Hacker News Post Calls for Regulation of Paid LLM Services
A Hacker News user argues that AI regulation debate has been dominated by doomerism and AGI while paid LLM services face little scrutiny over token metering, model downgrades, output tampering, quotas and billing for failed responses. The post also questions whether AI companies can be trusted after IPO and compares their trajectory to Uber and Lyft.
The author writes that real regulation should begin with fair accounting of input and output tokens. The post compares paying for a dollar of LLM compute to buying a gallon of gas and asks where the weights and measures department is to enforce fair counting. It asks whether a provider could play games with input or, especially, output tokens to devalue a customer’s dollar.
The post’s second concern is quality. According to the author, when a user requests a higher effort level from a frontier model, the service may downgrade the request to a lower-tier model or a different effort level, and the user may not know. The author argues that this amounts to selling a service the customer did not ask for, and that if providers consider downgrades a risk and frequently get them wrong, they should discount the cost of the downgraded response. The post dismisses terms of service as an adequate answer.
Third, the author says LLMs are routinely “adulterating” output for tracking and accountability purposes but asks who validates that these efforts do not lower response quality. The post says the author has noticed words being swapped in ways that made no sense, and that the changes stood out even without looking for them. Fourth, the post asks whether responses themselves could be seeded so that the author of a response can be identified, comparing the possibility to yellow printer dots used to deanonymize printed documents.
The fifth concern is quotas. On a metered service, the author asks what happens to requests that fail mid-response or responses that are cut midway, and why a customer should pay for a failed response. The post asks who validates that providers are not doing this at scale to save money.
The author argues that engaging in what the post calls “nonsense regulations” proposed by the major LLM-as-a-service duopolists distracts from real discussion about regulating the product customers pay for. The post says there is no reason to believe these companies will operate in good faith after an IPO, and asks where the incentive will be once competition is shut down to ensure that only OpenAI and Anthropic are available to purchase from.
As a comparison, the post points to Uber and Lyft, saying they have degraded service and increased costs over the years. It says they skirted regulation and then demanded that the world bend and write regulations fitting their revenue goals, or threatened to pull out where they could not. The post concludes by asking how the same companies can be trusted to regulate a threat they describe as severe if they refuse regulations around basic delivery of service.