One Fake News Article Can Upset Your AI Decisions

More and more small and medium enterprises (SMEs) are relying on AI for inventory forecasting, automating credit assessments, and determining order quantities based on demand predictions.

So, let me ask: What happens if a single piece of fake news gets mixed into the information your AI is reading?

The answer is that “the predictions can be completely overturned.” Moreover, the cost of such an attack is around $5 a month. With the price of just one subscription, your company’s AI decisions can be compromised.

This is not a hypothetical scenario; it is backed by recent research findings.

—

74-82% of Predictions Can Be Reversed by One Fake News Article

First, let’s verify the facts.

A recent study conducted an experiment injecting a single fake news article into AI prediction systems that incorporate news articles as information sources (such as RAG-type LLM-based predictions). The results were as follows:

• 74-82% of predictions contrary to the fake news were reversed
• The probability of a fake news article being incorporated into the system is about 90%
• All it takes for the attack is to publish one SEO-optimized fake article.

In other words, as long as the AI uses a system that “gathers information from the internet to make decisions,” simply contaminating that information source allows for control over the prediction results.

What specifically happens? For instance, consider a local food wholesaler using AI for demand forecasting. If a fake news article circulates claiming “Quality issues with ingredients from XX region,” the AI will interpret this as a sudden drop in demand and reduce orders. In reality, demand remains unchanged, leading to stockouts. Conversely, if there’s fake information stating “XX is experiencing a massive boom,” it could result in over-ordering and excess inventory.

In the food industry, this directly translates to waste losses, potentially resulting in losses of hundreds of thousands or even millions of yen in an instant.

Attack Cost: $5 monthly hosting fee + time to create one article. Damage Amount: Hundreds of thousands to millions of yen. This asymmetry is the essence of the problem.

—

Why Small and Medium Enterprises Are “Prime Targets”

At this point, some may wonder, “Is there any point in targeting a small company like ours?”

Unfortunately, there is. There are three reasons:

1. Weak defenses
Large corporations have dedicated information security teams. Most SMEs do not. Many companies lack a system to verify the input data for AI predictions. From an attacker’s perspective, this is like an unlocked house.

2. Attacks do not need to be “targeted” at specific companies
Fake news affects entire industries. By disseminating one piece of false information aimed at the construction materials industry, all companies using AI that picks up that information will be impacted. This means attackers only need to target the “industry”; they don’t even need to focus on individual companies.

3. Competitors can become attackers
If a competitor can disrupt your judgments for just $5, there’s no more cost-effective sabotage for a malicious competitor. Ethically, it’s unacceptable, but technically, anyone can do it.

—

Social Engineering—Targeting AI’s “Ears” Instead of Its “Eyes”

If fake news attacks distort the AI’s “eyes” (information input), social engineering targets the AI’s “ears” (internal data).

This is a classic method, but it has taken on new significance with the proliferation of AI.

For example, an attacker could use a clever phishing email to extract login information from employees and gain access to internal systems. From there, they could tamper with customer data or transaction data that the AI references. The AI would then regard the altered data as “correct” and make erroneous credit assessments or pricing decisions.

Recent studies suggest that the detection accuracy for phishing emails has reached over 90%. Conversely, this means that 1 in 10 phishing emails can slip through detection. If an SME receives 100 emails a day, it calculates to 10 phishing emails getting through daily.

Moreover, attackers can now generate phishing emails using AI. The cost for attackers has dramatically decreased. What used to take a human hours to write one fraudulent email can now be done in minutes for hundreds of emails. While the defense costs remain unchanged, the attack costs have only decreased. This structure poses the greatest danger to SMEs.

—

Data Poisoning—Rewriting the AI’s “Memory” Itself

Another attack method to be aware of is data poisoning.

This technique involves injecting malicious data directly into the AI’s training data. While fake news distorts “current judgments,” data poisoning skews the “learned judgment criteria.” The roots run deep.

Research has shown that simply altering a few percent of the training data for relatively simple models like logistic regression can significantly decrease the model’s accuracy.

In SMEs, it is common to use past data managed in Excel or spreadsheets for AI training. If the management permissions for this data are lax, either internal personnel or external intruders can distort the AI’s judgment criteria by changing just a few lines of data.

What’s more troublesome is that data poisoning is difficult to detect. With fake news, one can later realize, “That article was false.” However, with data poisoning, the AI’s output gradually deviates, making it hard to pinpoint the cause. By the time one notices, several months of judgments may have been skewed.

—

Organizing the Cost Structure

Now, let’s summarize the cost structure of attacks and damages.

The difference between attack costs and damage amounts is 100 to 10,000 times. This is the reality.

—

So, What Should We Do?

“I understand the scary story. So, what should we do?”

This is the crux of the matter. What SMEs should do is not to mimic the security systems of large corporations, as they lack the costs and human resources. Instead, it is more realistic to create a “system that minimizes damage even when attacked.”

1. Incorporate “Human Checkpoints” into AI Decisions

Do not execute AI outputs as they are. Set alerts for order amounts that fluctuate by more than 30% compared to the previous week, and have humans verify sudden changes in credit assessments. This “anomaly filter” can be implemented at the spreadsheet function level, costing almost nothing.

2. Limit and Fix the Information Sources Referenced by AI

Instead of gathering information from an unspecified number of sources online, explicitly list reliable information sources. This could include official data from industry associations or reports from contracted research companies. By narrowing down the information sources, the risk of fake news injection is significantly reduced.

3. Manage Access Rights to Training Data

Check who has access to the past data used by the AI. Just reviewing the sharing settings of Google Sheets can be effective. If it’s set to “everyone can edit,” it should be changed immediately.

4. Monthly “Cross-Check” of AI Outputs

Cross-reference AI predictions with actual results once a month. If accuracy suddenly drops, there may be a risk of data contamination. This also requires no special tools; simply creating a comparison table with actual data will suffice.

None of these measures require expensive security tools. What’s needed is to incorporate a “system that does not blindly trust AI” into the business workflow.

—

If You Use AI, Know the Conditions That Can Break It

AI is not infallible. While it may seem obvious, many companies forget this once they implement AI.

You should not delegate business decisions to a system that can be broken for $5. AI is a tool. Knowing its weaknesses and using it accordingly will yield vastly different results than using it without awareness.

The trend of SMEs utilizing AI will not stop, nor should it. However, whether you structure your business on the premise that “AI is correct” or on the premise that “AI can make mistakes” will decisively impact your company’s resilience.

An attack cost of $5 a month. This is a number that business owners should keep in mind.