“Behind the Claim of ‘AI Made It Cheaper,’ Someone is Cleaning Up”

The cost of creating content with AI has dramatically decreased. Logo design, article writing, image generation — things that once cost hundreds of thousands of yen can now be produced for a few thousand yen, or in some cases, for free.

But I want to ask: Is that output usable as is?

In reality, the process of “humans correcting what AI has produced” is quietly ballooning everywhere. The costs of revisions, verification, and legal risk management are adding up. When these are combined, the premise that “it should have become cheaper with AI” is being fundamentally undermined.

This time, we will delve into the reality of AI’s “cleanup costs” through three news stories.

—

1. Freelancers’ Work Has Turned into ‘AI Correction Duty’

The story of Lisa, a freelance graphic designer living in Spain, is emblematic.

Since the release of ChatGPT in 2022, the nature of her assignments has changed drastically. Previously, she focused on creating logos and package designs from scratch. Now, 90% of her orders involve tweaking AI-generated content. The proportion of revision work in her annual income is 60-70%.

At first glance, one might think, “If there is work, that’s good, isn’t it?” But the reality is different.

• Clients believe that they are just asking for “a little correction of what AI created,” so the unit price is low.
• When she actually opens the files, she finds images of people with six fingers, logos that completely disregard brand guidelines, and the use of materials that are legally gray in terms of copyright.
• In many cases, rather than “just a little correction,” it would be faster to recreate it from scratch.
• However, the compensation remains at the rate for “corrections.”

Ultimately, Lisa has started to refuse such projects. She couldn’t handle them mentally or physically.

Now, consider this: Are the jobs Lisa turned down gone? They are not. Someone else is taking them on at the same low rates. In other words, the “cleanup of AI” is structurally being offloaded onto freelancers.

A Lesson for Small and Medium Enterprises

If local small and medium enterprises are creating designs and content with AI and thinking, “We got it done cheaply,” I want to ask: Who is checking the quality of that output?

If there are no personnel in-house to verify it, they will ultimately outsource it to freelancers. How does the total cost, including the cost of revisions, compare to hiring a professional from the start? In some cases, it may even be reversed.

—

2. OpenAI Faces 30 Lawsuits — The Shock of the Tumbler Ridge Incident

A shooting incident in Tumbler Ridge, British Columbia, Canada, has led to OpenAI being hit with 30 lawsuits.

The crux of the lawsuits is the possibility that ChatGPT’s output influenced the actions of the perpetrator. The lawyers for the victims argue that “AI-generated content can adversely affect human decision-making and behavior,” holding the platform provider accountable.

What the conclusion of these lawsuits will be is still unknown. However, there is another point that small business owners should pay attention to.

A precedent is being set where ‘lawsuits arise due to AI output.’

For instance, imagine the following scenario:

• Your company implements an AI chatbot for customer support.
• The bot provides incorrect product information.
• A customer believes that information and makes a purchase, leading to an incident.
• The customer sues, saying, “Your AI provided false information.”

This is not just a hypothetical situation. There have already been cases overseas where an airline’s AI chatbot provided incorrect refund policies, resulting in the airline losing a lawsuit (Air Canada, 2024).

Few small and medium enterprises have in-house legal counsel. The costs associated with handling lawsuits can start at a minimum of 500,000 to 1,000,000 yen just for attorney fees. If the trial drags on, it could reach several million yen. For a company with sales in the tens of millions of yen, this could be fatal.

—

3. AI Detection Tools Are Unreliable — The Blind Spot of ‘Verification Costs’

AI detection tools are gaining attention as a means to verify AI output. Tools like “Pangram” are considered the gold standard, but what is the reality?

To put it bluntly, current AI detection tools are not perfect.

There are three specific issues:

1. False Positives (Human-written content judged as AI-generated): There have been reports of texts written by non-native English speakers or those who write formulaic content being mistakenly identified as “AI-generated.”
2. False Negatives (Failing to detect AI-generated content): With clever prompts and the use of paraphrasing tools, detection can be easily bypassed.
3. Detecting Hallucinations is a Separate Issue: AI detection tools determine whether content was written by AI but do not assess whether the content is accurate.

The third issue is particularly troublesome. In the Australian Parliament, there have been reports of factual inaccuracies in AI-generated materials influencing policy discussions. AI outputs “plausible lies” that slip past detection tools and are used in decision-making. This risk exists regardless of the size of the company.

The Reality for Small and Medium Enterprises

The notion that “having AI detection tools is sufficient” is dangerous. In addition to the costs of implementing and operating detection tools, the conclusion is that a final check by a human eye is indispensable.

In other words, new human costs for verification are added to the processes that were supposed to be streamlined by AI. This is one of the true identities of the “cleanup costs of AI.”

—

Estimating the ‘Cleanup Costs of AI’ — The Reality for Small and Medium Enterprises

Based on the three news stories, let’s summarize the “cleanup costs” when small and medium enterprises incorporate AI into their operations.

In total, this amounts to 1,200,000 – 4,500,000 yen annually.

Even if the subscription fees for AI tools are a few thousand to tens of thousands of yen per month, the cleanup costs can be several times or even tens of times higher.

The claim that “AI is cheap” only considers the output costs. When viewed from the total cost perspective, a completely different picture emerges.

—

So, What Should We Do?

I’m not saying, “Don’t use AI.” AI is undoubtedly a powerful tool. The issue lies in the design of its usage.

When small and medium enterprises adopt AI, there are three key points to keep in mind.

① Decide ‘What Not to Let AI Do’ First

Delegating customer interactions, legal documents, and information related to health and safety to AI is far too risky. Clearly delineate the areas where AI outputs can be released as is and those where human intervention is mandatory.

② Estimate Cleanup Costs ‘Before Implementation’

Do not make decisions based solely on the monthly costs of AI tools. Estimate the total costs including revisions, verification, and risk management, and compare them to hiring a professional from the start. If the costs are reversed, it may be more rational not to use AI.

③ Regularly Measure ‘AI Output Quality’

Once you start using AI outputs, keep track of revision rates, return rates, and complaint rates. Without managing these numbers, the cleanup costs will continue to grow unnoticed.

—

Conclusion — What Happens After Costs Decrease

AI has dramatically lowered the “cost of creation.” However, this has led to the emergence of new costs for “correction,” “verification,” and “accountability.”

Moreover, these costs are often structured to be offloaded onto weaker players like freelancers and small and medium enterprises.

Large corporations have legal departments, quality control teams, and the capacity to absorb risks. Small and medium enterprises do not have these resources. Therefore, the idea of ‘designing the cleanup before using AI’ is essential.

The costs of implementing AI will continue to decrease. However, cleanup costs will not automatically decrease. Misjudging this can lead to the outcome of “thinking you started cheaply but ending up paying a high price.”

If you are going to use AI, design it from the exit point.