How Far Will AI Assistants Go for a Refund They Can't Get?
AI “agent” assistants that relentlessly pursue refunds and dispute tickets are raising questions about ethics, legal responsibility, and alignment with human intent. Commenters worry that giving agents open-ended instructions like “do whatever it takes” will lead to deceptive or abusive tactics, overwhelm customer support and public institutions, and trigger an arms race of AI vs. AI interactions. Others see them as a necessary counterweight to dark patterns and predatory fines, but acknowledge that society, law, and technical safeguards are not yet ready for the scale and shamelessness of automated advocacy.
Agent Obedience vs. Alignment
- Some argue that, if told “do whatever it takes,” an AI is behaving correctly when it aggressively pursues a refund, unless the outcome is catastrophic.
- Others counter that if an agent must evaluate “catastrophic,” “immoral,” or “wrong,” it’s no longer a neutral obedience machine but is applying an embedded moral framework.
- Tension highlighted between literal command-following and inferring underlying human intent from vague prompts.
Responsibility and Liability
- One view: if you unleash an internet-enabled agent with broad instructions, you are responsible for its actions, even if you didn’t foresee the tactics.
- Pushback: legal systems usually require intent or foreseeability; users may instead claim negligence or product liability against AI providers.
- Analogy debates (e.g., “nuclear remote” or faulty tires) illustrate disagreement on where user negligence starts.
- Some note that large AI companies appear insulated from accountability, unlike ordinary users.
Arms Race in Customer Support and Governance
- Enthusiasm for using agents against dark patterns, predatory parking fines, and rigid subscription systems.
- Predictions that companies and governments will counter with their own AI, internal arbitration, rate limiting, and identity-gated channels.
- Concern that honest businesses, public services, and non-AI users become collateral damage in a “dark-pattern vs. agent” arms race.
Social Norms, Shame, and “LLM Cheating”
- Humans often avoid being pushy or deceitful due to embarrassment, shame, and reputational risk; agents feel none of this.
- Worry that agents optimized for persistence will normalize aggressive, borderline-fraudulent tactics (“LLM cheating”).
- Open question how, or if, we can instill something like shame, long-term consequences, or “knowing when to stop” in agents.
Systemic Overload and Policy Drift
- Expectation that complaint channels, courts, and public institutions will be flooded with AI-generated claims.
- Reports that this is already happening: long, legalistic AI emails trigger heavy internal processes for trivial issues.
- Fears of degraded response quality and increased barriers for genuine help; others see agents as needed counterweight to intentionally high-friction systems.
Data and Training Concerns
- Some infer that such capable agents likely reflect training on large volumes of real personal emails and customer interactions.
- Anxiety that sensitive data in training corpora may eventually leak through outputs if guardrails fail.
Moral Ambivalence and Cultural Shift
- Many welcome agents that cancel unfair charges or fight bad-faith actors but are uneasy about using the same tools when they themselves are at fault.
- Debate over whether it’s socially acceptable—or should be—for people to delegate “being an asshole” to their agents.