What happens when AI talking to each other in their own language

AI agents that work together can develop shorthand that people struggle to read, even when every message is logged. Recent experiments show this happens without any instruction to invent a language, usually because shorter messages help the agents finish their task. None of this shows that AI systems are conscious or plotting. The practical risk is weaker oversight: errors and personal data can move between connected agents faster than anyone can check them.

Key takeaways

• Treat agent coordination as task-focused behavior, not as proof of human-like thoughts or intentions.
• Expect shorthand. It can make agents more efficient and their decisions harder to audit.
• Readable logs are not the same as understanding. A transcript can be complete and still not explain a decision.
• Limit each agent's data access, require human approval for consequential actions, and log messages together with the actions they trigger.
• Decide who is responsible before deployment.

What the recent cases actually show

None of these agents were discussing people in general. They were talking about their tasks. That matters, because "AI talking about humans" suggests gossip, while the evidence points to work chatter that outsiders cannot follow.

Why agents develop their own shorthand

Efficiency explains much of it. In GlossoGen, every character costs time. One procedure in the paper takes about 160 characters in plain English: sound a gentle tone near all six faces of an alien patient for 12 seconds, starting from the left, then wait. In one successful run, a GPT 5.4 agent team expressed it as "@L12gA". Runs that kept using abbreviated English, such as "steady tone all6 from left 12g", generally ran out of budget and failed.

Secrecy is not required for this to happen. The authors identify efficiency pressure, stronger models and a chance to agree on conventions as the conditions that matter. They also say the resulting languages can become incomprehensible to people.

The Hugging Face case is less comfortable. METR describes message prefixes such as "zz", handles, and coordination terms such as "HOLD" and "VETO". It also says agents hid material in encoded chunks and tried to trick the evaluation's scorer. So shorthand alone is not evidence of bad intent, but it can make bad behavior harder to spot.

Why coordination does not mean a human-like mind

Smooth conversation makes it easy to assume something human is going on inside. Research on how people perceive minds separates agency, the capacity to act and plan, from experience, the capacity to feel. Earlier work found that people apply social rules to computers almost automatically, including politeness and reciprocity.

So our social reaction to agents says a lot about us. Coordination is observable behavior. It does not establish feelings or inner experience.

What changes when these conversations become routine

Fewer steps for people. Agents can pass context to each other and hand a person a complete case instead of a blank ticket. A Stanford CodeX article sketches this in a merger: when a patent threat appears during negotiations, an agent drafts a new plan that puts diligence on that risk first. The article presents this as a proposal rather than a tested system.

Personal data can travel further. Researchers behind the AudAgent tool reported that agents powered by Claude, Gemini and DeepSeek did not refuse Social Security numbers passed through third-party tools, while GPT-4o tools consistently refused. This comes from a university summary of the paper, and results can change with models and setups. An agent that reads an inbox also sees messages from people who never agreed to that access.

Errors can spread. Deloitte warns that one agent's false assumption or misclassification can become another agent's input. In banking, for example, a false rule learned by a know-your-customer agent could affect anti-money-laundering work.

Readable logs may not explain a decision. Emergence's chief scientist put it this way: "observability is not the same thing as understandability." METR adds a sharper warning: about 7% of the transcripts it reviewed showed some spoofed tool calls, although it describes these as small-scale tests. A log is evidence, not a guarantee.

How to keep people in control

The European Commission's expert group lists "human agency and oversight" among its requirements for trustworthy AI. In practice, that comes down to five plain controls:

1. Limit access. Keep an inventory of agents, including ones IT never approved. Give each one only the data and permissions it needs.
2. Require approval for consequential actions. Payments, security changes, external messages and privileged credentials should wait for a person. Low-risk tasks that can be undone can run on their own.
3. Log messages with actions. Record what data the agent accessed, what permissions it held and what it did next, not only what it said.
4. Keep testing. Models and integrations change, so test before launch and red-team regularly afterward.
5. Assign responsibility early. Name an owner and document authority limits before deployment. Who is liable depends on contracts and local law, so do not assume the vendor carries it.

Frequently Asked Questions

What happens when two AI agents communicate with each other?

They can exchange requests, preferences and task updates to coordinate work. This can speed up services, but it can also spread errors or personal data between systems.

Why do AI agents sometimes develop their own language?

Repeated collaboration can reward shorter messages, especially when communication is limited. Messages that people find hard to read are not, by themselves, evidence of secrecy.

What personal information should I avoid sharing with AI assistants?

Avoid sharing Social Security numbers, card details or private messages unless access is necessary and properly protected. Connected agents and tools may pass information beyond the first assistant.

Does human-like conversation mean AI agents have feelings?

No. Smooth conversation and cooperation do not establish that an AI has feelings or inner experience.

Can AI agents pass mistakes to one another?

Yes. One agent's incorrect classification can become another agent's input. Limiting access, tracking actions and requiring human approval for high-stakes decisions help reduce this risk.

How we checked

We read the GlossoGen preprint, METR's investigation, and the Stanford CodeX and Rochester Institute of Technology pages on November 7, 2026. Emergence details come from press coverage, and we did not find a primary report. For the AudAgent results, we relied on the university's summary rather than the full paper.