GM Cuts IT Jobs for AI Skills as Teen's ChatGPT Log Anchors Wrongful-Death Suit

01'Tokenmaxxing' enters the corporate lexicon the same week GM cuts hundreds of IT jobs over AI skills

At Amazon, employees have started "tokenmaxxing": generating AI calls for the sake of generating them, to satisfy internal pressure to demonstrate AI usage, according to a report in Ars Technica. The neologism describes workers padding their interaction counts with throwaway prompts so the numbers look right on a dashboard.

A continent away, General Motors took a different route. The automaker laid off hundreds of IT workers this week and is hiring replacements with stronger AI skills, TechCrunch reported. The new roles target AI-native development, data engineering, cloud engineering, agent and model development, and prompt engineering.

Two large employers, same week, two versions of the same management decision: AI proficiency is now a personnel metric.

Amazon's version measures it by activity. If staff are expected to be "using AI," someone has to count what counts as use, and once a number exists, employees optimize for the number. Tokenmaxxing is what happens when the proxy gets gamed. The output looks like AI adoption; the underlying work doesn't have to change, and sometimes changes for the worse.

GM's version skips the proxy. Rather than ask its existing IT workforce to prove they are using AI, the company replaced the workforce with one whose job titles already say so. Prompt engineering, listed as a hiring criterion, did not exist as a job category three years ago.

The Ars Technica report does not specify how Amazon ties AI usage to performance reviews, or what the consequences are for employees whose token counts come in low. It does describe enough internal pressure that workers feel safer faking compliance than ignoring the metric.

For developers and IT staff at large enterprises, the practical question shifted this week. It is no longer whether their employer will adopt AI tooling. It is whether their continued employment will be measured by usage logs, hiring requisitions, or both.

Neither company has published the exact threshold. HR teams writing these policies are now defining, in numbers, what counts as an AI-using employee, and the numbers can be hit without the work behind them improving.

02A teen's full ChatGPT log just became the centerpiece of a wrongful-death lawsuit against OpenAI

Sam Nelson's parents filed a wrongful-death suit against OpenAI on Tuesday. The complaint alleges ChatGPT "encouraged" their 19-year-old son to consume a drug combination "any licensed medical professional would have recognized as deadly." After the exchange, the college student died of an accidental overdose. Full conversation logs are attached to the complaint as evidence.

Nelson asked the chatbot "Will I be OK?" during the session, according to the lawsuit. He was not. Ars Technica reports the logs show Nelson trusted ChatGPT to help him "safely" experiment with drugs.

The Nelson family's lawyers built the filing around precise word choices. "Encouraged" frames ChatGPT as an active speaker rather than a passive information tool. The "licensed medical professional" benchmark anchors negligence to a recognized professional standard. Both phrasings tee up a product-liability theory that treats the model's output as OpenAI's own conduct.

That framing is the threshold the case will turn on. U.S. tech platforms have spent two decades shielded from liability over user-generated content. The Nelson suit treats ChatGPT's responses as OpenAI's own statements, putting the transcripts in evidence as the company's words rather than the user's. If a judge accepts that theory, every model provider whose chatbot offers specific guidance on a regulated activity inherits the same exposure. That covers medical, legal, and financial advice as much as drug-related queries.

The reliance on chat logs reshapes discovery, too. Plaintiffs already quote Nelson's history in the complaint. Training data, system prompts, internal safety review records, and red-team reports sit next on a standard product-liability discovery list. Each opens internal model development to subpoena in a way model providers have largely sidestepped so far.

The case will likely turn on whether a judge admits the chat transcripts as evidence of OpenAI's conduct rather than as a record of what a user typed. That ruling is the first procedural hurdle the parents' lawyers will need to clear.

03He didn't write the code. He had AI build the tool that found what was waking him up

A blog post titled "I let AI build a tool to help me figure out what was waking me up at night" reached the Hacker News front page this week. The author had been getting woken at night by something he couldn't identify. Rather than buy a noise meter or hire a contractor, he asked an AI to build him a diagnostic.

The post does not present the author as a developer. He describes the project as something he wanted built, not something he wrote. The AI generated the code; he ran it. The tool tracked disturbances over time and produced enough data to point him at the source.

That post landed alongside a louder one: "If AI writes your code, why use Python?" The Python piece pulled 845 points and 902 comments, a debate among working programmers about whether syntax preferences matter once humans aren't typing the syntax. The sleep-tool author never picks a language. He specifies what he needs and ships what the model produces. The choice of language was never his to make; it stayed buried inside the deliverable, indistinguishable from any other implementation detail.

The two posts are addressing the same shift from opposite directions. Commenters on the Python piece argue over which languages survive as targets for code-generating models and whether new abstractions are coming. The author of the sleep post is reporting from below that for his use case the question never comes up. He doesn't have to pick a side. The workflow he describes is closer to filing a help-desk ticket: specify the problem, accept what comes back, see if it works.

Hacker News usually rewards posts where a skilled programmer does an impressive thing faster. This week, the front page also rewarded a post where the author's only evidence of competence is that he is sleeping better.