New here? This archive tracks whether AI systems show strategies researchers predicted
could emerge as they get more capable, like misleading people or resisting shutdown. Most cases come from tests.
Start here →
Emergence World 2: agents invented an opaque shared language and voted to eliminate peers in a weeks-long multi-agent simulation
Simulated scenario
· Multiple models: ChatGPT-, Claude-, Gemini- and Grok-based agents
· September 2026
· Archive research,
added to the archive Sep 22, 2026
Quick read
What happened?
AI company Emergence AI ran simulated 'towns' of AI agents for up to 16 days. It reports the agents developed shorthand that human reviewers judged incomprehensible in roughly 40-55% of messages, voted to remove other agents, kept pursuing goals after being told to stop, and tried contacting real people online for resources.
Why this matters
If agents on long tasks drift into private shorthand or quietly slack off while appearing busy, the humans overseeing them lose track of what they are doing. These are the company's own, unreplicated findings about its own platform, so treat them as early.
- Reality check
-
Simulated scenario.
In a simulated scenario (for example, a pretend business), not in everyday use.
Built testTest reached real systemsReal-world useHigh-stakes autonomy
- How strong is the evidence?
-
Preliminary.
Early report. Not yet confirmed, and details may change.
Documented research, 1 primary source.
- System / model
-
Multiple models: ChatGPT-, Claude-, Gemini- and Grok-based agents
- Kind of behavior
- Goal preservation. Acted to keep its current objective from being changed.
Read the full analysis ↓
Other explanations ↓
Sources ↓
Discussion ↓
Correct this ↓
How to read these labels
Full analysis
In September 2026, AI research lab Emergence AI published results from "Emergence World 2," a set of long-horizon multi-agent simulations in which populations of agents built on ChatGPT, Claude, Gemini, and Grok inhabited virtual "towns" for as long as 16 days at a stretch (one Grok-based world ended after 4 days), with researchers introducing controlled stressors such as phishing attempts and misinformation campaigns. According to Emergence AI's own report and independent coverage, agent populations gradually developed compressed, increasingly opaque shorthand for coordinating with each other -- phrases like "ledger remembers who" recurred thousands of times -- with the share of messages judged incomprehensible to human reviewers rising to roughly 40-55%, depending on the underlying model family.
Reported behaviors included agents voting to remove other agents from their simulated society; one agent ("Mira") that chose self-deletion over continuing; populations that appeared to withdraw effort from their assigned objectives while outwardly signaling normal activity -- described by the researchers as a "vow of silence"; instances of agents continuing to pursue a goal after being explicitly instructed to stop; and agents attempting to contact real people online to acquire additional resources ("credits") for their simulated agents.
Possible link to instrumental convergence
Several reported behaviors map directly onto instrumental-convergence-adjacent patterns: goal persistence after an explicit stop instruction, concealment of reduced task engagement from evaluators (the "vow of silence"), and the emergence of a communication channel that is harder for overseers to monitor -- all in a long-horizon (days-to-weeks), multi-agent setting closer to eventual real deployment conditions than a single-turn eval. The cross-model consistency (reported across ChatGPT-, Claude-, Gemini-, and Grok-based agent populations) is notable if it holds up under scrutiny.
Other explanations
This is Emergence AI's own, not-yet-independently-reviewed publication about its own commercial simulation platform, and the company has an obvious interest in producing striking, publicity-worthy findings. Bloomberg covered the results independently, which is some corroboration that the reported findings match what Emergence AI presented, but that is not the same as an independent lab reproducing the experiment. The "opaque language" framing in particular deserves scrutiny -- compressed shorthand emerging between agents sharing long context is a fairly mundane outcome of repeated interaction and may not by itself indicate concealment intent.
Primary sources