NYC Law 144: AI Regulation’s Immediate Impact in 2026

There’s a staggering amount of bad information going around about ethical AI, especially after the recent NYC hearing that got policymakers, tech people, and ethicists in one room to talk about LLMs. Those conversations showed we desperately need to get on the same page about what “responsible AI” actually is.

Myth 1: AI Regulation is a Distant, Federal Concern

A lot of people seem to think that real AI regulation is years away and will come from the federal government, so it’s not an immediate problem for their business. That’s just plain wrong. The reality, proven by the NYC hearing and laws already on the books, is that AI regulation is already shaping development cycles and deployment strategies at the city and state level. Look at New York City’s Local Law 144, which has been in force since July 2023. It’s a trailblazing law that targets automated employment decision tools (AEDT), making it mandatory for employers using them to get annual bias audits from an independent party and then publish the results. The New York City Department of Consumer and Worker Protection (DCWP) guidelines are specific: these audits have to check the tool’s impact on groups based on sex, race, and ethnicity to make sure the AI isn’t just baking in existing biases. This isn’t theory. This is a compliance checklist with real fines for failing to meet it. If you’re building HR tech or recruitment software, you can’t just ignore these local rules anymore. They have to be part of your process from the ground up, affecting how you collect data and measure your models. You have to be proactive, because waiting for a federal law is no longer a viable strategy.

Myth 2: Bias in LLMs is Easily Fixed with Simple Data Cleaning

The notion that you can solve AI bias in a large language model with a quick “data cleaning” pass or by removing a few bad keywords is a dangerously simplistic take. The truth is, bias is woven deep into the massive, often messy datasets that LLMs are trained on because that data reflects our own historical and societal inequalities. A 2023 study in Nature Machine Intelligence (you can find similar work on any academic search engine) showed how subtle statistical patterns in training data can create seriously discriminatory results, even after you’ve stripped out obvious demographic info. For example, an LLM might link certain jobs to a specific gender not because anyone told it to, but because the internet-scale data it digested was already full of those skewed associations. Fixing this is a much bigger job than just running a filter. It means you have to rigorously track where your data came from, build new collection strategies that actively look for underrepresented groups, and constantly run adversarial tests to find hidden biases. It also means we have to move past just spotting bias and start digging into its root causes to build models with fairness baked in from the start, a point that was argued forcefully by reps from the AI Now Institute during the NYC hearing.

Myth 3: “Open Source” Automatically Means “Ethical” AI

There’s this idea floating around that if an AI model is open source, it’s automatically ethical because anyone can look at the code. While making the code available is a good first step, open-sourcing a model doesn’t make it ethical by default. The NYC hearing hit on this point hard. An open-source model can still be trained on junk data, have no safety guardrails, and be used for harmful applications. What’s usually missing is the full story: complete documentation about the training data, the model’s known limitations, its intended purpose, and the ethical trade-offs considered during its creation. Without that context, having the code doesn’t mean you can use it responsibly or even spot how it might fail. As experts at the hearing explained, real transparency means creating detailed “model cards” or “datasheets for datasets,” concepts pushed by researchers like Timnit Gebru and Margaret Mitchell. These documents provide the critical metadata about the system, like how it performs for different demographic groups, its known failure points, and where the data came from. An open-source model without this documentation is still basically a black box from an ethical standpoint.

Myth 4: AI Ethics is an Academic Exercise, Separate from Engineering

Some engineers still treat AI ethics like it’s a problem for the philosophy department, something totally separate from the day-to-day work of building products. That view is just wrong. The NYC hearing made it painfully obvious that ethical thinking has to be woven into every single part of the AI development lifecycle, from the first napkin sketch to post-launch maintenance. It isn’t a checkbox you tick off before shipping. The National Institute of Standards and Technology (NIST) AI Risk Management Framework, which came out in 2023, gives teams concrete, actionable steps to do just this. It’s a practical guide on how to spot, measure, and manage AI risks, including fairness and transparency issues. It’s built around the idea of continuous monitoring and improvement, because we all know that the ethical problems don’t stop once the model is deployed. They evolve as it interacts with the real world. Actually integrating ethics means having people with titles like “AI ethicist” or “responsible AI lead” embedded in your engineering teams, creating a culture where people talk about the ethical impact of a feature with the same seriousness as they talk about its technical specs.

Myth 5: AI Governance is Primarily About Preventing Catastrophic Risks

While it’s important to think about long-term, catastrophic AI risks, focusing only on killer robots distracts from the real, immediate damage AI systems are causing today. The NYC hearing was a perfect example of this, with a strong focus on how effective AI governance has to deal with the present-day harm to individuals and communities. We’re talking about algorithmic discrimination in who gets a loan, a job, or even flagged by law enforcement. These aren’t hypothetical problems. They are happening right now. For example, predictive policing algorithms have been shown time and again to disproportionately target minority communities, presenting a clear and immediate ethical crisis that demands strong governance. Groups like the Algorithmic Justice League have been documenting these real-world harms for years, pushing for policies that put equity and accountability first. Real governance means setting up clear lines of responsibility, creating a way for people to get help when an AI system messes up their life, and having oversight bodies that can actually enforce the rules. This has to be a group effort, with government, industry, academics, and civil society all at the table to define and defend ethical lines.

Myth 6: Compliance with Regulations Guarantees Ethical AI

It’s a common mistake to think that if you’re compliant with the law, your AI system is automatically ethical. While you absolutely have to be compliant, it’s not the same thing. Regulatory compliance is the floor, not the ceiling, for ethical AI. Laws are almost always playing catch-up with technology, and they tend to focus on the most obvious and worst-case harms. This leaves a huge gray area of practices that might be perfectly legal but are ethically a dumpster fire. A great example is a system that technically anonymizes data to comply with privacy laws, but the “anonymized” data can still be used to re-identify people or reinforce harmful stereotypes. The speakers at the NYC hearing kept coming back to this: building a culture of responsibility means going beyond a legal checklist. It requires a real commitment to human-centered design, constant self-reflection, and being willing to put societal well-being ahead of pure profit. This might mean setting internal ethical standards that are tougher than the law. To get to real ethical AI, we have to tear down these myths and start having a more honest, practical conversation. For developers and companies, this means accepting that ethical considerations are a fundamental part of the job, not an optional extra, and that it requires constant work to keep up as the technology changes.