The Neutral Tool Myth

Whenever AI causes harm, or simply just controversy, someone inevitably tries to cushion the blow by pointing out that it’s just a tool. The implication is that the technology itself is neutral, like a hammer or a calculator, and that any moral questions belong entirely to the user.

That idea is emotionally appealing as it relieves responsibility and suggests we can keep building and deploying powerful systems without having to argue about ethics, politics, or human consequences. But it’s increasingly false.

AI isn’t neutral in the way people mean it. Not because AI has opinions, but because AI is built out of choices. Choices about what data counts as knowledge, what outcomes count as success, what risks count as acceptable, what behaviors count as safe, and whose experiences count as normal. Those choices embed values in the system long before a user ever touches it.

The “neutral tool” myth breaks down for a simple reason: AI is not one tool, but rather it’s a pipeline of decisions. And every stage of that pipeline encodes priorities.

What People Mean By Neutral and Why it Doesn’t Fit AI

When someone says a hammer is neutral, they mean the hammer doesn’t decide what to build, the user does. AI is different because it doesn’t just execute the users intent, it often interprets your request, makes recommendations, and generates content that can influence beliefs, mood, and behavior. It can gatekeep opportunities (hiring algorithms), shape public perception (recommendation feeds), and decide what information is visible or invisible (ranking systems). It becomes a mediator between people and reality.

A better analogy than “tool” is institution: AI systems behave more like institutions because they enact rules at scale, sometimes without transparency and without appeal. Institutions are never neutral. They are governed by values.

Where the Values Enter: the AI “Value Stack”

To see why neutrality is a myth, it helps to map where values get embedded. Think of AI as a stack of layers, with each layer including decisions that look technical, but are also moral.

1. Data is a Value Choice in Disguise

AI systems learn patterns from data. That sounds neutral until you ask: which data? from whom? from what time period?with what biases? If trained on the open internet, it is exposed to both humanity’s brilliance and humanity’s garbage, except it cannot tell which is what.  If the internet is filtered, a value judgment about what is acceptable to include is being made. If certain sources (mainstream media, academic databases, social platforms) are emphasized, certain institutions and perspectives become more prominent. If engagement-optimized content is included, the model learns that attention-grabbing language is normal.

Even the act of labeling data, deciding what is “hate,” “harassment,” “misinformation,” “quality,” “helpful,” “harmful”, is inherently value-laden. Two people can sincerely disagree about what counts as hate versus criticism, or satire versus harassment. Those disagreements don’t disappear when they are turned into a dataset. They instead get frozen into the model. Neutral data does not exist. There is only data plus decisions.

2. Objectives Encode Priorities

Every AI system is trained to optimize something, and that something is a value choice. If optimized for user engagement the result is content that triggers emotion., and if optimized for helpfulness the result is confident answers (even when certainty is unwarranted). If optimized for safety the result could be refusals that frustrate legitimate users, and if optimized for profitability the result is almost certainly  that shape behavior in predictable (and often socially destructive) ways.

Even the phrase “maximize accuracy” contains values. Accuracy about what? Under what definition? With what tolerance for uncertainty? In whose language? For which communities? Using which sources?

A system optimized for minimizing harm will behave differently than a system optimized for maximizing expression. A system optimized for reducing legal risk will behave differently than one optimized for public benefit. These are not neutral settings. They are governance choices.

3. “Alignment” is Values, Formalized

Modern AI systems don’t just learn from raw text,  they are tuned using human feedback, like people rating outputs as better/worse, safe/unsafe, helpful/unhelpful. This is often described as “alignment.” Alignment is unavoidable. If you deploy a system widely, you must define boundaries for what should be refused, how sensitive topics should be handled, how harassment or requests for wrongdoing should be respond to, and how uncertainty should be dealt with.

But notice what’s happening: humans are teaching the system a set of norms. That is values-by-design. When people argue that AI is biased or censored, they are usually arguing about this layer, about whose norms were used, how they were applied, and whether they were consistent.

4. Interface Design Shapes Behavior

 Most of the impact of AI doesn’t come from the model alone. It comes from how the model is packaged. Does the chatbot show sources or speak without citations? Does it signal uncertainty or speak with confident tone? Does it encourage verification or encourage copy-paste answers? Does it default to share and repost, or build friction into virality?

A system that presents itself as an authority invites trust. A system that presents itself as a draft assistant invites checking. The interface is a moral choice because it shapes how humans relate to the output, whether as guidance, as truth, as suggestion, or as entertainment.

This is why “it’s just a tool” is misleading. Tools don’t typically try to persuade the user that they’re right. Many AI systems do, implicitly, by sounding fluent and certain.

5. Deployment Context Becomes Ethics in Practice

An AI model used for brainstorming is low-stakes. Use the same model for hiring, policing, medical triage, or news summaries, it becomes high-stakes. Values show up in deployment decisions: Where is AI allowed to replace a human? Where must a human remain in the loop? Who is accountable when it fails? What appeals exist for people harmed by automated decisions? How is performance monitored over time?

These are governance questions, but they are also values questions. They determine whether AI increases dignity and opportunity, or erodes them.

A common response is to concede that AI has values in it, after all, humans built it. Everything has values. The problem is not that values exist. The problem is that values are hidden behind the language of neutrality. When values are hidden, they become unaccountable. And unaccountable power produces predictable harms, especially for minorities, particularly during conflict, and certainly when systems scale faster than public understanding.

We don’t need value-free AI, which would be impossible to achieve. We need value-transparent AI, with systems whose priorities are legible, contestable, and open to correction.

Why Neutrality is Often a Power Move

Another reason the neutrality myth persists is because claiming neutrality is a way to avoid scrutiny. If a platform claims it is neutral, it avoids explaining why certain content is boosted or suppressed. If an AI company claims to be just a tool, it avoids responsibility for predictable misuse. If the company insists its model is objective, it is attempting to frame critics as political rather than principled.

But neutrality is not the absence of values, but the presence of default values often driven by incentives like profit, growth, or legal risk. Those defaults aren’t neutral. They’re just invisible.

A Practical Example: Bias vs. Safety

AI is often the focus of two common, contradictory complaints, that It is biased, and that it is censored. Both can be true at different times, for different topics, for different users. Why? Because the system is trying to satisfy competing values like helpfulness, safety, openness, liability, brand risk, political pressure, and user trust. The system cannot maximize all of them simultaneously, so tradeoffs are inevitable. The moment we recognize that tradeoffs exist, we admit that AI is embedded with values.

Jewish Ethics Offers a Useful Lens: Not Neutrality, But Responsibility

Jewish tradition is not built around the fantasy of value-free life. It assumes moral tension is real, that human speech can harm, and that power requires limits. It recognizes that communities must set boundaries to preserve life and dignity.

One need not be a Talmudic scholar to recognize the relevance: systems that shape human behavior require moral guardrails. In Jewish terms, we might call it pikuach nefesh when harm is physical, or lashon hara when harm is social and reputational, or tzelem Elohim when dignity is degraded. In plain civic language: safety, integrity, and dignity.

The point isn’t that Jewish ethics provides a single policy blueprint, but rather that it refuses the lazy escape hatch of neutrality. It asks: What does responsibility require, given real human weakness and real consequences?

So What Should We Demand Instead of Neutral Tools?

 If we retire the neutrality myth, we can replace it with better expectations that are realistic and actionable.

1. Transparency About Priorities – AI systems should clearly communicate what they optimize for, such as helpfulness, safety, speed, creativity, certainty, or some balance. Users deserve to know what they’re getting.
2. Honest Uncertainty – If a system doesn’t know, it should say so. Confidence should not be the default tone. Uncertainty should be a feature, and not seen as a failure, especially in sensitive domains.

3. Evidence and Provenance – When a system makes factual claims, it should show primary and reputable sources. This is one of the simplest ways to reduce confident error at scale.

4. Accountability for Harm – When AI is used in high-stakes settings, someone must be accountable. Not the algorithm or the user. Real accountability includes audits, monitoring, appeal processes, and human responsibility.

5. Cultural Humility – No single institution owns the truth. But that doesn’t mean every view is equally grounded. AI systems should be built to handle disagreement responsibly by presenting multiple credible perspectives, distinguishing facts from interpretations, and avoiding the false comfort of fabricated certainty.

 The Bottom Line: AI is Governance in Disguise

The neutral tool myth is tempting because it lets us avoid conflict. But avoiding conflict doesn’t prevent value choices, it simply ensures that the values are chosen by default, and often by whoever has the most power and the least scrutiny.

AI always has values built in because AI is not just computation. It is selection (data), optimization (objectives), mediation (interfaces), and distribution (deployment). It shapes what is seen, believed, rewarded, and punished. That is governance, whether we admit it or not.

The question is not whether AI will carry values. The question is whose values will be embedded, how transparently, with what accountability, and with what ability for communities to push back when reality is being bent in dangerous directions.

The sooner we drop the fantasy of neutrality, the sooner we can have the only conversation that matters, a conversation around what kind of moral and civic world do we want these systems to build, and what guardrails do we require to get there?