Recently, I have kept coming across reels that are misogynistic in nature. Some talk about harassment of women as a joke, or how child marriage is the right way for society to work. The worst one was normalizing marital rape, saying that it is a wife’s duty and the language used by the creator was deeply disrespectful.

What made me write about this was when I saw that many women felt uncomfortable and reported his account. The platform acted. But even after that, he got his account back. All the reels are still there and he is making more content proudly, even. This means that somewhere in that process, a human or an algorithm decided that content normalizing marital rape does not violate community guidelines.

Why is the entire system automated AND human review failing women again and again?

This might look like a technical failure but it is not. This is not just a bug this is a business model. A system that prioritizes engagement over safety, because outrage drives engagement. Otherwise, why are people who create such content still on the platform?

Many might argue : how do we define misogyny? It can range from something subtle to something very bold. How many accounts will we deactivate? And will it even change the mindset of the person?

They are right it is difficult. But this content its outright outrageous that should not even be thought over. And yes, while it may not change the creator’s mindset, at least that person will not be getting a global platform to share these views and influence others.

Platforms will not change voluntarily. They need to be made to. That starts with holding them legally accountable for content they actively choose to reinstate

yes, platforms have billions of users. You cannot hold them accountable for everything posted. And that is true. But I am not talking about everything. I am talking about one account. One that was reported by multiple women. One that was removed. And then one that was actively reinstated. That is not a moderation failure at scale. That is a deliberate decision on a specific case. And specific decisions have specific accountability

Why People Still Don’t Trust AI: Reflections from User Research

A lot of conversations about AI adoption seem to assume that mistrust is a temporary problem. The idea is that once systems become more accurate, more transparent, or more tightly regulated, people will naturally begin to trust them.

My experience participating in user research for an AI-powered legal tech product made me less certain about that assumption.

In a limited, early-career role, I was involved in observing how users interacted with an AI system, how they interpreted its outputs, and how comfortable they felt relying on it. I did not work on the core model or make high-level product decisions. But even from this narrow vantage point, it became clear that mistrust was rarely about the technology “not working.” Instead, it was shaped by how people understood the system, what they expected from it, who they felt was accountable, and how the tool affected their sense of professional competence.

This piece is a reflection on a few things that stood out to me. I don’t claim that these observations apply everywhere, or that user research alone can explain AI governance challenges. But they helped me see why trust in AI is harder to build — and harder to regulate — than it often appears.

What discussions about AI trust often assume

When AI trust comes up in policy or governance discussions, it’s often framed in fairly linear terms. There’s an implicit belief that:

• Transparency = trust
• Accuracy = adoption
• Compliance = legitimacy
• Users follow intent

These assumptions make sense. Policymaking depends on simplification. But watching real users interact with an AI system suggested that trust doesn’t work in such a straightforward way. Even when a system met technical or legal expectations, people’s reactions were shaped by factors that sat outside performance metrics or formal rules.

1. People often don’t know when to trust AI

One thing that surprised me was that mistrust wasn’t usually absolute. Instead, many users were unsure about when the AI should be trusted.

Some leaned too heavily on AI outputs, treating them as authoritative even in situations where the system was clearly limited. Others remained skeptical even when the system appeared to work well. This wasn’t simply a divide between “trusting” and “distrusting” users. It was more about mis calibrated trust.

From a governance perspective, this feels important. Many regulatory approaches implicitly assume “reasonable use,” but what counts as reasonable varies widely depending on experience, context, and confidence. Trust turns out to be situational, not binary.

2. Transparency doesn’t always reassure people

Another thing I noticed was that explanations didn’t necessarily make users feel more comfortable.

Transparency is often treated as a solution to AI mistrust, and it does matter. But in practice, some users ignored explanations altogether. Others found them confusing, or felt they didn’t address what they actually cared about. In several cases, explanations clarified how the system worked, but not whether it should be relied on in a specific situation.

This made me think about the gap between technical transparency and human reassurance. From a governance point of view, it raises questions about approaches that rely heavily on disclosure without considering how people interpret and act on that information.

3. Accountability mattered more than accuracy

What seemed to matter most to many users wasn’t whether the AI was accurate in a narrow sense, but whether someone was clearly responsible for its outputs.

Questions like:

• “What happens if this is wrong?”
• “Can I challenge the result?”
• “Who is accountable here?”

came up repeatedly, either explicitly or implicitly. Even a system that appeared to perform well could feel untrustworthy if users weren’t sure what recourse they had when things went wrong.

This suggested to me that trust in AI is closely tied to trust in institutions. Governance that focuses mainly on model behaviour may miss how strongly legitimacy depends on accountability structures around the system.

4. Fear of losing skills shapes trust in subtle ways

Another source of scepticism that stood out was concern about skill erosion.

Some users weren’t just evaluating whether the AI was helpful. They were also thinking about what relying on it might mean for their own expertise over time. In a legal tech context, professional judgement is deeply tied to identity and credibility. For some people, using AI felt less like assistance and more like the beginning of dependency.

This concern influenced how cautiously the system was used, or whether it was used at all. Importantly, this didn’t seem to come from misunderstanding the technology. It felt like a rational worry about autonomy, deskilling, and long-term professional value.

From a governance perspective, this raised an interesting question for me. Regulation often focuses on preventing harm caused by AI systems. It pays less attention to how AI might quietly reshape human skills and decision-making authority, even when systems are technically safe and compliant.

What this suggests about AI governance

Taken together, these observations made me sceptical that trust in AI can be built purely through better models, clearer explanations, or stricter rules.

Trust seems to be shaped by things like:

• How AI systems are introduced and framed
• How uncertainty is communicated
• Whether accountability is visible and meaningful
• Whether people feel their agency and skills are preserved

This has implications for AI governance, especially in regulated domains like law or public services. Frameworks that focus mainly on pre-deployment checks may struggle to capture how trust actually develops — or erodes — once systems are in everyday use.

In international or multilateral settings, where institutional capacity and professional norms vary widely, these challenges are likely to be even more pronounced.

What I’m still unsure about

I’m still unsure how governance frameworks should incorporate user behaviour without becoming overly prescriptive. It’s also unclear how responsibility should be shared between developers, organisations, and users when AI systems subtly influence judgement over time.

User research offers one useful lens, but it’s far from sufficient on its own. How to meaningfully integrate human factors into scalable and adaptive governance remains an open question for me.

Closing reflection

Participating in user research didn’t give me answers about how AI should be regulated. What it did give me was a clearer sense of why trust in AI is fragile — and why it can’t be reduced to technical performance or legal compliance alone.

If AI governance is going to be effective, it may need to treat trust not as an output to be engineered, but as a social and institutional relationship — one shaped by experience, accountability, and a sense of human agency.

That’s the question this experience left me most curious to explore.

Heuristic method

Moving towards 1966 when ELIZA a chatbot developed by Joseph Weizenbaum that mimicked human conversation using pattern matching. Your statement “NLP relied on handcrafted rules and syntactic structures, which we call heuristic learning” is mostly correct but could be refined for clarity. In its early days, NLP relied on handcrafted rules and syntactic structures, a form of heuristic learning where systems followed predefined rules rather than learning from data.

Statistical NLP and Machine Learning method

The 1980s saw a shift towards statistical methods with the introduction of probabilistic models like Hidden Markov Models (HMMs) for speech recognition. NLP researchers started using corpora and machine learning algorithms instead of manually crafted rules.

During the 1990s, statistical approaches gained momentum. Techniques like n-grams, part-of-speech tagging, and decision trees were widely adopted. IBM’s statistical machine translation (SMT) laid the foundation for data-driven language models.

Neural Networks and Deep Learning method

In the 2000s, something magical happened in the world of NLP. Computers were no longer just memorizing words — they started understanding sequences like humans do when reading a book. This was made possible by Recurrent Neural Networks (RNNs) and their smarter sibling, Long Short-Term Memory (LSTM) networks. These models could remember past words in a sentence, making speech recognition, sentiment analysis, and named entity recognition (NER) much more accurate. It was like teaching computers not just to recognize letters but to understand the flow of a conversation.

Then came the 2010s — the decade of deep learning breakthroughs. Imagine giving words their own unique identity, like fingerprints. That’s what Word2Vec (2013) and GloVe (2014) did, helping machines understand words in relation to each other rather than in isolation. But the real revolution came in 2017 when a paper titled “Attention Is All You Need” introduced transformers — a technology that changed NLP forever. These models didn’t just read words in order; they focused on important words in a sentence, much like how our brain jumps to key details in a conversation. This led to the birth of BERT (2018) and GPT-2 (2019), making AI sound more natural and intelligent than ever before.

And so, the journey of NLP continues. What began as simple rule-based systems has now evolved into powerful deep learning models that can converse, translate, and even generate creative content. With self-supervised learning, multimodal AI, and real-time NLP applications, the future is brighter than ever. Perhaps, in the coming years, machines won’t just understand language — they might truly grasp its meaning, emotions, and nuances, bringing us closer than ever to the dream of human-like AI.