Making Up for Losses with Volume: Why Large Language Model Hallucinations Don't Make Them Useless

You may have heard the joke about a business losing money on every sale. When faced with this grim reality, the optimistic partner cheerfully proclaims: “Don’t worry—we’ll make it up in volume!” The humor lies in the absurdity of the logic: scaling up a flawed system only compounds the problem. Yet, when it comes to large language models (LLMs), a similar critique—that they “hallucinate” and therefore cannot be trusted—might be overlooking a more nuanced truth. Under the right conditions, even apparent flaws can be addressed and harnessed.

What Are Hallucinations?

In the world of LLMs, hallucinations occur when a model generates text that is factually incorrect or completely fabricated, despite appearing plausible. These errors stem from the inherently statistical nature of LLMs, which predict the most likely next word or phrase based on their training data. For critics, this is a fatal flaw. If an LLM cannot reliably produce accurate information, how can it ever be considered useful?

But this criticism assumes that perfection is the only benchmark for usefulness. In reality, most creators of text—human or machine—are imperfect, yet still highly effective when used appropriately. To understand why hallucinations don’t necessarily render LLMs useless, it’s worth revisiting the business joke.

Hallucinations and "Making Up for Losses with Volume"

At first glance, scaling up a hallucinating LLM might seem like doubling down on failure. After all, if a model produces errors, generating more content doesn’t solve the problem—it amplifies it, much like selling more products at a loss. However, this perspective misses a critical difference: while scaling an unprofitable business necessarily increases deficits, scaling LLM output can be paired with methods that reduce error rates and increase reliability.

The Power of Voting and Aggregation

Imagine you ask a single person for advice on a complex problem. Their answer may or may not be accurate. Now, imagine asking ten people and taking the most common response. Assuming they’re independently reasoning, the majority opinion is more likely to be correct than any one person’s answer. This principle—known as the wisdom of the crowd—can also be applied to LLMs.

For instance, rather than relying on a single response from an LLM, we can:

1. Generate multiple responses and use voting to identify the most consistent or plausible answer.
2. Cross-check outputs against external databases or retrieval systems.
3. Combine LLMs with fact-checking tools that highlight discrepancies in their outputs.

In these scenarios, hallucinations don’t negate the model’s utility. Instead, they become noise that can be managed through redundancy and aggregation, much like how errors in individual human judgments are mitigated by collective decision-making.

The Real Measure of Usefulness

Critics often focus on what LLMs cannot do perfectly while ignoring what they can do well. Despite their imperfections, LLMs excel in:

• Drafting complex documents: Even if edits are needed, LLMs save enormous time by generating initial drafts.
• Creative problem-solving: Hallucinations are sometimes a byproduct of models “thinking outside the box,” which can inspire innovative ideas.
• Rapid summarization and synthesis: LLMs can process vast amounts of information to produce coherent summaries that would take humans much longer.

Additionally, LLMs do all of this exceedingly quickly and inexpensively, at least relative to humans.

Importantly, these strengths are not diminished by occasional hallucinations, provided users approach the outputs critically.

Learning From Other Imperfect Tools

Many tools we rely on every day are far from flawless. Spell checkers make mistakes, calculators rely on correct inputs, and even human experts occasionally get things wrong. Yet we do not dismiss these tools outright; instead, we use them in ways that maximize their strengths while mitigating their weaknesses.

LLMs are no different. While they may not yet be perfect fact-machines, they are powerful "dream engines" that synthesize and generate ideas, solutions, and insights far beyond what a single person could achieve alone in the same amount of time and at the same cost.

The Path Forward

The real question isn’t whether LLMs hallucinate but whether their outputs can be managed and improved to meet specific needs. In fact, exceedingly hallucination-free output is required in very few circumstances. Even in those cases, techniques like voting, external validation, and carefully designed workflows show that it is often possible to eliminate hallucinations to the degree required. These solutions are already proving that hallucinations are not insurmountable flaws—they are challenges to be mitigated.

Returning to the joke, the optimistic partner in the business may have been wrong about solving losses through volume. But in the world of LLMs, scaling intelligently—by leveraging redundancy, aggregation, and external validation—can turn individual weaknesses into collective strengths. Hallucinations are not the end of the road; they are part of the journey toward building more effective and transformative tools.

Conclusion

Hallucinations may make large language models imperfect, but they do not make them useless. By understanding their limitations and developing strategies to address them, we can harness LLMs’ immense potential. Far from being a flawed investment, LLMs—like all groundbreaking tools—are works in progress that already provide tremendous value when used thoughtfully. Instead of dismissing them outright, let’s embrace their potential and continue refining them – and the systems and workflows within which they are used – for the future.