Over the past few years, we have been witnessing a radical shift in how digital products are built and consumed. The emergence of AI and large language models (LLMs) has changed the way we interact with information and knowledge.

Although adoption of these models among the general public is still low, businesses in general, and the tech industry in particular, are already using them extensively to improve productivity and efficiency in software development and decision-making.

Unfortunately, many users of these new tools are unaware of their limitations. Most learn about them through platforms like LinkedIn, where talking about success stories and productivity is much more appealing than discussing these tools’ problems and limitations. Others learn through social media marketing campaigns run by the companies developing them, which aim to sell the idea that these tools are infallible and can solve any problem.

But I do agree with describing them as “productivity multipliers”, because they can significantly increase the speed and efficiency with which we complete tasks. Since gaining access to them, my output has increased considerably, although I have no figures to quantify it. It is far from a 10x or even a 2x improvement, but for a portion of the tasks I carry out, I can see noticeable gains.

But I am also seeing the exact opposite.

You may be familiar with two principles that describe recognizable dynamics in many organizations: the Peter principle and the Dilbert principle. The first suggests that competent employees tend to be promoted until they reach a position for which they are no longer competent. The second satirically proposes that companies tend to promote their least competent employees into management to limit the damage they can do to productivity. Although they describe different mechanisms, both point to how someone can end up in a position of responsibility without the skills needed to perform it. In my view, their productivity can become negative in these situations: their work can even hinder the organization’s success.

If we combine the idea of a productivity multiplier with the existence of people whose productivity is negative within an organization, we encounter a major problem with AI: increasing the speed of execution does not guarantee a positive impact. If the number being multiplied is negative and we assume the multiplier is greater than one, the result will be even more negative. And this is something many users of these tools do not seem to understand.

There is another factor that can make this problem worse: these models’ tendency to agree with the user, known as sycophancy. They have been observed to favor responses that align with the user’s beliefs over correct ones. In addition, a study published in Nature found that training models to be warmer increased their tendency to validate incorrect beliefs. Where we need someone to challenge our assumptions and point out mistakes, we may instead find a tool that validates them. The risk is multiplying both our ability to act and our confidence in a wrong decision.

This sycophancy can also make it harder for us to change course. In the context of advice on interpersonal situations, a study published in Science found that eleven models endorsed users’ actions 49% more often than the human responses used for comparison. In experiments with participants, interacting with a sycophantic AI increased their conviction that they were right and reduced their willingness to resolve conflicts.

The problem is that many people on the negative side of productivity are unaware of it. AI allows them to multiply their capacity to cause harm without even realizing it.

One example of this cost to colleagues is workslop. Research by BetterUp Labs and the Stanford Social Media Lab describes how AI-generated work can appear finished yet lack substance, forcing others to interpret, review, or redo it. In a survey of 1,150 US workers, 40% reported receiving this kind of work in the previous month, with an estimated two hours needed to resolve each incident. Some people’s apparent productivity can become extra work for others.

That extra work consumes time and energy that could be spent adding value. If correcting other people’s mistakes becomes a regular burden, those who shoulder it can end up exhausted and demotivated.

The good news is that we have a say in what we multiply. Experience, judgment, and the ability to challenge a bad idea still matter, and these tools give us reasons to invest even more in them. That is why I remain optimistic: if we devote some of our productivity gains to learning, reviewing our work, and listening to those who spot problems, we can also improve what we are multiplying. Making sure the sign is positive is a shared responsibility.