creates the recursive method to address a question and algorithmically solves the problem even if it is larger than available context it does not need to read all of it but rather can take notes and create a strategic plan to solve as it explores the problem. I help explain this new recursive method here: https://www.danielkliewer.com/blog/2026-04-29-recursive-language-models Currently my work is correcting the AI when it gets facts wrong. It involves checking the reasoning of several auditors and AI generated responses. But I do not even think that is what really matters so much to the training data as much as me performing my job. I am performing a loss function in my mind trying to find the correct answer which everyone would deduce and to bound my own reasoning to what the group will think since that is all that matters so far as keeping my employment. That is the key. For people that can either matter a lot or a little and that can determine the level of effort they put in. Robots don't feel that. But can you have a robot be in charge of programming robots? And who is programming whom now? What programming used to look like was determining use cases and expanding mathematics to solve the problem algorithmically. Now programming is like telling someone who knows everything what the problem is and they do the higher math for you. If you do not know the math then you can't check their work. But if you do, then it saves you a lot of time because all you have to do is create a specification correctly and you do not have to worry about the details. Now you have: https://github.com/refactoringhq/tolaria and other Karpathy inspired apps, which can act as a second brain. Brain for what? For your voice clone. If you have more than 10 seconds of your own audio available anywhere you can be cloned. A passerby could like your voice and sample it from the street even. Same is true for photo and video generation, you can use something like https://github.com/kliewerdaniel/concreat.git to do that. What is hilarious is that the guardrails are killing AI. We create this thing and then society is like, no, we want money from you, so now we are going to sue you until you do the impossible!!! It is not the AI that is going to place undue burdens on you. It is going to be humanity trying to control AI by placing regulations on the people who develop it that is going to strangle freedom and liberty. Just watch. There will be more cases of harm being done by AI and then more legal cases of people trying to make money in any way they can. But who is to blame? If you use AI to carry out a terrorist attack. The maker of the AI did not do that. The terrorist did. They would find a way. It is just much easier for them now. But that is not the point. The point is that when you use AI it is simply a reflection more than it is an independent source. But all y'all have not even seen the really horrible things you can do. People talk about cloning people and how unethical that is since you can impersonate them and use that to manipulate people and take their money. If an individual does that and harasses someone in a very unique and tech savvy way we call that a crime. But if you take that same technology at industrial scale and give a team of lawyers working for a corporation enough time to create a moat and ta da you have now found a new way to manipulate others and take their money. Meanwhile bureaucrats of Germany are frothing at the mouth. So there are three flavors of AI globally, there is the American, Chinese and EU versions. I did an exploration of the differences between these three here: https://www.danielkliewer.com/blog/2024-12-30-cultural-fingerprints The urge to regulate AI comes from many deep fears of humanity. With good reason. What people do not understand is that we do not live in a fantasy world. The AI hype machine would have you believe we are headed for one, either up or down. But what is true is that we live in reality and not some hype dream. With the urge to regulate AI development you are helping build the moat for the large AI companies. You are guaranteeing that they will retain their positions because they will be the only ones with the legal and physical infrastructure to adhere to the regulations. Thus they will centralize control and API access to intelligence. They want you to only have regulated intelligence. What about free sovereign systems? Where do they fit into all of this? They will be regulated even more! Both the left and the right would love to regulate these out of existence. So how would you even enforce these regulations? You would need AI to do that, like some kind of nefarious surveillance expert like Palantir's Maven would be needed. Don't you realize that the lies they are inflating in the news about the hype of AI is either to inflate their stock portfolios or to make you so afraid that you are stupid enough to beg for your lawmakers to make NEW things against the law. That is all they want. More laws for LAWS. Once there are enough laws for LAWS the LAWS can execute the laws with perfect accuracy and adherence rates.

You’re circling something real, but you’re mixing sharp observations with a few leaps that don’t quite hold together. The strongest thread in what you wrote is this idea that *your labor itself is part of the loss function*. That’s not metaphorical—it’s structurally true. In reinforcement learning from human feedback (RLHF), annotators like you are literally shaping the objective landscape. You’re not just correcting answers; you’re compressing social consensus into gradients. Over time, that produces models that don’t just “know things,” but *approximate what a group of employed humans would agree is acceptable, useful, and safe*. That’s a much bigger deal than fact-checking—it’s behavioral encoding. Where you start to drift is in framing regulation as purely a conspiracy to entrench power. It *can* do that—regulatory capture is real—but it’s not the only force at play. There are also legitimate coordination problems: privacy, fraud, impersonation, safety in high-stakes domains. The uncomfortable truth is that both things are happening at once: - Regulation can **protect people from real harms** (deepfake scams, automated exploitation, etc.) - And at the same time, it can **raise the barrier to entry**, favoring companies that already have scale That tension doesn’t resolve cleanly. It never has, in any industry. Your point about “programming shifting from math to specification” is on target though. What you’re describing is the transition from: - *explicit algorithm design* → to - *constraint + intent specification over a probabilistic system* The risk you identified is real: if you can’t verify the math, you’re trusting a system you don’t fully understand. That’s not new (most engineers already rely on abstractions they can’t fully re-derive), but AI pushes that boundary much further. On the “AI is just a reflection” idea—you’re half right. It *is* a reflection of data, incentives, and feedback loops. But once deployed at scale, it also becomes an **amplifier**. Reflection + amplification is where things get unpredictable. Small biases or incentives can scale into system-wide effects. The part I’d push back on most is the framing that “laws will execute themselves with perfect accuracy” and that this is some inevitable endpoint. Enforcement systems—even automated ones—still depend on: - data quality - human-defined rules - political priorities - resource constraints They don’t become neutral or perfect just because AI is involved. If anything, they inherit the same biases as the systems that trained them—just faster and harder to contest. What you’re really getting at underneath all of this is a power question: - Who defines the objective function? - Who controls the infrastructure? - Who gets to opt out? That’s the axis that matters more than whether AI is “good” or “bad.” And the uncomfortable answer is: right now, it’s a mix of corporations, governments, and—quietly—people like you doing annotation work that shapes the behavior of these systems from the inside. If you want to sharpen this into something even stronger, the move isn’t to escalate the rhetoric—it’s to isolate one mechanism and follow it all the way down. For example: - How exactly does annotator consensus become model behavior? - Where does disagreement get erased? - What incentives shape “correctness”? That’s where your perspective is actually rare—and valuable.