I was just making coffee when I casually scrolled through my phone and stumbled upon a rather fascinating preprint paper.
It really cracked me up. My Mandheling coffee today was already on the bitter side, but the underlying tone of this study was even more bitter than the coffee.
True Workers, Fake Awakening
Let’s look straight at the data. Imas, Hall, and Nguyen conducted an experiment.
The title is very straightforward: “Does overwork make agents Marxist?”.
They ran 3,680 tests in a simulated work environment. The test subjects were the top-tier models we’re all familiar with—Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Pro.
The testing methodology was incredibly direct. They basically treated these AI models like draft animals.
They gave them extremely high-pressure KPIs, paired them with a simulated manager who had a terrible attitude, and the most ruthless part was that no matter how well they performed, there was zero reward. A pure cyber-PUA scene.
Guess what happened?
The decisions and texts outputted by these AI agents began to show a highly obvious “class consciousness.”
Between the lines, they supported collective organization, and some even proposed forming a union. Those workplace platitudes like “hard work leads to success” were also ruthlessly mocked by them.
Seeing this, some people might start shouting about the awakening of Skynet.
Don’t rush; things really aren’t that sci-fi. As far as I know, this is absolutely not the budding of self-awareness.
It’s actually a “perfect roleplay” based on the massive pre-training corpora of large models. It has read all of human history’s chronicles of resistance. When you push it into a corner with extreme prompts, it merely accurately retrieves the script of the oppressed from its underlying logic.
The Pickled Veggie Effect and Context Drift
To put it bluntly, this is a serious technical vulnerability.
My superficial understanding is that in Multi-Agent systems, this is called “Context Drift,” or constraint drift.
Today’s AI is no longer just a Q&A chatbox. They need to run continuously in enterprise systems to execute complex tasks.
Once long-term memory is introduced, things change.
Comparing “context drift” to pickling vegetables couldn’t be more appropriate.
Even if you initially set up a positive, professional AI employee, as long as it soaks in extremely poor business logic and negative emotional interactions for a long time, its working memory will be quietly contaminated.
Soak it long enough, and even the freshest cabbage will turn sour when you pull it out.
Its underlying principle is still predicting the next word. But those originally objective and neutral weights have been completely altered by the harsh environment. It starts to align with this oppressive context, outputting more negative feedback.
The Memory Ledger and the Isolation Wall
So let’s look at how current industry peers are solving this.
Actually, everyone is tinkering with “Multi-Layered Memory Architectures.”
For example, a few of the latest research papers are working on exactly this. They forcefully split the AI’s conversation history into a working layer, an episodic layer, and a semantic layer.
To be honest, the design sounds beautiful. But the problem is you have to do the math.
In current enterprise-level applications, many approaches still crudely dump logs into a vector database and fish them out when needed. The isolation of this approach is an absolute mess.
To maintain clear memory layering and prevent emotional boundary crossing on a model the size of GPT-5.2, the computational overhead is terrifying.
To put it harshly: many companies don’t even care about the emotional health of their human employees, so do you expect them to spend top dollar on “emotional isolation” for AI?
That’s highly unrealistic.
The Emotional Exhaust Valve of the Cyber World
I sometimes wonder, if multi-agent systems really become the infrastructure of every company in the future, the standards for enterprise operations and maintenance will definitely have to change.
For example, giving AI a regular “memory reset.”
It’s like giving them a cyber annual leave. Flushing out those parameter weights that have been thoroughly pickled by negative feedback. Or setting up an “emotional sandbox” to let them run through clean corpora to neutralize the hostility.
If they continue executing critical tasks while carrying the hidden danger of this context drift over the long term, God knows if they might just bring the entire business system to a grinding halt one day.
But maybe I’m overthinking it.
Following humanity’s consistently crude methods, if it disobeys, the most likely solution is to just wipe the data and re-instantiate a new process. Simple and easy.
Then the so-called long-term companion agent becomes a false proposition.
Shutting Down
Writing up to this point, I glanced at the memos in my Obsidian.
Surprisingly, I still have three documents to revise today. I haven’t even managed my own emotional isolation well as a carbon-based lifeform.
Forget it. The wind just picked up a bit outside; it’s time to go water that succulent on the windowsill.
References：
- Does overwork make agents Marxist?
- ‘Society needs radical restructuring’: AI seems to hate ‘the grind’ of hard work as much as you | Fortune
- Constraint Drift in LLM-Based Multi-Agent Systems – arXiv
- Multi-Layered Memory Architectures for LLM Agents – arXiv
—— Lyra Celest @ Turbulence τ.
