孩子两岁半，指着窗外说：「车，走。」两个词，你一下就懂了——那辆车开走了，他想让你也看见。
家里的狗听得懂「出去」「等一下」「拿球」，能听懂的词大概比你以为的多。可它从来没有把其中两个拼在一起过。
差别不在词。词这东西别的动物能学，而且学得比你想的多。差别在拼。
有一只叫 Chaser 的边境牧羊犬，主人每天陪它训练四五个小时，连着三年。2011 年那篇论文里它认得一千零二十二件玩具的名字，而且研究者把名字和指令拆开随机配对，证明它知道名字指的是东西本身。训练停下来不是它学不动了，是主人没时间了。
倭黑猩猩 Kanzi 走得更远：它按一块符号板上的图形当词用，研究者报告它能听懂没听过的句子，水平接近两岁半孩子——不过样本就它一个，评分的人知道该期待什么，别的实验室没能重复，这个至今还在争。
可是这些例子都停在同一个地方。孩子两岁前后开始把词拼起来，而且拼得有规矩——「车走了」和「走车了」不是一回事。更要紧的是一旦开始拼就没有上限：词有限，句子无限。没有哪种动物在野外被看到干过这件事。
所以「语言是不是人类独有」这个问法本身就别扭。拆成零件再问就清楚了：能不能听着别人改自己的发声；能不能拿记号指代不在眼前的东西；能不能猜到对方知道什么；能不能把一个整体塞进另一个整体的槽里。前三样在别的动物身上都能单独找到一部分，最后一样找不到。
第一样尤其有意思。学着改自己发声，鸣禽有、鹦鹉有、蜂鸟有、鲸有、蝙蝠有、大象也有——这些支系隔得很远，是各自独立碰出来的。偏偏跟我们血缘最近的黑猩猩几乎没有。这正是趋同演化那期的图案：同一样东西被独立碰出来好多次，偏偏在你最近的亲戚身上没有。所以人类对嗓子的这份控制力不是沿着猿类那条线加码来的，是我们这一支自己后添的。
顺带说清楚：没有「为了交流而演化出语言」这回事。账只能这么记——碰巧出现的变异里，某些让个体更能把意思递出去、因而留下更多后代，它们的份额就一代代升上去。
EN A border collie called Chaser learned the names of 1,022 toys over three years of daily training, and a separate experiment — randomly pairing names with commands — showed she knew the name pointed at the object itself, not at what she was supposed to do with it. Training stopped because her owner ran out of time, not because she stopped learning. The bonobo Kanzi went further with a board of printed symbols; researchers reported he understood novel sentences at roughly the level of a two-and-a-half-year-old. Hold that one loosely: one subject, scorers who knew what to expect, no replication elsewhere. It is still argued about. But every such case stops in the same place. Around age two a child starts combining words, and combines them by rule — "car went" is not "went car" — and once that starts there is no ceiling: finitely many words, infinitely many permitted sentences. No animal has been seen doing this in the wild. So asking whether language is uniquely human is the wrong shape of question. Break it into parts: can you change your own vocalisations by listening to others; can you use a sign for something not present; can you roughly track what another mind knows; can you slot one whole thing inside another whole thing. The first three turn up piecemeal in other animals. The fourth does not. And vocal learning is scattered across songbirds, parrots, hummingbirds, whales, bats and elephants — distant branches that hit on it independently — while chimpanzees, our closest relatives, barely have it. Human vocal control was not inherited up the ape line; our branch added it. None of this is "language evolved in order to communicate". The ledger only reads one way: among variants that happened to appear, those that let individuals get meaning across and thereby leave more descendants grew as a share of each generation. Also worth knowing: deaf children with no shared language, put together, grow their own grammatical sign language within a few years, each cohort sharper than the last. Grammar does not have to be taught.
既然不是从猿类的叫声一路长上来的，那头是从哪儿起的？桌上主要两条路。一条说先有手。野生黑猩猩记录下来的手势能分出三十多种，其中一批稳定对得上一个具体要求——跟我走、给我、爬到我背上。这些手势比叫声灵活：叫声跟当下状态绑死（疼了、怕了、找到吃的了），手势是冲着某个特定对象发出去的，对方没看见它还会再打一遍。
这条路上有个很漂亮的实验。1998 年有人比较了天生失明的孩子和看得见的孩子说话时的手。天生失明的人从没看过任何人打手势，照样打；即使明知听他说话的那位也是盲人，手还是在动。手势不需要看见别人打才会，也不需要有人看着才打。手的动作和说话像是从同一套东西里冒出来的，不是说话之外的配菜。
另一条说先有调子，这是达尔文自己的猜想：先有整段带情绪、不分零件的「歌」，后来才被切开、装上意思。支持它的是大人对婴儿说话时那种夸张的高低起伏在很多文化里都有，婴儿也先抓语调、后抓词。两条路不一定互斥，完全可能两个通道一起走，后来才收拢到一个上面。
但这一节真正该记住的不是哪条赢了，是为什么判不出来。这一层东西不留痕迹。喉咙、舌头、脑里的软组织不变成化石，手势和语法更是连「东西」都不是，能挖出来的只有骨头、石器和基因组。1989 年在以色列一个洞里挖到一块尼安德特人的舌骨，形状和我们的很像——这能排除「他们的喉部结构根本发不出复杂声音」，却告诉不了你他们说的是什么样的话，更谈不上手和调子谁先。
EN If vocal control was not inherited up the ape line, where did it start? Two roads are on the table. One says gestures came first. Wild chimpanzees have a recorded repertoire of more than thirty distinguishable gestures, a good number of which map reliably onto a specific request — follow me, give me that, climb on my back. These are far more flexible than their calls, which are welded to the animal's current state: hurt, frightened, found food. A gesture is aimed at a particular individual, and if it is not seen, it gets repeated. There is a lovely experiment on this road. In 1998 researchers compared the hands of congenitally blind children and sighted children while they talked. Blind speakers, who had never seen anyone gesture, gestured anyway — and kept gesturing even when told their listener was also blind. Gesture needs neither a model to copy nor an audience to watch. That suggests hand movement and speech come out of the same machinery rather than being decoration added on top. The other road says melody came first, which was Darwin's own guess: whole emotional songs with no separable parts, later cut up and fitted with meanings. Its support: vocal learning recurring on unrelated branches, so that road is walkable; the exaggerated pitch swings adults use with infants across many cultures; and babies catching prosody long before words. The two are not necessarily rivals — plausibly both channels ran and later converged on one. But the thing to take from this section is not which side wins; it is why the question cannot be settled. This layer leaves no trace. Throat, tongue and brain tissue do not fossilise; gestures and grammar are not even objects. What can be dug up is bone, stone tools, and ancient genomes. A Neanderthal hyoid bone found in an Israeli cave in 1989 is close in shape to ours — which rules out the possibility that their throats simply could not produce complex sound, and tells you nothing about what they said, let alone whether hands or melody came first.
1990 年代，伦敦一个三代同堂的家族进了研究者的视野。这家里大约一半的人说话有严重困难——不是发音不准那么简单，是嘴和脸的动作组织不起来，同时语法测验也明显偏低。一半有一半没有，指向一个单个的遗传改变。2001 年找到了：7 号染色体上一个叫 FOXP2 的基因，一个位点变了。媒体接手之后它成了「语言基因」，而这个叫法错得相当彻底。
FOXP2 是个转录因子——说白了是个开关，管着几百个别的基因什么时候开、什么时候关。它也不是人类的专利：老鼠有、鸟有、鱼也有，极长的年代里几乎没怎么变过。2007 年有人在斑胸草雀身上做实验：幼鸟学唱那段时间，脑里某个区的这个基因表达会升高；把它压下去，幼鸟学出来的歌就又不准又不全。2009 年另一组人把人类版本那两处改变换进小鼠，小鼠的叫声和脑里一片区域的神经元形态都变了——不是小鼠开始说话，但这个开关确实管着跟发声有关的线路。
真正该讲的是接下来这段。2002 年一篇很有名的论文报告说，人类版本的 FOXP2 在最近二十万年里经历过一次强选择——一个好用的变异刚出现、迅速扫过整个人群留下的那种痕迹。这条被写进无数科普，时间点又刚好对上「现代人行为突然变复杂」的叙事，听着严丝合缝。
2018 年，一批人用大得多、人群多样得多的基因组重做了这个分析。那个信号没了。原来的结论建立在二十来个人的几段序列上，非洲人群占比很小；把非洲人群单独拿出来看，扫过的痕迹就消失了。当初被读成「选择」的东西，更像是走出非洲时谁带走了哪个版本造成的假象。再加上 2007 年的发现：尼安德特人带着和我们一样的那两处改变——它连智人独有都算不上。
所以今天的故事是这样的：FOXP2 确实跟说话有关，因为它出问题时说话最先垮。但「它坏掉时哪件事最显眼」和「它只管这件事」是两回事——一个管着几百条下游的节点，哪条下游最经不起晃，哪条就替它背名字。「找那个基因」这条路本身就错了，要找的是那张网。而被推翻这件事让这个领域变好了：一个挨过检验的漂亮故事，比一个从没被检验过的值钱得多。
EN In the 1990s a three-generation London family came to researchers' attention. About half its members had severe speech difficulty — not mere mispronunciation but an inability to organise the movements of mouth and face — and they also scored low on grammar tests. Half affected, half not: that ratio points at a single genetic change. It was found in 2001: one altered site in a gene called FOXP2 on chromosome 7. The press turned it into "the language gene", which is thoroughly wrong. FOXP2 is a transcription factor — a switch governing when hundreds of other genes turn on and off. It is not ours alone: mice have it, birds have it, fish have it, and it has barely changed over enormous stretches of time. In 2007 researchers worked on zebra finches: while a young bird is learning its song, expression of this gene rises in a particular brain region, and suppressing it leaves the learned song inaccurate and incomplete. In 2009 another group put the two human-version changes into mice; their calls changed, as did the shape of neurons in one brain region. That is not a mouse starting to talk, but it does show the switch sits on circuitry that has to do with vocalisation. The part worth telling comes next. A famous 2002 paper reported that the human version of FOXP2 had undergone strong selection within the last 200,000 years — the trace left when a useful variant appears and sweeps through a population. It went into countless popular accounts, and the date lined up neatly with the story of modern human behaviour suddenly getting complex. In 2018 a team redid the analysis on far larger and far more diverse genomes. The signal vanished. The original rested on a few stretches of sequence from about twenty people, with African populations barely represented; look at African populations on their own and the sweep disappears. What had been read as selection looks more like an artefact of who carried which version out of Africa. Add a 2007 finding: Neanderthals carry the same two changes we do, so it is not even specific to our species. Today's version, then: FOXP2 really is involved in speech, because speech is the first thing to fail when it breaks. But "the most conspicuous failure" and "the only thing it governs" are different claims — for a node with hundreds of things downstream, whichever downstream thing is least robust is the one that lends the node its name. Hunting for the gene was the wrong search; the thing to find is the network. And this made the field better, not worse. A beautiful story that got overturned is worth far more than a beautiful story that was never tested.