Aleksandr Timashov has spent years working at the intersection of data science and machine learning, most recently at Meta. We sat down with him to talk about something the industry often avoids: the ethical blind spots of the models reshaping how we work and think.
LLMs are everywhere now. What’s the problem nobody wants to talk about?
Bias. LLMs reflect the data they were trained on — and that data reflects the world, including hundreds of years of inequality. You can’t just ignore that. We have to actively curate training data, detect bias in models, and fine-tune them. It doesn’t happen by default.
They also just… make things up. How serious is that?
Very. LLMs are extremely good at generating information that looks convincing but is factually wrong. That’s dangerous at scale. I’m actually working on a guide for LLM evaluation right now — better assessment frameworks are one of the few real levers we have here.
Who’s most at risk from that kind of misinformation?
Anyone who trusts the output without verifying it. Which, increasingly, is a lot of people.
What about deliberate misuse — people using these models for harmful purposes?
It’s real. Deepfakes, automated cyberattacks, social engineering at scale. The uncomfortable truth is that most people use LLMs to do useful things — but some won’t. One approach I find interesting: using LLMs themselves to detect whether a prompt is aimed at malicious intent. A kind of ethical feedback loop built into the system.
Privacy is another concern. What does that actually look like in practice?
Models can retain fragments of personal data from training and surface them unintentionally. It’s not theoretical — it’s happened. We need much stronger data anonymization and stricter privacy guidelines baked into the pipeline, not bolted on after the fact.
If you had to pick one thing the industry should fix first, what is it?
Dataset curation. Everything downstream — fairness, misinformation, safety — starts with what you put in. If the foundation is bad, the model is bad. It’s not a technical problem, it’s a prioritization problem.
You mentioned probabilistic models as part of the solution. What do you mean?
Probabilistic frameworks let you quantify things that are usually treated as black boxes — fairness, ethical risk, uncertainty. If a model is not sure about something in an ethically sensitive area, it should be able to express that uncertainty rather than confidently produce a wrong or harmful output. Overconfidence is a real danger.
What would “responsible AI” actually look like five years from now?
Ethics embedded into the model’s reasoning, not added as a filter on top. Models that adjust their behavior based on cultural and societal context. And genuine collaboration — researchers, ethicists, policymakers, affected communities. No one group can solve this alone. The future of LLMs depends on whether we’re willing to treat this as a shared problem.
