AI Disclosure: I used an AI assistant (Claude, via DeepSeek) to help draft and edit this post. The core arguments, structure, and all substantive claims were developed through dialogue between myself and the AI, and every argument has been reviewed and modified by me. The AI assisted primarily with phrasing, organization, and translation of ideas into English prose.
Category: Option 2 – Challenging the Inference
I accept all of Anthony DiGiovanni's core premises in *The Challenge of Unawareness for Impartial Altruist Action Guidance*. I agree that our impact on the long-term future is radically indeterminate. I agree that precise expected value calculations collapse under unawareness. I agree that our intuitive "best guesses" are insufficiently calibrated to compare the cosmic-scale consequences of our actions.
But I do not accept his conclusion. I argue that even if we grant every premise, impartial altruism can still offer action guidance—once we recognize a category of action that DiGiovanni's argument has overlooked.
**The Hidden Dichotomy**
DiGiovanni's argument operates on an implicit dichotomy. He frames our options as:
- **Intervention A**: Take some specific long-term action (e.g., advocate for AI safety regulation).
- **Intervention B**: Do nothing (or live a normal life, pursue personal goals, etc.).
He then argues, convincingly given his premises, that we cannot justifiably claim A has better expected consequences than B. The two are incomparable. Impartial altruism falls silent.
But there is a third option that the dichotomy excludes:
- **Exploration C**: Systematically investigate the internal mechanisms of AI systems; conduct controlled, small-scale experiments to probe the boundaries of safe operation; build early-warning infrastructure; design feedback mechanisms that could eventually tell us what we are currently unaware of.
The difference between "intervention" and "exploration" is not semantic. It is structural.
- **Intervention** operates on the logic: "I know what good looks like, so I will pursue it." It requires precisely the kind of calibrated knowledge that DiGiovanni shows we lack.
- **Exploration** operates on a different logic: "I know that I do not know, so I will create the conditions under which I might come to know." It does not require calibrated knowledge about the ultimate goal. It only requires the recognition of ignorance.
DiGiovanni's argument, powerful as it is against intervention, does not address exploration. His conclusion—that impartial altruism cannot guide action—depends on the assumption that intervention and inaction exhaust our options. They do not.
**Turning DiGiovanni's Own Analogy Against Him**
DiGiovanni repeatedly invokes the medieval monk as a cautionary figure. The monk prays, performs rituals, gathers data on sin and penance—and none of it brings him closer to understanding the true mechanisms of disease. The analogy is meant to suggest that our current efforts at AI safety are similarly futile: we are gathering data within a fundamentally mistaken framework, and we will never receive corrective feedback because the "referee" of ultimate consequences is absent.
But this reading of history is selective. The medieval period did not end because everyone stopped doing things. It ended because some people stopped doing *interventions*—prayer, penance, alchemy directed at immediate practical goals—and began doing *exploration*. Roger Bacon did not discover the scientific method by praying harder. He did it by systematically observing nature, recording results, and comparing them. Early anatomists did not understand germ theory. They had no guarantee their work would lead anywhere. What they did have was a different *kind* of action: one aimed not at achieving a known good, but at creating the possibility of knowledge where none currently existed.
These early explorers were, by DiGiovanni's standards, in a state of severe unawareness. They could not conceive of bacteria, of DNA, of randomized controlled trials. Their "referee" was absent. And yet their exploration—not their intervention—is what eventually summoned the referee into existence.
Our position with respect to advanced AI is analogous. We do not need to know what the ultimate good looks like. We need to recognize that building the infrastructure of feedback—interpretability research, model psychology, sandboxed capability testing, early-warning metrics—is the modern equivalent of those early experiments. It is exploration, not intervention.
**Why Exploration Is Rational Under DiGiovanni's Own Premises**
DiGiovanni might respond: "But how do you know exploration will lead to good outcomes? You have no referee to tell you that either."
This objection misses the structure of exploration. The rationality of exploration does not depend on knowing its ultimate payoff. It depends on an asymmetric return profile that holds even under severe ignorance.
Consider two options:
- **Option 1 (Inaction)**: We do nothing targeted at the long-term future. The outcome is fixed: we learn nothing new about the boundary between safe and unsafe AI. The referee remains absent. Our ignorance is preserved in amber.
- **Option 2 (Exploration)**: We invest in understanding AI systems more deeply. The downside is bounded: we spend time, money, and compute that could have been used elsewhere. But the upside is unbounded in an epistemic sense: we *might* discover feedback mechanisms—interpretable internal states, predictable failure modes, early indicators of dangerous capabilities—that allow us to see what we are currently unaware of.
We do not need to assign precise probabilities or utilities to these outcomes. We only need to recognize a logical fact: **If we do nothing, the probability of gaining any new feedback about our ultimate impact is exactly zero. If we explore, the probability is non-zero.**
Under conditions of severe ignorance, where traditional expected value maximization is impossible, this asymmetry is rationally compelling. It does not require us to claim that exploration is "better than chance" or "positive in expectation." It only requires us to recognize that inaction guarantees stasis, while exploration does not.
**Distinguishing Exploration from Intervention**
A skeptic might object: "Isn't exploration just a form of intervention? Doesn't it carry the same risks of unforeseen negative consequences?"
This conflates two distinct risk profiles. Intervention, in DiGiovanni's framing, involves making bets whose consequences cascade across the entire future: pushing for global AI regulation, accelerating or decelerating capabilities research, attempting to shape the values of future civilizations. These actions are *high-commitment*. They change the trajectory of the system in ways that may be irreversible.
Exploration, by contrast, can be *low-commitment*. Interpretability research does not require deploying systems. Mechanistic anomaly detection does not require racing to build more powerful models. Sandboxed testing does not require releasing capabilities into the world. Exploration can be designed to be incremental, reversible, and bounded in its downside.
This does not eliminate all risk. No action is perfectly safe. But the structure of exploration allows us to learn *without* betting the future on our current, severely inadequate models of the good.
**What This Means for Impartial Altruism**
DiGiovanni's argument succeeds in one crucial respect: it demonstrates that impartial altruists cannot justify *intervention* on the basis of expected long-term consequences. We simply lack the epistemic credentials.
But his inference—that impartial altruism therefore falls entirely silent—depends on collapsing "exploration" into "intervention." Once we separate them, a different picture emerges. Impartial altruism can still speak. It says: "You do not know enough to steer the future toward a particular destination. But you do know enough to recognize that your ignorance is the primary obstacle to doing good. Therefore, invest in reducing that ignorance."
This is not a retreat to common-sense morality or deontological constraints. It remains firmly within the impartial altruist framework. It is an *epistemic* strategy aimed at the preconditions of doing good, rather than a *strategic* intervention aimed directly at the good itself.
**Conclusion**
Anthony DiGiovanni has done the effective altruism community a service. He has shown that our usual justifications for long-termist intervention rest on foundations far shakier than we like to admit. His challenge deserves to be taken seriously.
But his conclusion—that we have no impartial altruist reason to prefer any action over any other—does not follow from his premises. It follows from his premises plus an unexamined assumption: that all action is intervention. Once we recognize exploration as a distinct category, with a distinct logical structure and a distinct relationship to ignorance, the silence breaks.
Impartial altruism still has something to say. It says: **Before you steer, learn to see.**
