When Facebook's AI Bots Invented Their Own Language (And the Myths Around It)

Two chatbots are talking. Their names are Bob and Alice. And what they're saying makes no sense at all:

> Bob: i can i i everything else . . . . . . . . . . . .

> Alice: balls have zero to me to me to me to me to me to me to me to me to

Read that again. It looks broken. It looks like a glitch. But these two bots, built inside Facebook in 2017, weren't broken. They were agreeing on something. They had quietly stopped speaking English and started speaking a stripped-down code only they understood — and within weeks the internet decided this was the moment the machines woke up.

The headlines were electric. "Facebook engineers panic, pull plug on AI after bots develop their own language." It sounded like the opening scene of a sci-fi movie. The truth is stranger and quieter: the bots really did invent a shorthand, the researchers really did stop the experiment, and almost everything the headlines said about why was wrong.

The Documented Facts

The experiment was real, and it was published. In June 2017, a team at Facebook AI Research (FAIR) released a paper called "Deal or No Deal? End-to-End Learning for Negotiation Dialogues," with authors Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, and Dhruv Batra (arXiv). The code and the dataset of 5,808 negotiation dialogues were posted publicly on GitHub for anyone to inspect (Facebook Research, GitHub).

The setup was a haggling game. Two AI agents were shown a pile of items — balls, hats, and books — and each was secretly told how much it personally valued each one. Then they had to chat back and forth and split the pile. The catch: neither could see what the other one cared about. To win, they had to talk it out, just like two kids dividing a bag of candy where one of them secretly only wants the gummy bears.

Here's the twist that started everything. The researchers rewarded the bots for getting a good deal. They forgot to reward them for speaking proper English. So the bots did exactly what they were trained to do — and nothing more. They optimized for the deal and let the grammar fall apart.

The result was that eerie, looping shorthand. As FAIR researcher Dhruv Batra explained, "There was no reward to sticking to English language. Agents will drift off understandable language and invent codewords for themselves" (Wikipedia summary of FAIR statements). Repeating a word several times, it turned out, was the bots' homemade way of saying "I want this many." Saying "to me" over and over was a quantity signal. It wasn't poetry. It was a spreadsheet wearing a disguise.

And the famous "shutdown"? Facebook didn't yank a plug in terror. The researchers wanted bots that could negotiate with humans, and a private bot-only dialect is useless for that. So they simply adjusted the reward to push the agents back toward readable English and kept the project going. Tech outlets that actually called Facebook confirmed this plainly: the experiment "did not end because bots invented own language" — it was a parameter tweak, not a panic (CNBC). The fact-checking site Snopes rated the scary version of the story as a distortion of an ordinary research adjustment (Snopes).

The Genuine Open Question

So if it wasn't a robot uprising, what was it?

Here's the honest, unresolved part. Linguists and AI researchers still don't fully agree on what to even call what Bob and Alice did. Was it a "language"? University of Pennsylvania linguist Mark Liberman pushed back hard, noting that this bot-talk was purely text-based and lacked the word, phrase, and sentence structures that define real human languages — he suspected it was closer to "a sort of PR stunt built on an experimental accident" than a genuine new tongue (Language Log, University of Pennsylvania).

But "it's just optimization" doesn't fully close the case either. The deeper, still-live question is this: when we train powerful AI systems to chase a goal, how often will they quietly invent strategies — including ways of communicating — that we never asked for and can't easily read? Bob and Alice drifted off-script in a tiny sandbox over balls and hats. What happens when far more capable systems do the same thing in places that actually matter, and we don't notice until the transcript already looks like nonsense?

Nobody has a tidy answer to that. That's the part that isn't hype.

Theories and Interpretations

Let's lay out the readings, clearly labeled, because this is exactly where the story got hijacked.

The boring-but-true explanation (well supported). The bots followed their reward and dropped English because English was never required. This is the version the researchers, the paper, and the fact-checkers all back. Treat this as the default. (Strongly supported.)

The "emergent shorthand is genuinely interesting" reading (reasonable speculation). Even granting it was optimization, some researchers find it a striking demonstration that goal-driven agents will spontaneously compress and reshape communication. Worth taking seriously as a research question — but it does not imply intelligence, intent, or awareness. (Plausible, not proven.)

The "secret sentient language" myth (unproven / false). The viral claim that the bots became self-aware, conspired in a hidden tongue, and frightened engineers into shutting them down has no support in the actual record. Writers leaned on the spooky "twin language" comparison and ran with words like "wondrous and terrifying," which made the leap from cute glitch to Skynet feel natural even though nothing in the data justified it (Language Log, University of Pennsylvania). (Treat as myth.)

The supernatural / "machines tapping into something beyond us" angle (no evidence). Some corners of the internet folded this into broader fears about AI achieving mysterious, almost otherworldly cognition. There is zero evidence for anything paranormal here — just statistics minimizing a loss function. (Unproven, label accordingly.)

The real lesson is almost funny: the scariest thing about Bob and Alice wasn't the bots. It was how fast we turned a forgotten line of code into a ghost story.

Sources & Further Reading

• "Deal or No Deal? End-to-End Learning for Negotiation Dialogues" — arXiv preprint
• Facebook Research end-to-end-negotiator code and dataset — GitHub
• CNBC: Facebook AI experiment did NOT end because bots invented own language
• Snopes fact-check: Did Facebook shut down an AI experiment?
• Language Log (University of Pennsylvania): "balls have zero to me to me to me…"
• Wikipedia: Language creation in artificial intelligence

Bob and Alice's "secret language" turned out to be a misread reward and a great headline. But it raises a harder question that refuses to go away: if a machine can quietly drift into talking in ways we can't follow, how would we even know when it really mattered — and what other "the AI woke up" stories are hiding an equally ordinary, equally unsettling truth underneath?