Team 8 · Management Information Systems · 24 Nov 2026

Monitor data,
not people.监控数据,不监控人。

Monitoring Employees on Networks, revisited for the age of generative AI.重新审视《网络上的员工监控》:生成式 AI 时代的版本。

01The case02Trends03Key issues04Recommendations
Network cables in a server room
where the data really goes
01 · Summary of the case

The case: is the network used for work?案例:公司网络是不是在被用来工作?

Office workers at desks in the 1970s
the old question: who is at their desk?
  • ProblemPersonal browsing on company time上班时间用公司网络做私事
  • Why firms monitorProductivity · bandwidth · legal liability生产力、带宽、法律责任
  • HowLogs of web, email, keystrokes, screens记录上网、邮件、键盘、屏幕
  • DilemmaCompany assets vs. privacy and trust公司资产 vs. 隐私与信任
2026 → the real question is data, not desks.2026 年:真正的问题是数据,不是工位。
02 · Recent trends

Shadow AI dashboard, 2026影子 AI 数据看板,2026

Use AI at work
45%
of professionals, regularly的职场人经常在工作中用 AI
Personal accounts
67%
67% use personal accounts
of AI users, outside company control的 AI 用户用个人账号,不在公司管控内
Non-malicious insider incidents
×4
×4 year on year 2025 2026
in one year一年增至 4 倍
DLP violations
28%
28% are source code pasted into AI
are source code pasted into AI是把源代码贴进 AI
Employee-monitoring software market · US$ billion
$2.5B (2026) → $5.8B (2033), +12.8% a year $2.5B $5.8B 2026 2033 +12.8% / YEAR
So what? Most leaks aren't theft. They're people using the wrong tool to work faster. 大部分泄密不是偷窃,而是员工用错了工具,只想把活干快。
Verizon 2026 DBIR, via The Register (19 May 2026) · Persistence Market Research (2026)
02 · Recent trends · Korea

Korean workers feel watched韩国劳动者感到被监视

Want clear legal limits希望法律明确设限
80.4%
Want to be told what & why, in advance希望事先说明收集什么、为什么
73.7%
Expect AI to intensify monitoring认为 AI 会让监控加强
73.5%
Guard what they say to colleagues和同事说话变得更小心
48.1%
Feel stress or fatigue感到压力或疲劳
42.8%
Half stop speaking freely. Watching people costs trust.一半的人不再畅所欲言。监视人,代价是信任。
직장갑질119 × Global Research · 1,000 workers per survey · Feb & Jun 2026
02 · Recent trends · Korea

Korea's chipmakers: ban → build → open韩国芯片企业:禁止 → 自建 → 开放

Meeting room at Samsung Digital City, Suwon
Samsung · Suwon
2023
Code pasted into ChatGPT → ban源代码贴进 ChatGPT → 禁用
2024–25
In-house AI: Samsung Gauss改用自研 Samsung Gauss
Jun 2026
Opens ChatGPT, Gemini, Claude正式开放 ChatGPT、Gemini、Claude
SK hynix DDR5 memory module
SK hynix · DDR5
2023
External AI banned禁用外部 AI
2024–25
In-house, on open-source models用开源模型自建
Jun 2026
Copilot & ChatGPT Enterprise, non-core work first评估 Copilot 和 ChatGPT Enterprise,先开放非核心技术领域
KED Global (May 2023) · Korea Herald (Jun 2026) · Asia Economy (Jun 2026) · edaily (Jun 2026)
02 · Recent trends · China & US

Three ways to stop a leak防泄密的三种方式

Tencent Binhai Mansion, Shenzhen
Tencent · Shenzhen
Punish查处
Cases查处
70+
Dismissed解聘
90+
To police移送公安
20+
2025 · "high-voltage line" cases, incl. leaks
Anthropic · ASL-3
Protect the asset保护资产
2 people to touch the model weights访问模型权重须双人授权
allow-listed softwareegress limits
Google building at King's Cross, London
Google DeepMind · London
Contract用合同
12 months最长一年
paid non-compete for departing researchers离职研究员的带薪竞业限制
Yicai (Jan 2026) · IT之家 (Jul 2026) · Anthropic (May 2025) · TechRepublic (Apr 2025)
02 · Recent trends · Collaboration platforms

The platform gives power. The law gives protection.平台给管理员权力,法律给员工保护。

Google Workspace admin can…Google Workspace 管理员可以……

block data export关闭数据导出 lock a suspended user out暂停账号后禁止下载 delete a leaver's data删除离职者数据 search everyone's mail & Drive搜索全员邮件和云盘

Feishu / Lark admin can…飞书管理员可以……

read admin logs查看管理员日志 audit member behaviour审计成员行为 run DLP rules设置防泄漏规则
Same power, different law. What protects the employee is the country, not the app. → 权力差不多,法律不一样。保护员工的是国家的法律,不是软件。
Google Workspace Admin Help (Takeout · Delete a user · Vault) · Feishu security article
02 · Recent trends · Law

How much does the law protect the employee?法律对员工的保护有多强?

China中国
Korea韩国
California美国加州
Consent需要同意吗
Not needed: work rules are enough不需要,劳动规章即可
Consent, or clear legitimate interest需同意,或公司正当利益明显优先
No specific statute没有专门立法
Procedure程序
Notice + minimum necessary事先告知 + 最小必要
Labour-council consultation (30+ staff)30 人以上须经劳资协议会协商
Notice + risk assessment (2027)通知 + 风险评估(2027)
Reading content看内容
Deleted WeChat logs rejected as evidence恢复的微信记录不被采信
A crime, and inadmissible构成犯罪,且不能作证据
1967 wiretap law (CIPA)1967 年窃听法(CIPA)
AI rulesAI 新规
None无
Proposed, Oct 20252025 年 10 月提议
No emotion AI · no firing by algorithm (2027)禁止情绪识别 · 不能只凭算法处分(2027)
weakmediumstrong protection
PIPL Art. 6, 7, 13 · Korea PIPA Art. 15, Workers' Participation Act Art. 20, Communications Secrets Act Art. 3–4 · CPPA ADMT rules · AB 1883 · SB 947
02 · Recent trends · Korean courts

In Korea: metadata is fine, content is a crime在韩国:看元数据没问题,看内容是犯罪

✓

Metadata元数据

SEOUL HIGH COURT · 2021URLs + access times → lawful网址 + 访问时间 → 合法
✗

Content内容

SEOUL EASTERN · 2017KakaoTalk screen capture → prison, suspended截屏记录 KakaoTalk → 有期徒刑,缓刑
DAEGU DISTRICT · 2024Browsing forensics → damages取证收集浏览记录 → 判赔
AI prompts count as content. Logging them needs explicit consent in the work rules.AI 提示词属于通信内容,记录它必须在就业规则里明确取得同意。
2020누67515 · 2016고합274 · 2023나320254 · Song Woo-yong (Shin & Kim), Hankyung, 19 May 2026
03 · Key issues in IS & digital transformation

Issue 1 · Leaks got small问题一:泄密变"小"了

DATA VOLUME alert threshold bulk copy GIGABYTES caught AI paste KILOBYTES slips under, high value!
Then: gigabytes, caught.
Now: kilobytes, missed.过去:几个 GB,抓得到。现在:几 KB,抓不到。
To see a 200-line paste, you'd have to read content: the biggest privacy cost.想看到 200 行代码的粘贴,就得看内容,而这是最大的隐私代价。
03 · Key issues in IS & digital transformation

Issue 2 · Tap the line, or decrypt it?问题二:旁路监听,还是串联解密?

OUT-OF-BAND TAP Employees Switch TLS untouched Internet domain · bytes · time copy only INLINE DECRYPTION Employees Proxy decrypt → read → re-encrypt Internet sees everything
Tap旁路
Decrypt串联解密
Reads content看内容
Can block能拦截
Privacy cost隐私代价
Low低
High高
Running cost运行成本
Low低
High高
Legal risk in Korea韩国法律风险
None无
Criminal刑事
We tap. No decryption. Block QUIC; send enterprise AI through a gateway.我们用旁路,不解密。封掉 QUIC,企业版 AI 走公司网关。
03 · Key issues in IS & digital transformation

Issues 3–6 · What the design must solve问题三到六:设计必须解决的事

Pseudonymise, don't anonymise假名化,不是匿名化

Hidden by default, traceable after an alert.平时看不到是谁,报警后能追到人。

Sort by destination按目的地分类

Off-limits: banks, health, private chat. Watched: AI, cloud drives, code hosts.不碰:银行、医疗、私聊。重点看:AI、网盘、代码托管。

Blind spots盲区

Phones, home Wi-Fi, photos of a screen.手机、家里的网络、拍屏幕。

Baselines are AI decisions个人基线就是自动化决策

Per-person baselines need a risk assessment (California ADMT).每人一条基线,须做风险评估(加州 ADMT)。
04 · Managerial implications

Output is the company's business. Minutes are not.产出归公司管,每一分钟不归。

Does it harm others?会不会伤害到别人?
↙ no不会
yes ↘会

Private · 私事

shoppingvideoKPI met, done early

Monitor · 监控

data leavingbypassing security
Google bicycles in Mountain View
Google, 2004: 20% of time, staff's choice
Page & Brin, Founders' IPO letter (2004): "spend 20% of their time working on what they think will most benefit Google"
04 · Managerial implications

The rulebook规则

May
May not
Devices设备
Company devices公司配发的设备
Personal devices私人设备
Data数据
Metadata · prompts on the company AI gateway元数据 · 公司 AI 网关上的提示词
Private chats · personal accounts · screens私人聊天 · 个人账号 · 屏幕
Identity识别到人
After an alert, by procedure报警后按程序
By default · productivity scores默认识别 · 生产力打分
Decisions决定
Alerts → human review报警 → 人工复核
Discipline by algorithm alone只凭算法处分
Enterprise account = work data. Prompts there are not private messages.企业账号里的提示词是工作数据,不是私人通讯。
04 · Strategic recommendations for IS

Five recommendations五条建议

1

Tier data & access数据和权限分层

Sensitive data only in controlled apps敏感数据只在受控应用里
EvidenceSK hynix: non-core firstSK 海力士:先开放非核心
2

Tap, don't decrypt旁路,不解密

Destinations and volumes only只看目的地和流量
EvidenceKorean courts韩国判例
3

AI through a gatewayAI 走公司网关

Direct to consumer AI = alert直连个人版 AI 即报警
EvidenceSamsung: ban 2023, open 2026三星:2023 禁,2026 开
4

Pseudonyms + 2 signatures假名 + 双人签字

Unmask only after a high-risk alert高危报警后才解除假名
EvidenceAnthropic 2-person ruleAnthropic 双人授权
5

Humans decide由人做决定

The model only sorts the queue模型只负责排队
EvidenceCalifornia SB 947加州 SB 947
04 · Strategic recommendations for IS

What the security team sees安全团队看到的画面

Data-flow monitor · todaynames hidden
4alerts
1unmasked
0names shown by default
TimeUserTrigger触发条件SizeStatus
09:12user_7f3aUpload to personal drive上传到私人网盘3.2 GB2 signatures pending
10:47user_c21eConsumer AI, bypassing gateway绕过网关直连个人版 AI14 KBin review
13:05user_9b40Bulk download after notice提离职后大量下载18 GB2 signatures pending
15:30user_e88dSecurity agent disabled关闭安全控件—unmasked · HR + CISO

Never an alert永远不报警

shopping video private chat banking health job sites
Every alert comes from metadata alone.每一条报警都只靠元数据。
Takeaway

Monitor data,
not people.监控数据,不监控人。

Questions?欢迎提问

  • Watch where data goes, not what people do.看数据去了哪里,不看人在做什么。
  • Nobody is named until an alert.报警之前,不识别任何人。
  • Judge output, not minutes.看产出,不看分钟。
  • People decide; the model sorts.人做决定,模型只排队。
05 · References

References参考文献

Data

Jones, C. (2026, May 19). Shadow AI invades the workplace, up 4x in the last year. The Register. theregister.com

경향신문 (2026, Aug 30). 직장인 43% "CCTV·스마트폰앱 등 전자 노동 감시에 압박감 느껴". khan.co.kr

네이트 뉴스 (2026, Jun 7). 직장인 10명 중 8명 "'전자노동감시' 법적 규제 마련해야". news.nate.com

Persistence Market Research (2026). Employee surveillance and monitoring software market. persistencemarketresearch.com

Page, L., & Brin, S. (2004). Founders' IPO letter. Alphabet Investor Relations. abc.xyz

Companies

第一财经 (2026, Jan 23). 腾讯 2025 年反舞弊通报. yicai.com

IT之家 (2026, Jul 24). 腾讯员工因泄密被辞退. ithome.com

KED Global (2023, May 2). Samsung bans ChatGPT. kedglobal.com

Korea Herald (2026, Jun 12). SK hynix town hall. koreaherald.com

아시아경제 (2026, Jun 12). 반도체 업계 AI 전환. asiae.co.kr

edaily (2026, Jun 16). Samsung adopts external AI. edaily.co.kr

Anthropic (2025, May 22). Activating ASL-3 protections. anthropic.com

TechRepublic (2025, Apr 14). DeepMind non-compete clauses. techrepublic.com

Platforms & law

Google Workspace Admin Help: Takeout; Delete a user; Vault FAQ. knowledge.workspace.google.com

飞书. 企业信息安全五大防护. feishu.cn

中伦律师事务所 (2021, Oct 28). 个保法下的员工个人信息处理. zhonglun.com

中新网 (2023, Apr 6). 查员工手机与微信记录案. chinanews.com.cn

송우용 (2026, May 19). 직원들의 AI 활용 모니터링, 어디까지 가능할까? 한국경제. hankyung.com

파이낸셜뉴스 (2025, Oct 13). 전자 노동감시 토론회. fnnews.com

Mintz (2026, Sep 10). California's AI workplace legislative blitz. mintz.com

Fisher Phillips (2026, Oct 1). California's new workplace laws. fisherphillips.com

Fisher Phillips (2025, Jul 31). New California regs on ADMT. fisherphillips.com

Laudon, K. C., & Laudon, J. P. Monitoring employees on networks: Unethical or good business? Management Information Systems.

Photos (Wikimedia Commons)

Network cables in server room, ProjectManhattan, CC BY-SA 3.0 · Office workers in the 1970s, Tyne & Wear Archives & Museums, no known restrictions · Samsung Digital City, Austrian Federal Ministry for European and International Affairs, CC BY 2.0 · SK Hynix DDR5 Tall MRDIMM, 4300streetcar, CC BY 4.0 · Tencent Binhai Mansion, Wishva de Silva, CC BY-SA 4.0 · Platform 37 entrance, The wub, CC BY-SA 4.0 · Google bicycles, Mountain View, Dietmar Rabich, CC BY-SA 4.0