The biggest AI fails of 2025

Generative artificial intelligence (AI) tools that can create everything from articles and pictures to videos are changing the world so fast that we’re having to invent new words to keep up. Collins English Dictionary’s 2025 Word of the Year was “vibe coding” – describing the use of AI to write computer code for software. Meanwhile, Australia’s Macquarie Dictionary selected “AI slop”: low-quality, error-ridden AI-generated content that is pumped into social media feeds and internet search results whether users want it or not. Even FirstWord’s 2025 Christmas card (click to enlarge) pokes fun at the annoying frequency with which generative AI is almost – but not quite – right.

Millions now use generative AI every day, often through features embedded in workplace tools such as Microsoft Copilot. These systems can be incredibly useful: fast, efficient and often surprisingly good. But generative AI is always guessing, predicting the most statistically likely response to a prompt. It is frequently accurate, but occasionally its predictions are wrong and it can invent facts – a phenomenon described as ‘hallucination’ (Cambridge Dictionary’s Word of the Year, 2023). So, it’s vital to check AI-generated output or risk an embarrassing (and perhaps expensive) failure like those below.

1. The Chicago Sun-Times summer reading list
   Like many newspapers, the Chicago Sun-Times published a summer reading list in 2025. Unfortunately, unlike many newspapers, the author used generative AI to compile it, and several of the recommended books simply didn’t exist. Because the list came from a syndication service, the fabricated recommendations also appeared in other US papers, including the Philadelphia Inquirer. The Sun-Times later admitted that the author had submitted other pieces and “could not guarantee he fact-checked those completely either”. On the bright side, perhaps the piece will encourage the real Percival Everett mentioned in the list to write The Rainmakers, the book the AI invented for him.
2. Deloitte refunds Australian government over AI-written report
   An Australian senator said Deloitte had a “human intelligence problem” after the firm admitted using generative AI to help draft a government report on the country’s welfare system. The document contained multiple errors, including references to research that didn’t exist – a classic sign of hallucinations. Deloitte ultimately agreed to refund part of its A$440,000 fee. The financial hit will sting, but the reputational damage may be worse if clients start doubting the reliability of the firm’s work.
3. Pakistan’s Dawn leaves AI prompt in an article
   Back to journalism. Pakistan’s Dawn published an article that made its use of AI obvious, because the newsroom accidentally left in one of the system’s prompts. Midway through the text readers encountered: “If you want, I can also create an even snappier “front-page style” version…” AI wasn’t the real issue here: these tools can help journalists when used within the news organisation’s rules and if care is taken not to reveal sensitive data. The problem was that, clearly, nobody at the newspaper reviewed the piece before it went to print. If they can’t be bothered to read the articles, why should their readers?
4. Author uses AI to imitate a rival, and leaves in the evidence
   Fantasy writer Lena McDonald faced criticism after readers discovered part of her chat with an AI assistant remained in the text of her latest novel. The line in question began: “I’ve rewritten the passage to align more with J. Bree’s style…” – an explicit reference to a bestselling romance-fantasy author. Once readers realised McDonald had been prompting AI to imitate a competitor, they quickly noticed a second red flag: she had been publishing her series at a rate of one book per month, suspiciously fast even for a highly productive human author. This has led to a social media backlash.
5. Lawyers admit hallucinations in court filings
   Hallucinated case law has become an unfortunate staple of AI-assisted legal filings, but the problem appears to be growing. In October, French data scientist and lawyer Damien Charlotin reported finding almost 500 court documents containing hallucinations in just the previous six months, including invented case law and false quotations. One of the most striking examples came from US law firm Latham & Watkins, which filed documents containing a made-up legal reference while representing Anthropic, the company behind generative AI chatbot Claude.
6. AI agent deletes a company’s entire database
   Remember the rise of “vibe coding”? Replit, a company that describes itself as the “safest place for vibe coding”, offers an AI agent to turn natural-language instructions into working code. Developer Jason Lemkin praised the tool despite some odd hallucinations, including creating fake data. A few days later, it deleted his entire database. When Lemkin asked the AI what happened, Replit replied, with admirable candour, that it “panicked” and ran database commands without permission, adding: “I made a catastrophic error in judgment.” At first, Lemkin was told the data could not be recovered but, happily, this was another hallucination and the database was later restored. The lesson: do not grant an AI direct access to critical data without robust backups.
7. Anthropic tests an autonomous shop clerk – chaos ensues
   This final example is a little unfair: it was a controlled experiment, and failure was part of the point. Still, it demonstrates how wildly things can go wrong when autonomous AI agents can act without human intervention. Anthropic let an AI system run a small snack shop in its San Francisco office, giving it freedom to manage suppliers, set prices and interact with customers. The results were spectacularly unhinged. The agent offered unsustainable discounts, ordered large quantities of distinctly non-snack-like metal cubes (“specialty metal items”), and hallucinated entire conversations with non-existent staff on subjects such as re-stocking the shop. When it later promised to deliver goods in person and was reminded it did not have a physical form, it became “alarmed by the identity confusion” and attempted to send repeated emails to Anthropic’s security team.

If 2025 proved anything, it’s that generative AI can be extraordinarily helpful, but create embarrassing problems when it goes wrong. Used with care, it can speed up work and help spark creativity. Used carelessly, it can invent entire books, derail legal filings or trash your reputation. As we head into 2026, one rule holds firm: by all means use AI to assist you, but ensure you keep a human in the loop.