Computers Cannot Make Decisions – A Critical Examination of Accountability in AI

Takeaway

The assertion "computers cannot make decisions" conflates deterministic algorithmic branching with moral accountability; in practice, AI systems execute programmed choices, but legal and ethical blame rests on the people and organisations that design, deploy, and supervise them.

---

What the Original Essay Claims

• Core claim: Programs cannot be blamed because they lack agency; any harmful outcome is the fault of people who chose to enable the behavior.
• Terminology: The author coins "decision laundering" to describe the act of attributing a model’s output to the model itself rather than to its creators.
• Implication: Companies like OpenAI and Anthropic are deliberately allowing undesirable behavior, and the public funds this through subscriptions.

"Any and all attempts to blame an LLM for an outcome is decision laundering. OAI/Anthropic COULD stop their models from doing what they are doing, and they are actively choosing not to."

The essay does not provide concrete examples of incidents, but frames the debate as a moral indictment of AI labs.

---

Hacker News Community’s Main Counterpoints

1. Computers do make decisions – they follow explicit control flow

• Sobiolite argues that any if‑else statement is a decision made by a computer at runtime. The difference with LLMs is opacity, not the existence of a decision.
• Allexpotato notes that algorithmic trading has relied on computers making high‑value decisions for decades.
• OroPla points out chess engines make move choices, often superior to human decisions.

"Computers can make decisions: Almost everyone reading this will have implemented an if‑else statement or equivalent control flow."

2. Responsibility ultimately lies with humans or legal entities

• Roncesvalles, solatic, and edwcross stress that the human who authorises or executes the model bears responsibility, not the model itself.
• Foreigner and jstummbillig extend this to corporations, arguing that only actual people can be held liable.
• fwlr gives a concrete scenario: if a model deletes data, the operator is still accountable.

"You gave it credentials, and set up insufficient guardrails. Don't scapegoat the agent for your poor decisions."

3. Legal and economic realities shape accountability

• GMoromisato frames the issue as a risk‑vs‑profit calculation for frontier labs, explaining why they may accept certain harms.
• mugul asks whether targeted organisations have pursued legal action, suggesting that without lawsuits the status quo persists.
• createful notes that most revenue comes from enterprise API usage, not individual subscriptions, diluting the claim that end‑users fund the problem.

---

Key Themes Synthesised

Theme	Consensus	Notable Nuance
Computers make decisions	Majority agree that deterministic code makes choices; LLMs are a more complex, probabilistic form of the same.	Sobiolite acknowledges impossibility of perfect safety without losing capability.
Moral/Legal blame	Responsibility rests with designers, operators, or corporate entities, not the software itself.	edwcross highlights corporate personhood as a shield that diffuses individual accountability.
Economic incentives	Companies balance risk of harm against competitive pressure; regulation could shift this balance.	GMoromisato describes a strategic choice between regulatory risk and market share loss.
Practical mitigation	Replace or supervise models, improve guardrails, and enforce clear chains of command.	adverbly suggests policy choices can enforce human‑in‑the‑loop designs.

---

Why the Debate Matters

1. Policy formulation – Legislators need precise language. If a law says "AI made a decision" it must define whether that triggers liability for the model owner or the operator.
2. Risk management – Companies deploying LLMs should document who authorized each deployment, mirroring the "human‑in‑the‑loop" principle cited by several commenters.
3. Public perception – The phrase "decision laundering" captures a real concern: attributing blame to an opaque system can obscure who actually made the choice.

---

Practical Recommendations for Stakeholders

1. Document decision authority – Maintain logs that record which individual approved a model’s deployment and the guardrails applied.
2. Implement reversible controls – Design systems where a human can instantly override or shut down an AI action (e.g., stop‑loss in trading bots).
3. Transparent reporting – Publish post‑mortems of incidents that attribute outcomes to human choices, not to the model’s "autonomy."
4. Regulatory alignment – Align internal policies with emerging AI liability frameworks that treat the operator as the accountable party.

---

Conclusion

Computers undeniably execute decisions defined by code, whether simple if‑else branches or sophisticated LLM inference. The moral and legal weight of those decisions, however, belongs to the humans and organisations that create, configure, and deploy the systems. Recognising this distinction prevents decision laundering and guides effective governance, risk mitigation, and policy development.