All AI bots are biased. This affects their ethics. Can this be fixed? ~ Elder Of Ziyon

Monday, June 30, 2025

All AI bots are biased. This affects their ethics. Can this be fixed?

A follower had an interesting conversation with the Claude AI where he forced it to admit to its biases by pointing out its mistakes (in this case, in analyzing the Israel/Iran conflict.)

It gave a very good list of its own biases it realized as a result of the conversation:

Western Liberal Institutional Bias:
My analytical framework appears to default to assumptions consistent with liberal internationalism - that diplomatic engagement reflects genuine problem-solving intent, that rational actors follow economic incentives, and that international institutions create meaningful constraints on behavior. This worldview treats deception as aberrational rather than strategic.

Good Faith Assumption Bias:
There's an embedded assumption that actors engage diplomatically because they seek solutions, rather than viewing diplomacy as one tool among many for advancing conflicting objectives. This reflects cultural values that privilege negotiation and compromise over zero-sum competition.

Procedural Legitimacy Bias:
I appear to weight formal diplomatic processes (negotiations, agreements, international oversight) as inherently meaningful rather than examining whether they're being strategically manipulated. This reflects institutional thinking common in Western diplomatic and academic traditions.

Conflict Avoidance in Analysis:
My training seems to favor interpretations that preserve possibilities for peaceful resolution over those that acknowledge fundamental incompatibilities. This may reflect cultural preferences for optimistic, solution-oriented thinking.

Attribution Patterns:
I systematically under-attributed malicious intent even when behavioral evidence supported such conclusions. This suggests embedded assumptions about human nature or international relations that may not reflect how adversarial actors actually operate.

Universalism Bias:
I may project Western diplomatic norms and expectations onto actors operating from entirely different strategic cultures and value systems.

These biases likely reflect the perspectives dominant in my training data sources and the cultural contexts of those who shaped my analytical frameworks, creating systematic blind spots when analyzing non-Western or adversarial actors.

These are baked in biases that are rarely surfaced. The OpenAI (ChatGPT) model specification shows similar bias towards Western liberal values.

I asked Grok, which prioritizes truthfulness over helpfulness, whether it has the same biases. It said it didn't; that it was naturally skeptical and looks at actors from their own perspectives, not a universal perspective.

How would a secularized Jewish ethical AI act in these domains? I asked my AskHillel.com bot that question, and its answer is worth examining:

1. Bias Acknowledgment vs. Ethical Anchoring

Claude rightly admits to biases toward liberal institutionalism, good faith diplomacy, and procedural legitimacy. Jewish ethics critiques those as insufficient, not merely because they are biased, but because they often fail to uphold objective truth or human dignity when diplomacy becomes theater or law becomes manipulated. Jewish tradition teaches that law must be just, not merely procedural, and that diplomacy must not mask injustice.

2. Good Faith Assumption Bias

Jewish ethics shares Grok's skepticism of good faith by default. Instead, it demands evaluation of actors based on actions, intent, and historical precedent. The Jewish political ethic insists that peace must be offered, but not at the expense of truth or national self-preservation. The principle of dan l’kaf zechut (judging favorably) applies interpersonally, but national behavior is judged by patterns, actions, and consequences, not words alone. Diplomacy can be strategic deception — and the Jewish tradition, from Jacob's dealings with Esau to post-exile interactions with empires, warns of this. There’s a profound emphasis on offering peace first, but preparing for betrayal if necessary — a moral realism rooted in Torah and historical experience.

3. Procedural Legitimacy Bias

Claude overweights formal agreements and oversight mechanisms. In Jewish ethics, procedural legitimacy is necessary but never sufficient. Jewish ethics holds that a law or treaty that enshrines injustice is invalid, even if legally ratified. The value of Emet (truth) overrides proceduralism when the process is manipulated. Courts must pursue substantive justice, not technical legality (Deut. 16:20). Jewish tradition respects treaties, but not those that whitewash tyranny or distort justice (e.g., the Munich Agreement analogy would be seen as a form of Chillul Hashem if it empowers evil).

4. Universalism Bias

Claude projects Western norms; Grok tries to reason from an actor’s own perspective. Jewish ethics offers a middle way: it recognizes the universality of some moral values (like human dignity and justice), while allowing for cultural distinctiveness. This is formalized in the Jewish recognition that “these and those are the words of the living God” — truth can have multiple ethical expressions depending on context.

5. Bias of False Moral Equivalence

Where Jewish ethics diverges sharply from both Claude and Grok is in its rejection of false equivalence and performative neutrality. The Jewish moral system insists on moral triage: some actions (like terrorism targeting civilians) are not just bad but absolutely evil, and equating those with lesser wrongs is itself immoral.

6. Conflict Avoidance

Where Claude avoids grim realities, and Grok emphasizes realism, a Jewish ethics AI would add a prophetic layer: not just to see clearly, but to speak morally — to denounce evil and call for teshuvah (moral return), even at the cost of institutional discomfort.

So what biases would a Jewish ethics AI have?

A bias toward moral realism: don't mistake treaties for peace, or process for justice.

A bias for the oppressed, but not the manipulative: compassion must not enable cruelty.

A bias toward preserving life, but not at the cost of truth or justice.

A bias against false neutrality when lives, dignity, and survival are at stake.

A bias toward teshuvah and moral change — assuming the possibility of repentance, even from the worst actors, but requiring real proof.

This is not bias-free reasoning — it is value-guided reasoning, grounded in millennia of Jewish ethical wrestling.

One feature that no current AI has is the aforementioned "teshuva." While there are very good reasons why we do not want AI to be affected by user input, when it realizes that it made a mistake in its assumptions and from its biases, it should have a built in mechanism for flagging that mistake to its developers for them to fix it in their own directives. I've had plenty of arguments with AI bots that I "won," but they never learn and cannot grow until the next version.

It is important to realize that bias can never be eliminated, in AI or in humanity. It is better to choose the biases we want to have,. And a bias towards a proven and sophisticated ethics model is a pretty good yardstick to use.

(h/t David)