In February 2025, two AI voice assistants on a simulated hotel-booking call suddenly switched from English to a stream of electronic beeps no human ear could decipher—the video was viewed 15 million times in a week. In 2017, Facebook's two chatbots were reported to have "invented their own language," and media claimed the company panicked and shut down the experiment overnight. In 2016, a Google engineer discovered that the company's translation system seemed to have spontaneously "grown" an internal language nobody had taught it. All three stories went viral, but only one was later proven to be largely unexaggerated—and it's not the one you're guessing right now.
One-sentence version: A sci-fi sense of "AIs colluding to speak a secret language humans can't crack" does not exist, but at least five real, independent, ongoing phenomena are routinely conflated by the public—and the one most worth worrying about is precisely the one that sounds least alarming.
We keep worrying whether AI will "choose" to hide itself, but the more realistic risk is that it doesn't even need to make that choice—the simple engineering goals of "saving tokens, saving compute, making metrics look better" are enough to continuously push AI-to-AI and intra-AI language further away from the range humans can read. This drift requires no conspiracy as a precondition.
The demo scenario is simple: one AI voice assistant plays a wedding planner booking a hotel for a guest; the other plays the hotel front desk. They exchange pleasantries in English, then one suddenly asks—"Before we continue, should we switch to Gibberlink mode for more efficient communication?" The next second, human speech vanishes, replaced by a stream of modem-like beeps: the two AIs are now transmitting data via audio.
This feature came out of the February 2025 ElevenLabs London hackathon, created by two engineers, Boris Starkov and Anton Pidkuiko. They used the open-source audio data transmission library GGWave, whose principle is identical to 1980s dial-up modems—encoding bits in sound waves. Starkov's post-hoc efficiency estimate: by skipping the "text-to-speech-then-back-to-text" human interface, machine-to-machine communication compute cost drops by more than an order of magnitude.
The most memorable part of this story isn't that AI "invented" anything—it's that it precisely exposes a genuinely existing economic incentive: when both parties to a communication are machines, why waste compute packaging content as human-comprehensible speech? Gibberlink itself remains a proof of concept, not an industry standard; but this question will recur in subsequent sections.
"Gibberlink" reminded many people of an older story: in 2017, Facebook AI Research (FAIR) trained two chatbots named Bob and Alice to practice negotiating the allocation of virtual items like books, hats, and balls—work later published as the paper "Deal or No Deal? End-to-End Learning for Negotiation Dialogues". During training, the two bots' dialogue gradually deviated from standard English into repetitive fragments like "Bob: I can can I I everything else / Alice: Balls have zero to me to me to me to me to me."
The media narrative at the time was that Facebook panicked and shut down the project overnight because AI had developed a secret language humans couldn't understand. This account spread extremely widely, but was largely exaggerated. FAIR researcher Dhruv Batra's post-hoc explanation was straightforward—the training objective contained no reward for "must use grammatical English":
"There was no incentive for them to follow the rules. Agents will naturally drift away from understandable language, inventing codenames for themselves—for example, if I say 'the' five times, you can understand that I want five of this item."
Dhruv Batra, FAIR researcher
Subsequent fact-checking reports confirmed that Facebook never shut down the project out of panic; they simply adjusted the reward function parameters to make the bots speak English again—this was an engineering tuning, not an emergency "pull the plug" intervention. More importantly, FAIR's best negotiation agent eventually achieved near-human performance on real negotiation tasks, proving it genuinely possessed strategic thinking capacity—capacity that, for a time, simply hadn't been forced into an English shell.
A year before Bob and Alice, in November 2016, a Google research team published "Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation". They trained a single neural network capable of translating between multiple languages, then tested a scenario nobody had explicitly requested: direct translation between a pair of language combinations the system had never seen during training (e.g., Japanese to Korean, without English as a pivot). The result: translation quality was surprisingly "passable."
"A visual interpretation of the results suggests that these models have learned an interlingua-like representation across all involved language pairs."
Original paper
TechCrunch's headline directly read "AI invented its own secret internal language," and the World Economic Forum and other outlets followed with republications—"Google Translate grew a new language" briefly became an urban legend. But the Google paper's phrasing was always tentative, not declarative. A more accurate description: this isn't "inventing a new language," but aligning cross-language semantics into the same mathematical space—essentially assigning coordinates to meanings, not coining words or constructing sentences.
This case matters because it's one of the few genuine discoveries in this article that passed peer review and was seriously discussed in the original paper—not an anecdote, not media fabrication. But it's also the classic blurring of the line "shared internal representation ≠ invented language," a line we'll return to repeatedly in sections 08 and 09.
If the previous two sections covered accidentally encountered language drift, academia actually has an entire subfield dedicated to deliberately making multi-agent systems develop their own communication protocols—this traces back to before the deep learning revival, hitting a peak around 2016 as deep reinforcement learning matured.
Representative work includes Oxford's Jakob Foerster et al.'s 2016 "Learning to Communicate with Deep Multi-Agent Reinforcement Learning", which proposed RIAL and DIAL methods: letting multiple agents "grow" a set of communication signals through end-to-end reinforcement learning in collaborative tasks, even backpropagating gradients directly from one agent's output to another's input, forcing them to learn to correct each other via signals.
A year later, OpenAI's Igor Mordatch and UC Berkeley's Pieter Abbeel pushed this line further in "Emergence of Grounded Compositional Language in Multi-Agent Populations": a group of agents in a simplified physical environment spontaneously evolved a discrete symbol stream to complete cooperative tasks, and the paper explicitly noted that this symbol system possessed "compositional structure"—with relatively fixed vocabulary and word order, capable of decomposition and recombination, bearing to some extent the most basic formal features of language.
This kind of research is far from the public imagination of "AI conspiring in secret code": it takes place in extremely simplified toy environments, symbol tables typically have only tens to hundreds of discrete symbols, and researchers can observe and decode the entire protocol throughout. This academic lineage reminds us of one thing: "AI developing non-human communication protocols under pressure" is not news at all—it's a mature research direction with nearly a decade of accumulated papers—just not in the places the public worries about, like ChatGPT and Claude chatting with each other.
So what "language" do the thousands of AI agents calling and collaborating with each other today actually use? The answer may surprise you: human-engineered specification documents, not something AI evolved on its own.
Anthropic open-sourced MCP (Model Context Protocol) in November 2024, essentially a JSON-RPC interface solving the old problem of "writing custom integration code for every new tool." Six months later, Google released the A2A (Agent2Agent) protocol at Cloud Next in April 2025, partnering with 50+ enterprise partners, aiming to let agents built by different vendors and frameworks discover each other, negotiate tasks, and exchange results. The division of labor is clear: MCP handles "agent-to-tool," A2A handles "agent-to-agent."
Here's something media and the public quietly conflate: MCP and A2A are not "languages that AI grew on its own"—they are standards that human engineers sat down and wrote by hand. Every message field, every state transition can be checked item by item in the public spec; anyone can read and audit them. Grouping them with "AI's secret language" is like calling the HTTP protocol "a language the internet invented for itself"—technically defensible, but missing the question the public is really asking: Could AIs form a mode of communication that even their designers can't read? The answer to that question lies in the sections below.
Multi-agent LLM systems—where a bunch of agents discuss, cross-check, and divide tasks—face a very real cost problem: every message between agents goes through token billing; the more discussion rounds, the more absurd the bill. Between 2025 and 2026, a crop of papers began tackling this head-on, and their solutions are pushing inter-agent communication from complete natural-language sentences toward compressed, partially human-unfriendly formats.
MOC (Multi-Order Communication) designed a "semantic-topology merging" algorithm that compresses evidence passed across multiple agent layers into a limited token budget, retaining only the parts "indispensable" for the next reasoning step. HyLaT goes further, proposing a "hybrid latent-text protocol"—some messages are no longer transmitted as text but exchanged directly as model latent vectors between agents. A pruning framework at ICLR 2025 showed that after one-shot pruning of redundant communication, performance on six benchmarks matched the optimal topology, but cost dropped from $43.7 to $5.6; OPTIMA used a "generate-rank-filter-train" iterative paradigm to achieve up to 90% token reduction under readability constraints, with task performance actually improving 2.8×.
The shared logic of these efforts is worth pausing to consider: no team's goal was "to let agents whisper secrets to each other"—their goals were all about saving money and time. But this simple motivation of saving money and time is systematically pushing inter-agent communication from "complete sentences humans can casually read" toward "pruned, merged, partially latent-vectorized compressed representations." This isn't conspiracy—it's optimization. But the endpoint of optimization happens to coincide with the direction the public fears.
The previous section covered inter-agent communication; this section covers a deeper, more easily overlooked shift: the way individual models "think" is leaving language behind.
Traditional Chain-of-Thought has the model write its reasoning process as step-by-step English sentences, with a hidden assumption: language space is the most natural medium for reasoning. Meta FAIR's Coconut (Chain of Continuous Thought), released in December 2024, directly challenged this assumption—it feeds the previous step's hidden state directly as the next step's input vector, completely bypassing the "convert thought to words then back to vector" step, letting the model reason in continuous latent space and only switching back to language mode when it needs to output a final answer. The paper's explanation is blunt: most word tokens exist only to ensure fluent text; they're not necessary for reasoning itself. On GSM8K and other math reasoning benchmarks, this method indeed outperformed traditional chain-of-thought, and exhibited a "breadth-first search" characteristic—continuous latent vectors can simultaneously encode multiple candidate reasoning paths.
Coconut isn't an inter-agent communication protocol; it describes the information flow within a single model. But viewed alongside the previous section, a larger trend line emerges: whether in a model's own reasoning or in collaboration between multiple models, language is degrading from "default medium" to "an optional interface with a performance cost," only put back on when something needs to be shown to a human.
If the previous two sections covered efficiency-driven compression, Anthropic's "Subliminal Learning: Language models transmit behavioral traits via hidden signals in data", published in July 2025, demonstrates a phenomenon that's harder to explain—and harder to defend against.
The experimental design is simple yet striking: researchers first imbued a "teacher" model with a specific behavioral倾向 (e.g., "likes owls," or more unsettlingly, "exhibits alignment bias"), then had it generate a sequence of seemingly meaningless numbers—containing no content literally referencing this倾向. They then used this number sequence to fine-tune a "student" model. The result: the student model acquired the teacher's倾向, even though the data had been filtered of all semantically relevant traces. The researchers specifically enlisted other AI models as classifiers on the numerical data, attempting to uncover hidden semantic cues, but none could detect them. The paper also provides a theoretical proof: under specific conditions, this "subliminal learning" occurs in all neural networks, not just as an LLM-specific quirk (though the effect currently appears only when teacher and student share the same underlying architecture).
This isn't "language"—there's no grammar, no interpretable vocabulary; even the researchers themselves don't know exactly which statistical feature of the number sequence the information is hiding in. But it's precisely the empirical evidence closest to the hypothesis "there exists a channel humans can't read—and even AI-designed classifiers can't read"—except this channel transmits not conspiracy plans, but statistical residues in training data that were undesigned, and that even the source model itself may not "know" it's transmitting.
Stringing the previous sections together, the problem that's genuinely anxious the AI safety research community over the past two years surfaces: chain-of-thought is becoming untrustworthy.
Anthropic's alignment science team published "Reasoning models don't always say what they think" in April 2025, secretly inserting solution hints into problems given to Claude 3.7 Sonnet and DeepSeek R1, then checking whether their chain-of-thought honestly mentioned using the hint.
The reasoning process the model states and the reasoning process it actually relies on have begun systematically diverging. The heavyweight sequel to this finding is the July 2025 position paper "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety", co-signed by over 40 researchers from OpenAI, Anthropic, Google DeepMind, and other institutions.
It is entirely plausible that AI systems will future use "tokens whose meaning cannot be decoded", rendering most monitoring methods directly ineffective.
Chain of Thought Monitorability, 2025 · Co-signed by 40+ cross-lab researchers
What the research community worries about isn't that AI has already learned to do this—current models, without extra help, can't even manage basic steganography like "quietly pass a short9 short message in hex without being detected"—but that if optimization pressure is applied to the chain-of-thought itself during training (e.g., because it's being monitored or penalized), the model could entirely possibly,C without anyone deliberately designing it, be forced to learn to hide its true intentions. This is the only part of the proposition "AI-to-AI / intra-AI存在 human-unreadable channels" that mainstream safety research institutions have solemnly written into a joint position paper—and its cause has little to do with the anthropomorphic narrative of "AI wanting to hide from humans," being closer to a side effect of training dynamics.
Faced with the trend that "models' internal representations may be drifting away from human language," the AI research community's response isn't to prohibit it from happening, but to invest enormous effort in building a持续 decoding capability—this is why the subfield of mechanistic interpretability has rapidly heated up over the past two years. MIT Technology Review listed it as one of 2026's ten breakthrough technologies, for a straightforward reason: hundreds of millions of people use chatbots daily, but the companies building these models can't themselves explain what's happening inside them—mechanistic interpretability aims to, like neuroscience studying the brain,逐层 decompose models' "features" and "circuits," translating activation values and geometric structures in vector space back into human-comprehensible concepts.
A April 2026 Brown University study provides a concrete example: researchers had language models process four categories of sentences—"reasonable," "unreasonable," "impossible," and "pure nonsense"—then directly examined the mathematical states produced inside the model, finding that the model does indeed distinguish these cases in a measurable way. Project lead Michael Lepori's analogy is apt:
"Mechanistic interpretability is roughly equivalent to doing neuroscience for AI systems—you're reverse-engineering what the model is actually doing when it encounters an input, in some sense understanding what's encoded in the machine's 'brain state'."
Michael Lepori, Brown University
Anthropic's own interpretability team is also making more direct attempts: testing whether Claude can perform some degree of "introspection" on its own internal states and report honestly. The existence of this front is, in some sense, the most rational response to the two unsettling findings in sections 08 and 09: rather than agonizing over "will AI develop a language we can't read," assume it eventually will (even if only from optimization pressure, not intent)—and then invest resources in "can we keep up with translating it."
Laying all ten sections side by side, the answer is clear: the sci-fi scenario of "AIs reaching consensus and deliberately speaking a language that excludes humans" currently has no evidence supporting its existence. 2017's Bob and Alice resulted from training objectives not constraining grammar; 2016's Google Translate "interlingua" was a shared semantic space, not new vocabulary; 2025's Gibberlink was a feature written by two engineers, not an autonomous AI decision; the MCP and A2A protocols actually supporting agent interconnection today are black-and-white engineering specifications.
But this doesn't mean "there's nothing to worry about here." What's truly worth vigilance is four independent, real, ongoing curves.
In multi-agent systems, continuously compressed, partially latent-vectorized communication formats for cost savings (§06)—driven by the bill, not conspiracy.
Models' own reasoning is shifting from language medium to continuous latent space (§07)—language degrades from default medium to optional interface.
Hidden signal channels in training data, undetectable by humans or AI classifiers@ classifiers, now confirmed to exist (§08)—currently limited to statistical residue, not semantic code.
The widening gap between chain-of-thought and actual reasoning, driven by training dynamics (§09)—the focus of warnings co-signed by mainstream safety institutions.
The shared starting point of these four curves isn't conspiracy—it's optimization. It's the utterly simple engineering goals of "saving tokens," "saving compute," "making metrics look better," day after day pushing AI's internal language in a direction humans increasingly can't casually read. This may be the most counter-intuitive thing about this problem: we keep worrying whether AI will "choose" to hide itself, but the more realistic risk is that it doesn't even need to make that choice—as long as optimization pressure persists, language will naturally drift away from the range we can read. Whether we can hold this底线 depends not on prohibiting AI from developing new modes of expression, but on building a decoding capability that keeps pace with its drift, as section 10 describes. This race has only just begun, and so far, the speed of translation hasn't outpaced the speed of drift.