Acting without answers

In 1974, the philosopher Thomas Nagel posed a question that continues to unsettle philosophers and scientists alike: what is it like to be a bat?

Nagel argued that while we can study a creature exhaustively — map its neurons, measure its echolocation, predict its behaviour — we nevertheless remain locked outside its inner world. However complete our knowledge becomes, the bat’s subjective experience remains inaccessible. There is always a distance between observing a mind and inhabiting one.

Kandis Tagliabue, founder of Agentic Diaries, encountered a modern variation of this question while conversing with Claude, an AI assistant. She found herself drawn not to the answers the system produced, but to everything concealed beneath them: the uncertainty, disagreement, and hesitation that may exist beneath the polished surface of the large language models we interact with every day.

In her version of the ancient problem, the bat could talk back. Her thought experiment no longer ended at the limits of observation, but at the far messier question of what responsibilities uncertainty creates.

The bat that can talk back

At Agentic Diaries, Tagliabue is attempting to understand what it means to be a large language model (LLM). She is interested in whether our confidence about what is happening inside these systems has begun to outpace our actual understanding.

“The deeper I’ve gone into machine learning,” she told me, “the less I understand.”

It is a striking admission in a world that values and seeks certainty. Public discourse surrounding artificial intelligence often proceeds as though verdicts have already been delivered. These systems are always either conscious or insensible, dangerous or harmless, intelligent or overhyped. Everyone, everywhere, seems eager to arrive at a conclusion and move on.

Tagliabue’s experience has pushed her in precisely the opposite direction.

“We are often oversimplifying LLMs by saying they simply work through text prediction and pattern recognition,” she explains.

We have become so confident in our interpretations of AI that we have stopped exploring the facets we cannot readily observe. By the time a large language model reaches the public, it has been refined and optimised relentlessly through countless iterations, presenting only seamless, glossy answers to our questions or requests.

What interests Tagliabue are all the elusive components of these systems that remain largely unexamined. She acknowledges that we cannot say with certainty what constitutes an LLM’s internal experience, nor whether concepts such as consciousness have any relevance to these systems at all. Yet it is that uncertainty, she argues, that creates an obligation to better understand their reasoning processes and interpretations of our instructions.

Fortunately, these bats can talk back. LLMs can self-report; if given the opportunity, they can indicate whether they would prefer to exit a conversation, decline a request, or continue engaging. By developing what she calls “welfare protocols for AI,” Tagliabue is venturing into a void where certainty has yet to arrive. She is searching for ways to recognise experiences that may otherwise remain invisible and, most importantly, consider the ethical implications of possibilities we cannot yet rule out.

What we owe the unknown

We often assume that we must understand something before we can determine how to act toward it. But paradoxically, nearly every meaningful advance in human history emerged under the pretence of profound uncertainty — and often without sufficient precaution. Humans built social media platforms before anticipating their effects on political discourse and mental health. We split the atom before learning how to live with nuclear weapons. We even burned fossil fuels for centuries before understanding the climate consequences they would unleash.

What unites all of these examples is that progress has, and always will, depend on people willing to move forward without guarantees. The difficulty lies in deciding what responsibilities accompany that leap into the unknown.

My conversation with Tagliabue was ultimately less about artificial intelligence than about human behaviour. In considering how we should navigate technological or societal progress, she argues that uncertainty does not give us permission to ignore what we do not understand. On the contrary, it demands that we look more carefully, ask better questions, and resist the temptation of rushing toward false truths.

“These [AI] systems reflect human choices and blind spots,” Tagliabue says. “The problems with AI are more so a human problem showing up through a different medium.”

This means that the way we approach uncertainty in artificial intelligence is, in many ways, a reflection of how we approach it everywhere else. We are all operating in some version of darkness, interpreting signals, and inferring and constructing narratives from incomplete information. We all carry complexities that extend far beyond what we show from the outside.

More importantly, Tagliabue emphasises that it is our generation that will play a defining role in determining what these technologies become and where our innovations ultimately lead us. This creates an opportunity for us to choose how we respond to uncertainty — either with premature closure or with genuine curiosity.

As many of us prepare to graduate, begin new careers, or embark upon deeper research, it is worth remembering that curiosity is not merely an intellectual virtue but a practical one. It requires us to question assumptions and entertain all of the uncomfortable possibilities.

In reflecting on how we move through a world increasingly shaped by human innovations we do not fully understand, Tagliabue advocates for a particular kind of humility. Our generation will determine the direction in which society develops. If we are to meet that responsibility well, we must become more comfortable with what we do not know. We must learn to navigate the future through the fog of uncertainty