You're Holding the AI Wrong

Shortly after the iPhone 4 launched in June of 2010, users discovered that holding the phone in their left hand caused the device to lose reception. Apple's dismissive response to its flawed antenna design has been immortalized in tech canon as "you're holding it wrong." Many felt gaslit, as if the company wasn't taking accountability for its flawed product. The saga was dubbed "Antennagate" and it serves as a helpful framing for the narrative around adopting artificial intelligence (AI).

As the AI hype cycle slides into the "Trough of Disillusionment," it's worth paying attention to how savvy sales and marketing tactics reframe the technology's value proposition. This piece will breakdown how the large language model (LLM) industry simultaneously heralds the dawn of super intelligence while blaming users for its shortcomings.

Just write better prompts

The first attempt to pin the unreliability of LLMs on users focused on prompting, despite the fact that instability is core to their statistical design. Not happy with the quality of the output? Well, you're just prompting it wrong. While there is truth to the idea that more specific prompts produce better outputs, and challenging LLMs can lead to a facsimile of critical thinking, this argument largely exists to make users believe the problem isn't with the technology but how they are using it. The magic is real, you just haven't found it... yet.

Most GenAI projects fail

As new research comes out suggesting the vast majority of generative AI projects fail, the industry needs a new story. A recent MIT study found that 95% of organizations are seeing zero return on their investment. Similarly, research from RAND states, "by some estimates, 80% of AI projects fail--twice the rate of failure for information technology projects that don't involve AI." There are other examples with similar findings, yet no major credible studies that show resounding success rates for AI initiatives.

Corporate applications of the technology remain the last hope for profitability for frontier model providers and, to a lesser extent, the countless companies that repackage and resell their products. This piece won't harp on how unlikely a meaningful return is on the monumental investment that went into the underlying technology, as the flawed economics have been dissected at length. Instead let's focus on the new tactics companies are using to continue shifting the blame of poor outcomes onto their customers. When the lack of ROI becomes impossible to ignore, the industry pivots from blaming the user’s input to blaming the user's infrastructure.

Just design better processes

As evidence mounts against the effectiveness of AI-driven automation, the defense mechanisms become more convoluted. The latest framing repackages 30-year-old corporate jargon, arguing that failure doesn't stem from the underlying technology but rather the processes your business depends on. Success with AI requires "Business Process Re-Engineering (BPR)," which demands companies restructure business processes and operations around AI agents.

BPR frames AI's failures as signs of process immaturity rather than technological flaws, with its proponents attempting to sanitize the inherent chaos of LLMs with the deterministic language of classical engineering. While it is true that LLMs mark a paradigm shift in processing unstructured data, it is hyperbolic to claim that deterministic guardrails can eliminate their shortcomings. These checks might flag an improperly structured response, but they cannot detect a factually hallucinated one that happens to follow the correct format.

Ultimately, this approach leads to a logical dead end. If we must define every possible deviation to ensure accuracy, we aren't actually "implementing AI," we are simply writing traditional, deterministic code using much more expensive and unreliable tools.

It's not me, it's you

LLMs have been marketed as a panacea, a technology on the precipice of super-intelligence unlike anything that came before it. This is why people are trying to automate anything and everything with it. If we consider that from the onset the technology has been presented as something that it's not, the real questions become:

• Would corporations be as interested in LLMs today if they hadn't been misled about their capabilities both yesterday and tomorrow?
• Similarly, would companies feel as compelled to bend LLMs to fit ill-suited automations and use cases?
• Would so many people even think it made sense to try?

Power corrupts critical thinking and rewards consensus over independent thought. We are seeing this at scale with the tomfoolery coming out of Silicon Valley. Long-winded arguments that focus on process while falling short of admitting the technology has fundamental limitations serve as mental gymnastics to rationalize a broken corporate dynamic. The primary intent is to maintain the narrative of AI superpowers while hedging against the mounting evidence that things aren't what they seem. The buyers and sellers of the technology are aligned on a need to project its success both internally and externally.

The architecture of hype

The funding mechanisms behind Silicon Valley do not prioritize truth, but instead the projection of inevitable progress. This ecosystem thrives on elevating individual visionaries while reducing the thousands of contributors to mere footnotes. This enables the industry to celebrate outliers as if they were the rule, rather than exceptions heavily dependent on luck.

The AI industry has adopted this exact playbook. It presents an idealized version of what LLMs appear to be, minimizes our collective contributions that make them possible, and tantalizes our imaginations with what they could one day become. But the very generality that makes these systems powerful means their fundamental flaws will never be completely resolved. In comparison, Antennagate was relatively straightforward. Consumers bought a new iPhone that wasn't a very good phone. Apple's initial defensiveness didn't stand up to scrutiny. The AI industry is selling human-level intelligence (and beyond). When the technology doesn't meet the expectations the industry set, outcomes are framed as a personal failure of its users and the goalposts are moved. This is harder to call out than a phone not being a good phone.

As the hype wave crests and the promises don't come to fruition, it is getting harder to blame the people trying to use the technology. If a process needs to be completely ideal for AI to work, LLMs are no different from previous automation technologies. They are merely expensive, unreliable, ROI-negative, and cannot be held in your left hand.