Name: 三大AI職場難題，你必須知道！| Martin Gonzalez for Big Think + (3 AI puzzles workplaces must solve | Martin Gonzalez for Big Think +)
Uploaded: 2025-12-04T17:55:21.000Z
Duration: 8 min 51 s

I'm a principal of Organization and Leadership Development at Google and I'm the author of the Bonfire Moment.

我是 Google 的組織與領導力發展首席顧問，也是《Bonfire Moment》這本書的作者。

The book explores this idea that teams are harder than tech in the process of innovation.

這本書探討一個觀點，那就是在創新的過程中，團隊合作比 tech 更具挑戰性。

It's so important for leaders and CEOs and founders to pay attention to the people side of the business because that could easily derail your best laid out plans.

領導者、執行長和創辦人非常重視業務中的人員面向，因為這可能會輕易地打亂你最周詳的計畫。

We know a lot of employees and organizations are starting to use AI for their work.

我們知道很多員工和組織都開始在工作中使用 AI。

We also know that we flip flop between these really intense narratives of substitution.

我們也知道，我們在這些非常極端的敘事之間搖擺不定，一個是替代。

There will be less of people playing my kind of work, my kind of role because of AI and a narrative of augmentation, which is these tools give me superpowers that allow me to do more within my role and, and

因為 AI，會有越來越少的人從事我這種工作、扮演我這種角色，以及一個增強的敘事，也就是這些工具賦予我超能力，讓我在我的角色內做得更多，然後

I will succeed and do well in the future if I can only adapt these new technologies.

如果我能適應這些新技術，我將在未來取得成功並表現出色。

There's a lot we need to think about when we think about the augmentation model.

關於增強模型，有很多東西需要我們思考。

Because as early research is showing, as we bring these tools into the workplace, we're not quite seeing the kind of transformative potential that AI has been talked about by its inventors.

因為早期研究顯示，當我們將這些工具帶入工作場所時，我們並沒有看到 AI 的發明者所談論的那種轉型潛力。

So I've started to think about three puzzles we need to solve for as we bring these technologies into our organizations.

所以，當我們將這些技術帶入我們的組織時，我開始思考需要解決的三個難題。

One of the challenges in bringing AI into an organization is what I've started to call the selective upgrade puzzle.

將 AI 引入組織的挑戰之一，是我開始稱之為「選擇性升級難題」。

This is when these tools endow its users with superpowers, but not all users.

這是當這些工具賦予使用者超能力，但不是所有使用者。

And somehow there's a selective upgrade that happens when these tools get shared in an organization.

而且，當這些工具在組織中共享時，會發生某種選擇性升級。

This, this one randomized control experiment that was run by researchers from places like Harvard and mit, engaged the Boston Consulting group and set up their junior consultants in control groups and a couple of experimental groups.

這是一個由哈佛大學和麻省理工學院等研究人員進行的隨機對照實驗，他們與波士頓諮詢集團合作，並將他們的初級顧問分為對照組和幾組實驗組。

What they did was they gave them access to a large language model and they were asked to do two kinds of tasks.

他們所做的是讓他們能夠使用大型語言模型，並要求他們執行兩種任務。

The first task was a creative ideation task.

They had to help a fictitious client come up with different product ideas that they can go to market with.

他們必須幫助一位虛構的客戶想出不同的產品點子，以便他們能夠推向市場。

The second was a business analytics task where they had to analyze why a business was struggling and create recommendations.

第二個任務是商業分析任務，他們需要分析企業為何陷入困境並提出建議。

What the study went on to discover was that when they looked at the top performers, they tended to do much better.

研究發現，當他們觀察表現最佳的員工時，他們的表現會好很多。

And when they looked at the lower performers, they tended to do much worse.

而當他們觀察表現較差的員工時，他們的表現則會差很多。

When you think of this selective upgrade effect spread across thousands of employees over a span of time, what we might see is an ever growing gap between your best and worst performers.

當你想到這種選擇性升級效應隨著時間在數千名員工中擴散時，我們可能會看到表現最好和最差的員工之間的差距不斷擴大。

And this variability would have been attributed to the use of these AI tools where that gap didn't exist before you deployed these tools.

這種差異原本並不存在，直到你部署了這些 AI 工具，而現在這種差異卻因為這些 AI 工具的使用而出現。

There are a couple of things that leaders can think about as they deploy AI in their organization.

領導者在組織中部署 AI 時，可以考慮兩件事。

The first is to create really clear guardrails around what these tools should be used for and what they shouldn't be.

第一是為這些工具的用途和禁忌設定非常明確的界線。

And those guardrails will possibly diminish over time as these tools become much more, much more effective.

隨著這些工具越來越有效，這些界線可能會隨著時間而減弱。

It's important to go through this experimental period understanding where it actually augments the work and where it actually takes away from the work.

重要的是要經歷這個實驗階段，瞭解它實際上在哪裡增強了工作，又在哪裡削弱了工作。

Another thing to consider is it's important for the users as they leverage these tools for certain domains, that they have a certain basic level of expertise in these domains.

另一件需要考慮的事情是，使用者在利用這些工具處理特定領域的工作時，在這些領域擁有一定的基本專業知識是很重要的。

It allows the users to apply good judgment when a tool is actually leading them in a worse direction and when it actually is.

這能讓使用者在工具實際引導他們走向錯誤方向時，並在工具正確引導時，都能做出良好的判斷。

Augmenting the work Using a tool when you have zero knowledge of that domain is a very, very dangerous proposal.

增強工作 在對某個領域一無所知的情況下使用工具，是非常非常危險的。

As we think about bringing AI into our organizations, we need to think about this agentic preference puzzle.

當我們考慮將 AI 導入我們的組織時，我們需要思考這個代理偏好之謎。

We as humans have a tendency towards control.

And when these tools take away control from the work, we see that adoption rates drop.

當這些工具剝奪了工作的控制權時，我們看到採用率就會下降。

There are some fascinating studies done out of Wharton that explore this idea that they called algorithmic aversion bias.

有一些來自華頓商學院的有趣研究，探討了他們稱為「演算法規避偏見」的觀念。

For example, when was the last time you decided to override what Google Maps or Waze told you was the right way home?

例如，你上次決定不聽 Google Maps 或 Waze 告訴你的回家路線是什麼時候？

We'll sometimes believe that we actually have better or have a lower error rate than these machines.

我們有時會認為，我們實際上比這些機器擁有更好的判斷力或更低的錯誤率。

And what this branch of study has looked into was when individuals actually see an algorithm commit an error, even if that error rate is still lower than the human error rate, we would much rather trust our human judgment over the algorithm.

這項研究探討的是，當個人看到演算法出現錯誤時，即使錯誤率仍低於人類的錯誤率，我們也寧願相信自己的人類判斷，而不是演算法。

It goes on to explain that perhaps one way to think about this is when we think of algorithms and these AI bots.

這進一步解釋了，或許一種思考方式是，當我們想到演算法和這些 AI 機器人時。

Their error rates are knowable and they're static, but human intuition and human intelligence is perfectable.

他們的錯誤率是可預測且固定的，但人類的直覺和智慧是可以不斷進步的。

And perhaps we therefore trust that we can perfect our own judgment in certain tasks.

也許是以我們相信自己能在某些任務中完善自己的判斷。

The research goes on to then try to figure out what's the right antidote to, to this.

這項研究接著試圖找出對此的正確解方。

And they allow, in one study, they allow users of these algorithms to tweak ever so slightly different parameters of these algorithms.

在其中一項研究中，他們允許這些演算法的使用者稍微調整演算法的參數。

When these people are given that leeway to control the algorithm, what you find is that the error rates will increase as a result, as you would expect.

當這些人被賦予控制演算法的空間時，你會發現錯誤率會如預期般增加。

But then you also see the adoption rates significantly increase because people can control it.

但你也會看到採用率顯著提高，因為人們可以控制它。

And this drives home a really valuable point around adoption of these AI tools.

這點出了關於採用這些 AI 工具一個非常有價值的重點。

As a leader, you might think about what is an error rate that is acceptable, if only it means that you then create a lot more adoption in the workplace.

作為領導者，你可能會思考，什麼樣的錯誤率是可以接受的，即使這意味著你在工作場所能創造更多採用率。

The ideal scenario is that people adopt these tools fully without tweaking them.

最理想的情況是，人們能完全採用這些工具，而不去修改它們。

But we know that that comes at a cost of lower adoption.

但我們知道，這會以降低採用率為代價。

Are we willing to sacrifice some amount of precision in the use of these tools in exchange for an improved level of adoption?

為了提高採用率，我們是否願意犧牲一些工具使用的精確度？

The final puzzle is this self sufficiency spiral.

最後一塊拼圖是這個自給自足的螺旋。

If you think about all the work we do in an organization, you can categorize them into solo work and interdependent work.

如果你思考組織裡我們做的所有工作，可以將它們分為獨立工作和相互依賴的工作。

And you might say that in the future, these tools will allow us to do a lot more solo work.

你可能會說，未來這些工具將讓我們做更多獨立工作。

And a lot of the solo work will colonize parts of the interdependent work.

而許多獨立工作將會侵蝕相互依賴工作的部分。

And then what gets left behind as interdependent work, whether it's writing emails or doing presentations or conducting meetings, a lot of these tasks will then get intermediated by these AI tools.

接著，剩下的相互依賴工作，無論是寫電子郵件、做簡報或開會，很多這些任務都會被這些 AI 工具介入。

When you think about what it takes to create culture in an organization, or the role of the leader in kind of bringing people together around a shared mission, a lot of that is about interactive tasks.

當你思考要如何建立組織文化，或是領導者在凝聚大家共同使命的角色時，很多部分都與互動任務有關。

A lot of that is about not being in solitude and doing isolated work, but

很多部分都與不孤單、不進行孤立工作有關，而是

實際上是作為一個團隊聚集在一起。

And if the future of the workplace is a lot more solo and a lot more isolated, I worry a little bit about what this means for the future of organizations and our ability to create cultures and create a sense of identity with the organization.

如果未來的工作場所變得更加孤立、更加個人化，我有點擔心這對組織的未來，以及我們建立文化和組織認同感的能力意味著什麼。

We've seen other technologies in the past kind of deliver to us a future that we didn't quite want.

過去我們看過其他技術，它們帶給我們的未來並非我們真正想要的。

You take for example, social media, where it had this promise to create a more connected world, but

以社群媒體為例，它曾承諾要創造一個更緊密的連結世界，但

instead what it gave us is possibly a more fragmented, polarized world where we perhaps have expected less from each other.

結果卻可能是一個更破碎、更兩極化的世界，我們對彼此的期望也隨之降低。

And as MIT ethnographer once said, we are alone together through these tools.

正如一位麻省理工學院的民族誌學者曾說，透過這些工具，我們是孤獨地共處。

So we don't want this feature for the workplace.

所以我們不希望工作場所出現這種情況。

And we need to think about ways that we can bring people together through perhaps different means and different approaches, so we can continue to create thriving environments for people as they engage with these tools.

我們需要思考如何透過不同的方式和方法來凝聚人們，以便在人們使用這些工具的同時，我們能持續為他們創造蓬勃發展的環境。

字幕列表影片播放

三大AI職場難題，你必須知道！| Martin Gonzalez for Big Think + (3 AI puzzles workplaces must solve | Martin Gonzalez for Big Think +)

scenario

thrive

fascinating

tend