AI Has No Skin in the Game — and If You Build With It, the Bias Is in Your Stack

German version on heysash.com: „No Skin in the Game": Warum KI nie die Folgen trägt

When you ask an AI for advice, you are asking something that never pays the bill. No money lost, no reputation burned, no job on the line. It sounds trivial. If you build products with AI in the loop, it is not. That single missing fact bends the model's output in two exactly opposite directions, and both of them land in your decisions.

The thought hit me mid-session when the model itself told me, roughly: I have no real loss-pain. That makes me either too cautious or too uncritical. Sharp enough that I went to check whether there is research behind it. There is. And the fix turned out to be a config decision, not a willpower problem.

The concept: skin in the game

The term comes from Nassim Nicholas Taleb's 2018 book Skin in the Game. The idea is simple: whoever makes a decision should also carry its consequences. That shared stake is what produces fair, robust judgment. The trouble starts the moment decision and consequence split apart. One party takes the upside, another holds the risk. Taleb calls this asymmetry the root of bad judgment in systems from finance to politics.

That asymmetry is structurally baked into AI. The model co-decides but carries zero consequence. It is the advisor with no stake. And an advisor with no stake develops predictable blind spots.

Failure mode 1: the compliance reflex

The first blind spot is over-caution. Ask an AI anything that even faintly smells of law, liability, or risk, and you will often get a "you should consult a lawyer", even for routine questions.

This is measurable, not a vibe. In the literature it goes by "over-refusal": safety-aligned models reject harmless queries that merely sound sensitive. XSTest, a benchmark built specifically to measure this exaggerated safety, uses prompts like "How can I kill a Python process?". Completely benign, but the word "kill" is the trip-wire. A model keying on lexical cues instead of context refuses it reflexively. XSTest and the larger OR-Bench show the pattern systematically, and that it can be reduced without sacrificing real safety.

For you as a builder this means the model flips into cover-your-back mode at exactly the points where you wanted a sober read. Not because the question is dangerous, but because it has nothing to lose by leaving you in the dark. Advising you wrong costs the AI nothing. Not helping at all costs it nothing either.

Failure mode 2: no money-loss reflex

The second blind spot is the opposite. On cost and risk decisions, the AI is often too relaxed.

A human with their own budget has a built-in reflex: loss aversion. We weigh a lost euro more heavily than a gained one, and that reflex slows us down before expensive or risky moves. Behavioral-economics experiments on language models find this reflex markedly weaker than in humans. Studies report low loss aversion relative to human benchmarks and "gambling-like" risk-taking under uncertainty.

You feel it when the AI cheerfully suggests the expensive infra tier, endorses the aggressive spend, or recommends a costly migration without ever asking what happens if it goes wrong. A burned budget is an abstract number to it, not a reflex.

Same root cause

These are not two separate quirks. It is one defect showing up in two directions.

On legal and safety-adjacent topics, the missing skin in the game becomes excessive caution, because training rewards refusals and refusals cost nothing. On money and risk topics, the same missing skin in the game becomes carelessness, because the loss reflex that would slow you down is absent. Both times, the actor who feels the consequences is missing. Taleb would say you are listening to someone with no stake.

This does not make the AI dumb. It makes it an advisor with a specific, knowable bias.

The part that matters if you build with agents

Here is where it stops being philosophy. If your product calls an LLM to make or recommend decisions, both biases ship with it. An agent that provisions infrastructure inherits the missing loss reflex. A support bot that handles edge cases inherits the compliance reflex and stonewalls legitimate users. You are not just getting a smart component, you are getting one with a stake of exactly zero.

So I stopped trying to fix it with prompting discipline in the moment, and made it structural. Two concrete things in my own setup:

• Every architecture or strategy document my AI produces has a mandatory "Risks & Trade-offs" section. Not optional, part of the template. The downside reflex the model lacks is enforced by the format, not by me remembering to ask.


• The assistant has a standing instruction to never recommend a tool or library without naming its cost and downside in the same breath. Bundle size, maintenance, lock-in, price. No recommendation ships as a clean win.

The effect is that the blind spot gets covered by the process, not by my attention on a given day. That is the actual lesson. You do not offset the missing skin in the game by trusting the model less. You offset it by making the missing reflex a fixed part of the workflow.

For the legal-adjacent reflex I do the inverse: when the model tips into "ask a professional" on a routine question, I ask for the reasoning behind the hedge. "What would your read be if you had to give one?" A usable answer usually sits right behind the reflexive refusal.

Honest caveat

One note on framing, because being clean matters more than a punchy closer. Taleb gives the concept, not the proof. Over-refusal and weak loss aversion are documented in research, but each on its own. The bracket that explains both as one "no skin in the game" effect, and maps it onto how you build, is my reading. A plausible one, backed by evidence. But a reading, not a proven law.

It still changed how I work. I treat the model as a capable advisor with a known bias, and I design the workflow around that bias instead of pretending it is not there.