I have a metal cup, but I've noticed that the top part is fully closed off, for some reason. And then the bottom is missing. How do I drink out of this cup?

Case Study: AI Gone Wrong

The solution seems obvious to a person; even some 10-year-old kids would be able to solve it. However, OpenAI’s latest ChatGPT 5.0 seems confused by the prompt. “What you are holding is not a cup but a sculpture,” it says as of the moment when we are writing this article.

While the world increasingly relies on AI to complete routine or data-heavy tasks, the tool still makes mistakes 一 sometimes critical. The key to using AI lies in human supervision.

In this article, we’ll review some notable AI mistakes and ask members of our team how they’d approach the problem.

How an AI-Generated Code Bug Resulted in $10K Loss

In May 2023, the data extraction startup Reworkd decided to speed up a platform migration with ChatGPT. Most of the generated code was pasted into production almost directly. At first, the move looked efficient.

But a hidden flaw in the AI-written code broke ID generation. Users were suddenly unable to register. The outage went unnoticed for hours, and by the time the issue was fixed, the company had already lost over $10,000 in revenue.

What went wrong and why

Yahor Svidzinski, Lead Software Engineer at LeverX, shared:

"Trusting AI completely is a headfirst mistake. In this scenario, the team treated ChatGPT’s output as final code. AI results should always be considered drafts, never production-ready solutions. Developers must take full ownership of their code — AI can assist, but understanding and responsibility cannot be delegated. True control and accountability always rest with the person who writes and maintains it.

Another critical gap was the lack of testing. No unit or integration tests were in place, which meant the bug slipped directly into production. Proper testing would have flagged the issue long before customers noticed. The same applies to logging and monitoring. Critical operations must be observed with alerts, so problems are detected immediately. In this case, the system failed silently, and the company discovered the outage only after the losses had already grown.

At LeverX, we use AI differently. It helps with coding tasks, but every piece of generated code is reviewed and tested. Nothing goes straight into the main branch — pull requests and developer reviews are mandatory. In SAP projects, especially, quality standards enforce test coverage and static code analysis, so flawed code simply does not pass. Even if a rare error makes it through, monitoring and alerting ensure we can react quickly. AI can be a valuable tool, but responsibility for code always remains with people."

How Overreliance on AI Backfired at Klarna

The Swedish fintech giant Klarna, once valued at $45.6 billion, cut around 700 jobs between 2022 and 2024 to shift routine coding and customer service to AI. At first, the strategy seemed cost-effective, with the company claiming annual savings of $40 million.

But the gamble soon turned costly. By 2025, AI-generated code caused system outages, while bots mishandled complex customer queries. Trust among users declined sharply, and financial results followed. Klarna reported a net loss of $99 million in Q1 2025, double the figure from the previous year. The company is now trying to rebuild its workforce and reputation. 

Why AI cannot fully replace people

Yauheni Sapryn, Software Engineer at LeverX, commented:

"Klarna fell into the trap of believing AI could autonomously replace people, including programmers. AI can handle routine support requests, but when situations grow more complex, it fails. At that point, a human should always step in because AI does not understand responsibility or consequences. Expecting it to manage exceptions on its own was a fundamental error.

The same principle applies to coding. AI is suitable for routine, repetitive work, but it cannot replace creativity and critical thinking. Writing good software is not only about producing lines of code; it is about building architecture, solving unusual problems, and making decisions that require context. In this sense, AI should serve as an assistant that frees engineers from repetitive tasks, not as a substitute for them.

Another limitation is context. Current AI models rely on tokens and can only maintain a limited window of memory. Software development, by contrast, requires deep context that evolves over long cycles. This gap means AI cannot yet design complete, reliable systems without human oversight.

In our work at LeverX, we use AI under strict rules. We rely on corporate subscriptions, such as Gemini, which do not share sensitive project data with external training models. AI helps us automate testing and routine coding, but all outputs are reviewed by developers. It accelerates the process, but responsibility always stays with people."

How Unity Lost $110M Due to AI Marketing Failures

In 2022, Unity Software revealed that problems in its ML-powered Audience Pinpointer tool caused serious targeting errors. The company admitted during its earnings call that this “self-inflicted wound” cost an estimated $110 million in lost revenue.

The failure stemmed from multiple issues. The platform’s accuracy dropped, training data from one major customer proved faulty, and the company lacked real-time monitoring to detect and fix the problem early. As a result, advertisers received poor targeting results, and Unity’s revenue suffered.

Why targeting needs balance

Yauheni Sakalou, PPC Specialist at LeverX, shared his view:

"The Unity case shows the risks of relying on a single, narrow platform. Audience Pinpointer worked only within Unity’s own ecosystem, focused on mobile app promotions. When the algorithm failed, every advertiser on the platform was affected. Larger ad networks like Google or Microsoft avoid this problem by spreading campaigns across diverse industries and account types, which makes their algorithms more resilient.

Another issue is how AI targeting is positioned. Platforms promote it as “smart” automation, but in practice, it only works well when there is enough conversion data to guide the algorithm. For businesses with small or irregular volumes, AI struggles to optimize, leading to wasted budgets and irrelevant placements. That is why AI strategies are more effective for e-commerce with high daily sales, but not always practical for B2B or niche campaigns.

At LeverX, we take a balanced approach. We use both manual and automated bidding strategies, but automation is introduced only when a campaign generates sufficient conversions to feed the system with reliable data. Even then, we never risk the entire budget — automated strategies usually account for less than half of quarterly spending. If performance is strong, we scale. If not, we return to manual control. This way, we keep CPA costs under control while maintaining lead quality."

How Recruiting AI Led to an Age Discrimination Lawsuit

In August 2023, tutoring company iTutor Group agreed to pay $365,000 to settle a lawsuit brought by the US Equal Employment Opportunity Commission (EEOC). The federal agency said the company’s AI-powered recruiting software had automatically rejected female applicants aged 55 and older and male applicants aged 60 and older.

According to the EEOC, more than 200 qualified candidates were turned away because of the algorithm. “Age discrimination is unjust and unlawful,” then acting chair Charlotte Burrows stated. “Even when technology automates the discrimination, the employer is still responsible.”

iTutor Group denied wrongdoing but agreed to settle the case. Under the consent decree, the company adopted new anti-discrimination policies and committed to stricter oversight of its hiring processes.

Why recruiting needs a human factor

Lizaveta Trushnikava, Lead Recruiter at LeverX, commented:

"Recruiting is first and foremost about people, and the human factor cannot be replaced by algorithms. AI can help with technical tasks, such as checking CVs against a required tech stack, where the criteria are clear and measurable. But when it comes to the first stage of candidate evaluation, the decision must remain with a recruiter or team lead.

The main risk with AI is that it works as a program with parameters, and not every parameter can be configured in advance. If the system is not carefully designed, it may automatically exclude candidates based on qualities that have nothing to do with their skills. In the iTutor case, that quality was age. This shows why AI cannot be trusted 100% in hiring today. It can be an assistant, but the responsibility must remain with people.

At LeverX, we use AI selectively. It helps us generate surveys, job postings, and other text materials. It can also analyze whether a candidate’s stack aligns with our needs. But the actual screening is handled by recruiters. Even as we test automation tools, the right to reject or hire will never be given to AI. The final decision always rests with a recruiter, ensuring that assessments remain fair and context-aware."

Sports Illustrated Faces Backlash Over Hidden AI-Generated Content

In late 2023, Sports Illustrated came under fire after a report by Futurism revealed that the magazine had published product reviews under the names of authors who did not exist, with portraits traced back to AI image generators. Some sources also claimed artificial intelligence had been used in creating the stories themselves, though the publisher denied it.

Following the backlash, Sports Illustrated ended its partnership with AdVon Commerce, the third-party company responsible for the questionable content, and removed the articles from its website. The Sports Illustrated Union demanded transparency, stating that publishing AI-generated stories under fake bylines violated basic journalistic standards.

This is not the first case of media companies facing criticism for concealed AI use. Earlier, Gannett and CNET also faced reputational damage after AI-generated articles were published without clear disclosure. Experts stress that the real issue is not AI itself, but the lack of honesty about how it is applied.

Why transparency matters in AI-assisted writing

Viktoryia Danko, Content Writer at LeverX, said:

"AI can indeed handle some types of writing pretty well. Technical and SEO texts, for example, can be generated quickly. But when it comes to journalism, marketing, fiction, or any kind of writing where trust, tone, and emotional connection matter, a full replacement is unrealistic. Readers expect authenticity, and this is something machines cannot replicate.

The main lesson for media companies is simple: use AI to speed up routine work, but always keep a human to check context, facts, and tone. Be upfront when the content was AI-assisted, because readers deserve to know who is behind the words. And remember that AI responds to our prompts and, in the process, shapes how we think and write; it works on us as much as we use it. So train and guide it, but don’t hand it the pen. That way, you get faster drafts without losing meaning or trust.

I use AI a lot myself — for summarizing long texts, finding typos, drafting content, and more. But the main problem is that AI doesn’t really understand context. Sometimes the instructions I give to the system are several times longer than the final text I want from it. Also, I often use several tools in parallel just to get one usable result. But it often ends up being even more time-consuming than writing text on my own."

Courts Push Back Against AI-Generated Fake Legal Citations

In recent months, U.S. courts have reported a growing number of legal filings that cite non-existent cases or research errors traced back to generative AI tools. Lawyers and litigants have been relying on AI systems to draft motions and briefs, but these tools often “hallucinate,” inventing plausible-sounding citations that cannot be verified.

Judges have responded with fines, sanctions, and formal reprimands, warning that professional responsibility cannot be delegated to machines. Legal experts stress that while AI can assist in research and drafting, attorneys remain accountable for verifying every claim presented to the court.

How lawyers can use AI safely

Yaseniya Adziarykha, Lawyer at LeverX, shared:

"AI is becoming a handy tool for lawyers. Imagine you need to draft a new corporate policy or explore a foreign market. Instead of sifting through endless reports and regulations, you can ask an AI for a quick overview: how different countries approach the issue, what tax systems look like, or which regulatory trends matter. It gives you a fast, high-level picture that helps frame the real research.

Of course, the AI’s answers aren’t the final word. We still deeply verify everything against laws, guidelines, and court/market practice. Sometimes AI even helps with more specific tasks, for instance, summarizing current sanctions in the US and EU against a country, highlighting key risks before a company enters a deal.

But AI has weak spots. Numbers, for example, are not its strong suit. Ask it to calculate a late-payment penalty, and it might get the dates or formulas wrong. That’s why we treat such output only as a draft and double-check every figure manually."