AI在职场中的三大挑战：谷歌专家揭示的“选择性升级”、“自主偏好”与“自给自足螺旋”难题

I'm Martin Gonzalez, a principal of organization and leadership development at Google and the author of The Bonfire Moment.

The book explores the idea that teams are harder than tech.

In the process of innovation, it's crucial for leaders, CEOs, and founders to pay attention to the people's side of the business because neglecting it could easily derail their best-laid plans.

We know many employees and organizations are starting to use AI for their work.

We also observe a constant shift between two intense narratives: one of substitution, where jobs will disappear, roles will be replaced, and fewer people will perform certain types of work due to AI;

and another of augmentation, where these tools provide superpowers, enabling individuals to achieve more within their roles. This suggests that future success depends on adapting to these new technologies.

However, there's much to consider regarding the augmentation model, because early research indicates that as these tools are introduced into the workplace, we are not yet seeing the transformative potential of AI that its inventors have proclaimed.

Therefore, I've begun to consider three puzzles we need to solve as we integrate these technologies into our organizations.

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

This occurs when these tools endow some users with superpowers but not all, creating a selective upgrade effect within an organization.

For example, one randomized control experiment conducted by researchers from institutions like Harvard and MIT involved the Boston Consulting Group. They divided junior consultants into control and experimental groups.

The experimental groups were given access to a large language model and they were assigned two types of tasks:

first, a creative ideation task, where they had to help a fictitious client generate product ideas for market;

second, a business analytics task, requiring them to analyze a struggling business and propose recommendations.

The study found that top performers tended to do much better with the AI tools, while lower performers tended to do much worse.

If this selective upgrade effect is extrapolated across thousands of employees over time, we might observe an ever-growing gap between the best and worst performers.

This variability would be attributed to the use of AI tools, a gap that didn't exist before their deployment.

Leaders should consider a couple of things when deploying AI in their organization.

First, establish clear guardrails for what these tools should and shouldn't be used for.

These guardrails might diminish over time as the tools become more effective,

but it's crucial to use this experimental period to understand where AI truly augments work and where it detracts from it.

Second, it's important that users leveraging these tools in specific domains possess a basic level of expertise in those areas.

This enables them to exercise good judgment when a tool might be leading them in the wrong direction versus genuinely augmenting their work.

Using a tool with zero domain knowledge is a very dangerous proposition.

As we consider integrating AI into our organizations, we must address the agentic preference puzzle.

Humans inherently tend towards control, and when these tools diminish our control over work, we observe a drop in adoption rates.

Fascinating studies from Wharton explore this concept, which they termed Algorithmic Aversion Bias.

For instance, when was the last time you decided to override the directions given by Google Maps or Waze on your way home?

We sometimes believe our judgment is superior or has a lower error rate than these machines.

This field of study investigated how individuals react when they perceive an algorithm making an error; even if the algorithm's overall error rate remains lower than the human error rate, we often prefer to trust our human judgment over the algorithm.

The explanation suggests that while the error rates of algorithms and AI bots are knowable and static;

but human intuition and intelligence are perfectible. Therefore, we might trust our ability to perfect our own judgment in certain tasks.

The research then sought to find an antidote, and in one study, they allowed users of these algorithms to slightly tweak different parameters.

When people were given this leeway to control the algorithm, the error rates predictably increased.

However, this also led to a significant increase in adoption rates because people felt in control.

This highlights a valuable point regarding the adoption of AI tools:

as a leader, you might consider what level of error rate is acceptable if it means significantly higher adoption in the workplace.

The ideal scenario is full adoption without any tweaking.

But we know this often comes at the cost of lower adoption.

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

The final puzzle is the self-sufficiency spiral.

If you categorize all the work done in an organization, it falls into solo work and interdependent work.

One might argue that in the future, AI tools will enable us to perform significantly more solo work.

And much of this solo work will colonize parts of what was traditionally interdependent work.

Consequently, many remaining interdependent tasks, such as writing emails, preparing presentations, or conducting meetings, will then be intermediated by these AI tools.

When considering what it takes to foster culture within an organization or the leader's role in uniting people around a shared mission, much of this revolves around interactive tasks.

It's about coming together as a group, not working in solitude or isolation.

If the future workplace becomes significantly more solo and isolated, I am concerned about what this implies for the future of organizations and our capacity to build cultures and a sense of identity with the organization.

We've witnessed other technologies in the past deliver a future we didn't quite desire.

Take social media, for example: it promised a more connected world but instead arguably gave us a more fragmented, polarized world where we perhaps expect less from each other.

As an MIT ethnographer once said, 'we are alone together' through these tools.

Therefore, we don't want this future for the workplace, and we need to devise ways to bring people together through different means and approaches, ensuring we can continue to create thriving environments for individuals as they engage with these tools.