decorative feature graphic to accompany post

The GenAI Paradox: billions spent, but where’s the value?

Goldman Sachs recently published a new issue of their newsletter, Top of Mind. The lead story is an interview with Daron Acemoglu, a Professor at MIT. Professor Acemoglu argues that genAI is wildly overhyped and its impact on GDP growth over the next decade will be extremely limited.

David Cahn at Sequoia published simple maths: given the capex spent on AI, the returns must be enormous (the “$600bn question”).

Meanwhile, we’ve heard little of Google’s rollout of genAI in search recently, after it embarrassingly had to deal with reports of once-trusted Google search allegedly telling its users to eat rocks or kill themselves. Many of those reports were fake, but not implausible.

Finally, I was extremely interested to see that the company with the highest revenues from genAI is not OpenAI, Anthropic, Microsoft, Amazon… but Accenture:

“Accenture’s annualized GenAI bookings, from its Q3 stats, are $3.6 billion. To put this number in context, OpenAI’s annualized revenue is $3.4 billion.”

Accenture, to be clear, is selling its services telling companies how to use AI.

All of this adds up quickly: there is an understanding that AI should be useful and it should be the future, but an awful lot of money, energy, and time is being spent on delivering a future that so far is only capable of generating occasionally accurate text.

All of this begs the question: what is going on?

What does genAI do, other than generate occasionally accurate answers?

There’s an expectation for genAI to improve, but there’s little consensus on what better looks like. In the Goldman report, Professor Acemoglu makes the point:

“What does it mean to double AI’s capabilities?”

The best benchmarks are arguably the crowdsourced assessments, but these are also precisely why Acemoglu has a problem: spending billions of dollars for a higher benchmark is not useful. This is why you need $3.6bn of spend with Accenture to tell you what to do. The output is occasionally accurate text.

I keep coming back to the line “occasionally accurate text” because it maps entirely with our experience. Ed Zillow’s excellent piece Pop Culture picks apart the Goldman report and includes this critical section:

“How does GPT – a transformer-based model that generates answers probabilistically (as in what the next part of the generation is most likely to be the correct one) based entirely on training data – do anything more than generate paragraphs of occasionally-accurate text? How do any of these models even differentiate when most of them are trained on the same training data that they’re already running out of?”

This aligns with our experience. First, I need to remind you of our credentials: Ellipsis was an extreme early adopter of AI. We first used genAI to improve existing content in 2021 – 18 months before ChatGPT’s release. I’ve worked on our next-generation FALCON AI for 3.5 years, and it was in the New York Times in March 2022, before AI was cool.

I’m reluctant to talk about my strengths, but I have a knack for identifying genAI use cases that actually work. There’s an alternative universe where Ellipsis is an AI consulting agency and we’re taking some of Accenture’s $3.6bn (and dare I say, doing a better job, too).

But: our first exploration of genAI resulted in occasionally accurate text. That is still the case 3 years later.

While genAI has undoubtedly improved, its core output remains fundamentally the same. The challenge lies not in generating text, but in producing consistently accurate and valuable content. This limitation has led to a paradoxical situation where billions are invested in technology that, at its core, still struggles with reliability.

The industry’s focus on benchmarks and capabilities often overshadows the more critical question: How can we harness genAI to create tangible value for businesses and society? We’re clearly in an AI bubble, and what happens next depends on solving this.

12–18 months until the bubble deflates

This disconnect between investment and tangible value isn’t sustainable. As the industry grapples with AI’s current limitations, signs of a potential market correction are appearing. Our client experience reflects the tension between AI’s promise and practical applications.

Clients want simultaneously more AI and less AI. There’s an understanding that there are surely efficiency gains, but also an understanding that this multi-billion dollar new technology lies to your face and refuses to back down.

Hence, we have a nuanced take on genAI. It’s good for some things and not for others. The first thing to ask when it generates text is “is this true?” because there’s a good chance it’s not.

Google’s recent troubles are of no surprise to me. A genAI summary in search outsources the task of assessing the truthfulness of the response to the user. My wife saw one of these responses on her phone for the first time recently. Her first question: how do I turn this off?

The experience of using genAI is so discordant with the hype. Google wants to upend the internet in order to serve users occasionally accurate text. Prof Acemoglu gives the AI bubble 12–18 months to live:

“My guess is that if important use cases don’t start to become more apparent in the next 12–18 months, investor enthusiasm may begin to fade.”

As the hype cycle progresses, we anticipate a shift towards more realistic expectations and targeted use cases. Companies that can identify and implement valuable AI applications will likely emerge as leaders. However, for many businesses, the challenge remains in distinguishing between AI’s promise and practical limitations.

Urgently seeking nuance

I remain bullish on AI, but I have an increasingly nuanced take. It is currently confusing to the market that Ellipsis talks about AI so much but has a team of writers. Are they just on ChatGPT all day? No, because that would be a terrible idea. AI is good at some things, bad at other things. We like using it for extremely specifically defined tasks. We do not like using it to contribute potentially accurate information to the blandification of the internet.

Our approach to AI is pragmatic and focused on delivering real value. We leverage AI’s strengths while mitigating its weaknesses, ensuring that our content remains authentic, accurate, and engaging. This balanced strategy allows us to harness the efficiency gains of AI without compromising on quality or truthfulness.

As the AI landscape evolves, we urgently need more nuance in the conversation here. It’s not “AI good/AI bad”, but about effectively and safely leveraging the technology where it can help achieve new and better things. To that end, I do agree with Professor Acemoglu: ultimately, this isn’t about replacing huge swathes of labour, but about the much more boring and incremental unlock of doing existing work better and faster. Much of that is new work that wouldn’t otherwise be done.

Hence, the impact on GDP growth is much more limited; the world is probably on the whole slightly better but not wildly so.

I got a pitch last week from a writing agency we used to work with. We stopped working with them because their writing quality was bad. They’re now offering a lower-cost “supervised AI content” service and a premium “human only content”.

This is an egregiously terrible idea. I do feel a bit sorry for the agency: they operate at the lower end of the market anyway, and they’ve likely seen a drop in demand for their services. Against this change in the market, you can see how they’re trying to roll out a reasonable adaptation.

But it’s pretty dystopian stuff: replace low-paid workers’ badly written articles with similarly bad AI-generated articles. GenAI makes it easier to create bad content. ChatGPT will confidently tell you anything you like and will do a better job than a poor-quality writer – as long as you don’t care about accuracy, truthfulness, or safety. ChatGPT spam is not progress. What’s the point?

Beyond the bubble

As we navigate the choppy waters of the GenAI hype cycle, it’s crucial to steer a course between blind optimism and undue scepticism.

We’re likely to see a market correction as reality catches up with expectations. This isn’t a failure of AI, but a recalibration of the approach. Companies that survive will be those that have developed a nuanced understanding of AI’s capabilities and limitations, integrating it thoughtfully into their processes instead of treating it as a panacea.

The key questions aren’t about AI’s capabilities, but how to harness them to create value. How can we use AI to solve real-world problems more effectively? How can we ensure AI enhances rather than replaces human judgement? And crucially, how can we maintain ethical standards and truthfulness in an age of easy-to-generate content?

The GenAI paradox isn’t a roadblock – it’s a call to action. The real AI revolution won’t be measured in benchmarks or billions spent, but in the meaningful improvements it brings to our work and lives.

Picture of Alex Denning

Alex Denning

Alex Denning is the Founder and Managing Director of Ellipsis®, a world-class SEO Content agency. Alex is the inventor of FALCON AI®, the next-generation AI SEO platform that helps you predict how your content will perform – before you create it. Alex is an international speaker and publishes regularly on the future of SEO. @AlexDenning on Twitter