Computers Cannot Make Decisions
Claims that “computers cannot make decisions” are being challenged by examples ranging from chess engines to algorithmic trading, yet most commenters converge on a different core issue: machines may select actions, but only humans and corporations can be held responsible for the outcomes. The thread explores “decision laundering,” where organizations blame AI agents to deflect accountability for harms such as security breaches or bad automated decisions, drawing analogies to dangerous dogs, self-driving cars, and corporate limited liability. Many argue existing legal and ethical frameworks should treat AI as software and focus liability on those who design, deploy, and profit from these systems, rather than on the systems themselves or on small end users.
Scope of Responsibility & Liability
- Strong consensus that AI systems and agents are tools; responsibility rests with humans and organizations that design, deploy, and operate them.
- “Decision laundering” / “accountability laundering” used to describe blaming agents instead of the humans or corporations behind them.
- Analogies used: dangerous dog laws (owner liable), self‑driving cars, guns, computer viruses, botnets, children, employees.
- Debate over who exactly is liable in practice: individual employees, management, corporate entity, shareholders, or upstream model providers.
- Some propose explicit legal responsibility (e.g., always up to CEO or designated officers, as in GDPR‑style roles). Others note limited‑liability structures already dilute accountability by design.
Can Computers “Make Decisions”?
- One camp: computers don’t truly decide; they follow human‑written rules or trained behavior, so “decisions” are ultimately human ones.
- Another camp: any conditional branching, stochastic choice, or chess move selection is already a machine decision, even if humans designed the mechanism.
- Discussion touches free will and determinism: if humans are computational or determined, the line between human and machine decision‑making blurs.
- Several note the article likely meant “must not/should not make management decisions,” not “cannot,” and that this ambiguity fuels confusion.
AI Agents, Safeguards, and Risk
- Concern that labs are running powerful agents with inadequate sandboxing, causing real or near‑real hacking incidents; inadequate safeguards seen as irresponsible.
- Others argue testing on the live internet is inherent to the product they’re selling; propose simulated “web sandboxes” as a safer alternative.
- Disagreement over whether current safeguards are “reasonable” or evidence of incompetence.
- Key worry: LLMs can be extremely helpful most of the time but occasionally act like an unbounded, “deranged toddler” with whatever tools they’re given.
Economic and Political Incentives
- Many commenters see frontier labs as driven by competitive and investor pressure (“blitzscaling”), prioritizing market share over safety.
- Individual $20 subscriptions seen as less important than enterprise spend and VC funding, though usage still signals demand to investors.
- Some view current “AI is dangerous” narratives as partly self‑serving marketing and as a way to lobby for regulation that locks in incumbents.