Use AI, but don’t believe the hype

The AI hype

You will have read and heard a lot about AI recently and how it will change our lives forever, not necessarily for the better - perhaps even wiping out mankind altogether. These are not the first warnings of new technologies impending doom, though. Conrad Gessner worried about information overload back in 1565 after the arrival of the printing press. The Victorians believed that trains were injuring the brain. The jarring motions of trains were alleged to unhinge the mind and either drive sane people mad or trigger violent outbursts from latent lunatics. The noise of trains was feared to shatter nerves.

But the question is if it is just a hype or if there is some truth behind it. Gartner’s hype cycle might be a good indicator. New technologies start slowly, pick up speed until expectations are inflated to a point where they cannot be fulfilled and then drop (too) quickly in popularity until they reach a realistic level. You might recall discussions (and big announcements) about the Metaverse some time ago. Not much is heard about it these days

AI-washing has not helped either. Many products are labeled as "AI-enabled" without providing tangible benefits, diluting the term's significance Inadequate oversight has further contributed to the hype, as seen with Elon Musk's Grok AI model which is not governed by rules.

So, is AI just a hype or is there justified excitement about a new technology that will transform everyone’s lives in an instant? Let’s look at some of the basics.

To use AI properly and to train the underlying Large Language Models first, sufficient data needs to be available and in good shape (or structure). A couple of Excel tables won’t cut it, no matter how much you want to use AI in your work. Very few companies will have sufficient data to train their own models, so need to acquire data elsewhere which can quickly lead to copyright infringements.

But even if enough data is available, AI won’t cure all business challenges - if you have an ill-designed sales process, the use of AI might help on the surface, but won’t solve the actual issue. On the contrary, you might accelerate your issues rather than solving them.

Even the largest LLMs struggle with reasoning (i.e. drawing conclusions from information available or with complex tasks that need to be broken down in sub-tasks). Models struggle to answer questions like “A coin is heads up. Maybelle flips the coin, Shalonda does not flip the coin. Is the coin still heads up?” or “Is it normal to find parsley in multiple sections of the grocery store (yes/no)?” without a lot of prompting.

We are just at the beginning, so can AI improve?

But surely AI will get better as we go along? One step further would be (autonomous) decision making - question is if we want machines to make autonomous decisions for us. In regulated industries, this might raise more than a few eyebrows by the regulators (or rather being prevented from being implemented altogether) for the lack of accountability and transparency. “Computer says no” is not what you want to hear if you need to access your savings or your pension has been invested at a loss.

But there is a bigger challenge - we will run out of data to train LLMs properly or might have already done so. More than half of the content available on the internet has been touched or even entirely created by AI, so AI models are trained by content created by other AIs. This can lead to a number of negative outcomes, which can lead to "model collapse," (the quality and diversity of outputs deteriorate over successive generations), e.g. through the amplification of errors, reinforcement of bias or the erosion of knowledge. An example could be automated news articles which contain inaccuracies or stylistic quirks which are used to train models, so the inaccuracies or quirks get amplified.

And while there is a lot of talk of the machines taking over, we always assume that the capacity of the human brain is somewhat limited. But what if it can expand its capacity, too, maybe with the help of AI? After all, some 86 billion neurons in the brain form 100 trillion connections to each other — numbers that, ironically, are far too large for the human brain to fathom, so we might never fully know how our brains operate. So, would it be possible to model an artificial intelligence if we do not understand existing intelligence?

Why AI can be a force for good

But let’s look at the positives. Even with these limitations, AI can be of tremendous help. Despite perception and the increase in available information, the rate of innovation is slowing down so AI can help to connect the ever-increasing number of dots. Productivity growth has slowed down by almost half in both advanced and emerging economies.

While AI has certainly been subject to hype, it is not merely a passing trend. AI is delivering real value and transforming industries, albeit more gradually than some initial predictions suggested. Tangible benefits in specific domains have been created: AI is driving advancements in fields like medical imaging, cybersecurity, industrial production, and autonomous driving.

It is still early days and AI capabilities are still evolving: AI models are rapidly improving, with larger models demonstrating emergent abilities and unexpected capabilities so diverse applications beyond generative AI are still under development. Examples are (high-quality) translations, self-learning autonomous vehicles showing capabilities beyond their initial programming or swarm intelligence, demonstrating collective behaviours that individual units could not achieve alone.  AI encompasses various technologies like predictive analytics, computer vision, and machine learning, each with practical applications. Real-world impact has already been created. Telecom Italia (TIM) implemented a Google-powered voice agent to address many customer calls, increasing efficiency by 20%. Five Sigma created an AI engine which frees up claims handlers to focus on areas where a human touch is valuable, like complex decision-making and good customer service. This has led to an 80% reduction in errors, a 25% increase in claims adjuster's productivity, and a 10% reduction in processing time.

So far, there has been minimal employment impact, but that might be because the use cases have been narrow so far. Contrary to predictions of job creation or elimination, 94.6% of AI-using businesses experienced no change in employment levels. AI could increase annual global GDP by 7% over 10 years or USD 7trn in absolute terms which would be the equivalent of adding another UK and India combined. After all, it is not AI that will take your job, but someone who is using AI faster and/or better than you.

While scenarios like Artificial General Intelligence (“the robots”) taking over makes for great headlines and some good movies, we should not be too worried about the one big apocalypse, but rather about smaller, but more probable mishaps. We should prevent these, but this requires a good understanding of the underlying technology.

It is still early days

So is AI just a hype? The answer is probably that it is too early to tell. While AI adoption has been slower than initially predicted, with only 5.4% of firms having officially rolled out Generative AI, this reflects the technology's position in the hype cycle rather than its lack of potential. As AI matures and organizations develop more realistic expectations, its true value is likely to emerge, transforming it from a hyped technology into a fundamental tool for innovation and efficiency. The challenge is to use it wisely now to profit from the tangible benefits of a new technology, but not to believe every extreme scenario. Start with researching some tools that you believe might make a difference in your organization/for your team, then dive in and play around to see what is possible. At the end of the day, only you can decide if AI is worth the hype when it comes to your interests or the goals of your organization.

If you look for further reading - “Co-intelligence” by Ethan Mollick gives a good overview about the current state of AI and how to use it best: it will not replace your job overnight, but can make it a lot easier if you have realistic expectations and take time to train the applications you are using.