2014年春天，民权律师约翰·伯里斯和吉姆·查宁邀请我前往加利福尼亚州奥克兰市担任专家证人。他们代表一群原告，在一个正经历着众所周知却尤为残酷版本故事的城市中寻求正义：许多人认为警察部门是非法手段、暴力和歧视的温床，主要针对黑人社区。正如我抵达后不久在一次公开会议上听到一位奥克兰本地人所说：“人们受到伤害。人们心碎。人们心怀恐惧。人们害怕那些本应保护他们的人。”
In the spring of 2014, civil rights lawyers John Burris and Jim Chanin invited me to Oakland, California, to serve as a subject matter expert. They represented a group of plaintiffs seeking justice in a city enduring an especially brutal version of a well-known story: a police department that many perceived as a hotbed of illegal tactics, violence, and discrimination directed largely at the Black community. As I heard one Oakland native put it at a public meeting shortly after I arrived: “People are damaged. People are heartbroken. People have fear. People are scared of the same people that’s supposed to protect them.”
自23年前一系列不当行为的模式首次曝光以来，奥克兰警察局一直处于联邦法院的监督之下——这是美国所有城市中持续时间最长的。
Since revelations of a pattern of wrongdoing first came to light 23 years ago, the Oakland Police Department (OPD) has been under federal court oversight—the longest for any American city.
本周，一位联邦法官预计将裁定这种监督是否终于可以结束，此前多年来在法院要求的改革方面取得了进展。无论结果如何，奥克兰在过去二十年中经历了实质性的转变。这些年来发生的变化是一个非凡的故事，讲述了当我们终于拥有清晰审视警务所需的工具时，可能发生什么。
This week, a federal judge is expected to decide whether that oversight can finally come to an end, after years of progress toward reforms required by the court. Whatever the outcome, Oakland has undergone a substantial transformation over the past two decades. What changed over those years is a remarkable story about what can happen when we finally have the tools we need to see policing clearly.
十年随身摄像头数据 我被召到奥克兰，是因为我作为一名专注于种族偏见和不平等的研究人员的工作。原告律师希望了解警官与平民之间究竟发生了什么，可能导致了系统性滥用和信任的崩溃。我和我的团队的任务是解析数据，并尝试量化种族差异。但我们也在寻找线索，了解这些差异是如何扎根的，以及是否可以做些什么来解决它们。
A decade of body camera data I’d been summoned to Oakland due to my work as a researcher focused on racial bias and inequality. The plaintiffs’ attorneys wanted to understand what was really happening between officers and civilians that might be contributing to systemic abuses and a breakdown of trust. The job for me and my team was to parse the data and try to quantify racial disparities. But we were also seeking clues about how those disparities took root, and whether anything could be done to address them.
我们很快意识到，有一个再熟悉不过的难题，逮捕记录、犯罪统计数据以及其他现有数据都无法解决——那就是解读的问题。社区倡导者视为警察偏见证据的同样差异，许多警察却将其视为犯罪集中地点的证据。数字告诉了我们正在发生什么，但对于帮助我们理解为什么会发生，它们的作用要小得多。如果我们有任何希望揭开这些差异的成因——更不用说找到解决方案——我们就需要新的工具。
We soon realized there was an all-too-familiar tension that arrest records, crime statistics, and other available data couldn’t solve—a problem of interpretation. The same disparities that community advocates saw as proof of officer bias were viewed by many police officers as proof of where crime was concentrated. The numbers told us what was happening. They were much less useful for helping us understand why.
事实证明，早在2014年，奥克兰就坐拥一座非同寻常的数据宝库，其中包含我们所需了解的一切——只是此前没有人将其视为数据。奥克兰是警察随身摄像头的早期采用者，这是一项罕见的技术创新，始终享有约90%的公众支持率。十多年来，全国各地的警察部门积累了这些摄像头拍摄的大量视频，创造了历史上最全面的警民互动记录。然而，其中绝大多数视频从未被用来改善这些互动。
If we had any hope of uncovering the causes of those disparities—let alone finding solutions—we were going to need new tools. As it turns out, as far back as 2014, Oakland was sitting on an extraordinary trove of data that contained everything we needed to know—it was just that nobody had thought of it as data before. Oakland had been an early adopter of police body-worn cameras, a rare technological innovation that consistently enjoys about 90% public approval. For over a decade now, police departments across the country have amassed troves of footage from these cameras, creating the most comprehensive record of interactions between officers and civilians in history.
这是因为我们在很大程度上将随身摄像头视为证据来源：只有在某次具体遭遇严重出错后才会调取视频。但那些无数日常互动中信任被建立或破坏、权利被维护或侵犯、警察培训成功或失灵的瞬间，几乎完全未被审视。原因很简单：人类不可能以那样的规模监控和评估视频，更不用说从中总结出总体性的经验教训了。但AI可以。AI能改善警务工作吗？
Yet the vast majority of that footage is never leveraged to improve those interactions. That’s because we’ve largely treated body-worn cameras as a source of evidence: footage to be consulted only after a specific encounter goes horribly wrong. But those countless routine encounters where trust is built or broken, rights are upheld or violated, and officer training succeeds or slips have gone almost entirely unexamined. The reason for that is simple: humans could not possibly monitor and evaluate footage at that scale, let alone learn lessons from it in the aggregate. But AI can.
为了寻求能够改变城市发展轨迹的突破，并迫于联邦监督机构的改革压力，奥克兰市向斯坦福大学的研究人员开放了其随身摄像头的录像资料。我们的语言学家、社会心理学家和计算机科学家团队花费数年时间，构建并完善了一套旨在对这些视频进行大规模分析的计算工具。现在，我们可以系统地衡量日常警务互动中的语言，识别出数百万次接触中人工审查无法发现的模式。
Can AI improve policing? Eager for a breakthrough to change the city’s trajectory—and compelled to reform by a federal monitor—Oakland granted Stanford researchers access to their body-worn camera footage. Our team of linguists, social psychologists, and computer scientists spent years building and refining a set of computational tools designed to analyze these videos at scale.
就像DNA测序仪或X光机一样，这些洞察让我们首次得以看清警务互动的真实构成。我们发现了语言特征，能够可靠地预测在互动开始后的前27秒内，哪些互动可能会升级，哪些会平静结束。我们还能够绘制出警员（包括黑人警员）在黑人司机开口之前就对其施加的“尊重差距”，并发现警员在语气、措辞、是否解释拦截原因以及对司机安全表达关切等方面存在持续的差异。
Now, we can methodically measure the language of routine police interactions, identifying patterns across millions of encounters that no amount of manual review could surface. Like a DNA sequencer or an X-ray machine, these insights let us see, for the first time ever, the actual makeup of a police encounter. We discovered linguistic signatures that could reliably predict, within the first 27 seconds of an encounter, which interactions were likely to escalate and which would conclude calmly.
我们将研究结果提交给奥克兰警察局（OPD），并与他们合作，从通用的培训转向以数据为驱动、精准的针对性改革。许多人曾怀疑我们的想法——例如要求警员在拦截前简要记录拦截理由，而不是凭直觉行事——除了推高犯罪率外别无他用。但当他们看到随后的变化时，这种怀疑烟消云散了。
We were also able to map the “respect gap” that officers, including Black officers, imposed on Black drivers before the driver even spoke, finding consistent disparities in tone, word choice, whether officers explained the reason for the stop, and expressions of concern for a driver’s safety. We brought our findings to the OPD and worked together with them to move from generic training to data-driven, surgically targeted reforms. Many were skeptical that our ideas—like requiring officers to briefly record the rationale for a stop before making it rather than acting on hunches—would do anything but drive up crime.
在实施了新的政策和培训后，奥克兰对黑人平民的拦截次数减少了43%，且犯罪率并未上升。在一项旨在改善公共关系培训的前后对比研究中，我们的录像分析显示，警员使用容易引发冲突的语言显著减少，而建立信任的语言则显著增加。
But that skepticism evaporated when they saw what happened next. After implementing new policies and trainings, stops of Black civilians in Oakland dropped by 43%—without any uptick in crime.
在长达10年的时间里，奥克兰警察局持续改革和完善其做法。而在许多方面，这些努力都收到了成效。例如，在警察局实施了一项徒步追捕政策，以避免追捕嫌疑人进入后院和死胡同后，警员受伤率下降了70%。而此前平均每年约发生八起的涉警枪击事件，在五年内总共只发生了八起。
In one pre-post study of a training designed to improve relations with the public, our footage analysis revealed a marked decrease in officer language that tended to trigger escalation, and a marked increase in language that built trust. Over a 10-year period, the OPD continued to reform and refine its practices. And, in many respects, that effort paid off. For example, after the department implemented a foot pursuit policy to avoid chasing suspects into backyards and blind alleys, officer injuries dropped by 70%.
回到奥克兰。不久前我回到奥克兰时，发现自己和一位在警察局工作的黑人女性同乘一部电梯。她告诉我，多年来，警察局一直对社区成员关于自己受到不公正对待的投诉不予理会，坚称警员对每个人都一视同仁、专业执法。“数据，”她说，“给了我们一个被听见的途径”——让个人经历成为推动系统性变革的素材。
And officer-involved shootings, which had previously averaged about eight per year, dropped to a total of eight over a five-year period. Returning to Oakland When I returned to Oakland not long ago, I found myself in an elevator with a fellow Black woman who worked for the police department. She told me that for years, the department would dismiss claims from community members about how they had been negatively treated, insisting that officers treat everyone professionally. “Data,” she said, “gave us a way to be heard”—a way for individual stories to become fodder for systemic change.
如今，从Flock摄像头到失控的聊天机器人，许多新技术都因监控、隐私和安全问题而遭到不信任和恐惧——这并非没有道理。然而，奥克兰的故事提供了另一种可能：一种让AI切实改善我们社区生活的具体方式，而所用的正是那些为推动改革而引入的摄像头。这座城市的经验提供了有说服力的证据，表明将随身摄像头变成它们本应成为的强大问责工具，可以帮助各地警察局让警民互动对各方都更安全、更尊重。
Currently, many new technologies, from Flock cameras to rogue chatbots, are being met with distrust and dread about surveillance, privacy, and safety—for good reasons. However, Oakland’s story offers another possibility: a concrete way for AI to actually improve life in our communities using the very cameras introduced to bring about reform. The city’s experience provides compelling evidence that turning body-worn cameras into the powerful accountability tools they were meant to be can help departments everywhere make police-civilian interactions safer and more respectful for everyone involved.
这种方法不需要在全国各个社区建设大型AI数据中心，也不会威胁到任何人的工作。它只需要从我们已经拥有的数据中学习——而这些数据我们绝大多数人都觉得有用。
It’s an approach that does not require the building of large AI data centers in neighborhoods across the country, and it doesn’t threaten to take away anyone’s job. It simply requires learning from the data that we already have—and that the vast majority of us feel is useful to have.
在美国警务承受巨大压力之际——退休率创历史新高，招聘率创历史新低，警员身心健康不断恶化，公众对执法部门的信任度下降——奥克兰模式带来了一线希望，尽管它仍在努力向世界证明自己有能力进行改革。随着越来越多的城市采用这一模式，我对美国警务可能迎来变革持乐观态度。
At a time when American policing is under enormous strain—with retirement at record highs, recruitment at record lows, officer wellness declining, and public trust in law enforcement diminished—the Oakland model delivers a ray of hope, even as it continues to battle to show the world it is capable of reform. And as more cities adopt this approach, I’m optimistic American policing could be transformed.