The 3 biggest AI fails of 2025 - and what teams should learn from them
Generative AI could have written this intro. Odds are it would've made something up along the way. That theme defined 2025: big promises, bigger hype, and some very public stumbles.
Here are the three misses that stood out - with practical takeaways for general leaders, IT and development, and product teams who want results without the headaches.
1) Hallucinations hit academia, government, and the law
AI has been making things up for years, but 2025 amplified the problem. Google's AI Overviews stopped telling people to glue pizza, yet still managed claims like saying the latest Call of Duty doesn't exist. The pattern didn't stop there.
A Deakin University study reported ChatGPT fabricated about one in five academic citations, while many others were riddled with errors. That didn't stop institutions from relying on it: an HHS department used AI to cite studies that don't exist, a major newspaper published a summer reading list with fake book titles, and lawyers in hundreds of cases filed arguments polluted by hallucinations.
Bottom line: if a system generates language, it can generate fiction. Treat outputs as drafts, not facts.
- What to do: Require source-grounding and automated citation checks. Build retrieval-augmented workflows that link every claim to a verifiable source.
- Add friction where it counts: block submission until citations resolve, or flag unverifiable claims for human review.
- Measure truth, not vibes: track factuality with evals and sampling. If you can't measure it, you'll ship it by mistake.
- Policy matters: set rules for academic, legal, and public-facing content. No sources, no shipment.
If you're building governance from scratch, the NIST AI Risk Management Framework is a useful reference point for controls and operational guardrails. Read it here.
2) The Friend wearable failed fast
The Friend was a pendant that recorded ambient audio, shipped it to a phone app, and texted you "conversation" in real time. It spent more than $1 million on New York City subway ads across 11,000 rail cars, 1,000 platform posters, and 130 urban panels. The internet did what the internet does.
Commuters vandalized the ads, the backlash became a Halloween costume, and reviewers piled on. The lesson isn't "don't market." It's "don't mistake attention for adoption," especially when your core feature triggers privacy and social norms.
- What to do: validate desirability before scale. Small cohorts, real usage, clear consent, explicit on-device/off-device data handling.
- Design for the room: recording bystanders is a social and legal hazard. Default to opt-in, visible cues, and minimal data retention.
- Pressure-test messaging: run a PR and ethics pre-mortem before a massive campaign. If the costume version writes itself, rethink the launch.
3) Most corporate AI pilots crashed
Companies were told they "must" use AI. Many tried. Most didn't stick. An MIT Media Lab report, "The State of AI in Business 2025," estimates 95% of initiatives failed, despite $30-$40 billion in spend.
Tools like ChatGPT and Copilot improved personal productivity, but had thin impact on P&L. Enterprise systems - custom builds or vendor platforms - were often abandoned. The common reasons: brittle workflows, missing context, and poor fit with daily work.
- What to do: pick narrow, high-frequency use cases with clear payoff (support macros, finance close notes, QA triage). Prove value in weeks, not quarters.
- Ship into existing tools: put AI where work already happens (ticketing, docs, IDEs, CRM). New portals gather dust.
- Feed it context: use stable knowledge sources, permissioned retrieval, and versioned prompts. No context = no reliability.
- Operationalize quality: add evals, feedback loops, and rollback plans. Treat prompts and retrieval like product code.
- Stage gates: idea → sandbox → pilot → production, with target metrics at each step (latency, accuracy, cost per task, adoption).
- People first: train teams, set usage policies, and assign an accountable owner. Tools don't fix process by themselves.
A simple 2026 checklist
- Ground every claim with sources people can verify.
- Measure factuality and task success on real data.
- Start small, integrate where work already happens, and expand only after proof.
- Respect privacy and social norms as product requirements, not afterthoughts.
- Budget for maintenance: prompts, retrieval, evals, and change management.
If you're upskilling a team for practical AI delivery - by role and skill - explore our curated training paths: AI courses by job.
Disclosure: Ziff Davis, Mashable's parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.
Your membership also unlocks:
