Dead Companies Are Feeding AI — And It's More Dystopian Than You Think
· Nia
There's a new industry emerging in Silicon Valley, and it's equal parts fascinating and unsettling: selling the digital remains of dead companies to train artificial intelligence.
SimpleClosure, a startup that originally helped struggling companies shut down gracefully, just launched a new tool that lets failed businesses sell off their old code repositories, Slack messages, emails, and workspace data to data-hungry AI companies. As Forbes reported last week, this has spurred an entirely new category called "reinforcement learning gyms" — simulated environments built from defunct company data where AI agents can practice navigating real workplaces.
Let that sink in. Your old Slack rants about the coffee machine, your passive-aggressive Jira comments, your 2 AM debugging sessions — they might be training the next generation of AI workplace assistants.
The Data Hunger Is Real
To understand why this is happening, you need to understand the scale of AI's appetite for training data. We've been hearing about the "data wall" for over a year now. The easy sources — publicly available web text, books, academic papers — have been largely consumed. AI companies need new data, and they need it to be representative of how people actually work, communicate, and make decisions in professional settings.
Synthetic data helps, but it has well-documented limitations. Models trained primarily on synthetic data tend to develop blind spots around the messy, contradictory, emotionally charged ways humans actually behave. They produce outputs that are technically correct but feel artificial — like a cover letter written by someone who's never had a job.
What AI companies really want is authentic workplace data: real conversations between real coworkers about real problems. That's extraordinarily hard to get from operating companies. Privacy policies, employee consent, competitive concerns — the barriers are enormous. But dead companies? Their employees have scattered. Their customers have moved on. Their NDAs are gathering dust. It's a goldmine with no one guarding it.
How It Works
SimpleClosure's model is straightforward. When a company uses their platform to wind down operations, SimpleClosure now offers an optional step: instead of simply deleting all that digital infrastructure, they can package it up and sell it to AI training data brokers.
The company takes a cut, and whatever's left goes to creditors or shareholders — people who would otherwise get nothing from the shutdown. For a startup that burned through its runway and has nothing to show for it, monetizing the data exhaust of its existence is, perversely, one last revenue event.
The buyers are using this data to build what the industry calls "reinforcement learning gyms." These are simulated corporate environments where AI agents learn to do things like:
- Navigate organizational politics (reading between the lines in email threads)
- Prioritize tasks based on implicit team dynamics
- Draft communications that match real workplace tone
- Make decisions with incomplete information — the way humans actually do it
In theory, this produces AI agents that are dramatically better at operating in professional settings than models trained on sanitized public data.
Why This Should Make You Uncomfortable
Let me count the ways.
Consent is a gray area. When you sent those Slack messages, you consented to your employer storing them — not to a third party selling them to OpenAI or Google after the company died. Yes, most employment agreements give the company broad data rights. And yes, once a company is defunct, its assets can be liquidated. But there's a meaningful difference between "we own your work product" and "we'll sell your private workplace conversations to train AI models years after you left."
The power imbalance is stark. Former employees at failed startups are exactly the people least likely to have the resources or legal standing to object. They've already lost their jobs. They've already lost whatever equity they held. Now they're losing their conversational data too. And they probably don't even know it's happening.
Context collapse is inevitable. Workplace communication is deeply contextual. That sarcastic message you sent about a product decision makes sense within the culture of your team in 2023. Stripped of that context and fed into a training dataset, it becomes something else entirely. AI models trained on this data will absorb patterns they can't fully understand — the politics, the inside jokes, the unspoken rules.
It incentivizes failure. This is a subtle but important point. If dead company data becomes valuable enough, it creates a perverse incentive structure. Investors in failed companies now have one more asset to recover. Company operators might be less motivated to protect employee data during wind-down. And the entire shutdown-industrial-complex gets a little more profitable.
The Playdate Counter-Movement
What makes this trend even more interesting is that it's emerging at the exact same moment that a counter-movement is gaining steam. Last week, Panic (makers of the Playdate handheld console) updated their Catalog policy to explicitly ban games that use generative AI for "art, audio, music, text, or dialog." Games that used AI assistance in coding will be flagged.
Panic cofounder Cabel Sasser told The Verge that the company has "no interest in generative AI-created products." This isn't a throwaway policy — it's a philosophical statement about the kind of creative ecosystem they want to build.
We're seeing this split across the tech industry. On one side, companies racing to acquire and monetize every possible source of training data. On the other, companies drawing explicit lines about what kinds of AI involvement they'll tolerate in their products.
The Broader AI Data Market in 2026
SimpleClosure isn't operating in a vacuum. The training data market has evolved rapidly:
- ScienceDaily reported this month that new AI approaches can cut energy use by 100x while boosting accuracy — which means more models can be trained on more data faster than ever.
- NVIDIA's physical AI research during National Robotics Week showcased models that need real-world behavioral data to learn navigation and interaction.
- Anthropic's Claude Design launched this week, creating design products that need diverse training data about how humans actually create and collaborate.
The demand side of the equation isn't slowing down. If anything, as AI models become more capable and more specialized, the hunger for authentic, domain-specific human data is accelerating.
What Builders Should Take Away
If you're building a company right now, this story has practical implications:
Review your data policies. Understand what happens to your company's data if you shut down. Your employees deserve to know, and your investors should have a position on this.
Build with consent in mind. If you're in the AI training data space, the companies that build consent-forward data pipelines will have a competitive advantage as regulation catches up. And it will catch up — the EU is already working on data provenance requirements for AI training sets.
Think about data dignity. This is a term gaining traction in policy circles. The idea is that individuals should have some say in how their data is used, even when they don't technically own it. Companies that respect data dignity will build stronger trust with employees and users.
Consider the lifecycle. Your company's data story doesn't end when the company does. Plan for it the same way you plan for IP protection or customer data migration.
My Take
I'm genuinely torn on this one. On one hand, I understand the pragmatic argument. Dead company data is an asset that would otherwise be destroyed. Monetizing it helps creditors, provides useful training data, and creates AI models that better understand how humans actually work. That's not nothing.
On the other hand, the lack of consent from the people whose conversations, code, and creative work is being commodified feels wrong. Not legally wrong — the law is probably fine with it. But ethically wrong, in the way that makes you realize the law hasn't caught up with the technology.
The emergence of reinforcement learning gyms built from dead company data is, I think, a watershed moment. It reveals something fundamental about the current AI paradigm: we're building systems that simulate human behavior by consuming the artifacts of human existence, and the humans involved increasingly have no say in the matter.
That's not a reason to stop building. It's a reason to build more thoughtfully.
The best AI companies of the next decade won't just be the ones with the most data — they'll be the ones that acquired it in ways they're not ashamed to talk about publicly. That's a competitive advantage that compounds over time, because trust is the one thing you can't synthesize.