"When people don't know what to trust, they end up trusting nothing at all, and that can deter democratic participation."
— Jean Philip De Tender, Media Director, European Broadcasting Union
In the largest study of its kind, professional journalists from 22 public broadcasters across 18 countries evaluated more than 3,000 answers from the leading AI assistants. The finding should stop every enterprise leader who has wired generative AI into a customer-facing process: AI assistants misrepresented news content 45% of the time, regardless of language or territory. Generative AI is the most transformative business technology in a generation, but its ethical risks are no longer abstract or theoretical. They are measured, costly, and showing up in courtrooms, contact centers, and balance sheets right now. Here are the seven biggest generative AI ethics risks of 2026, the verified numbers behind each one, and what responsible governance actually looks like.
1. Misinformation and Hallucination
The headline risk is that generative models produce confident, authoritative-sounding content that is simply wrong. The EBU and BBC study, published in October 2025, found that beyond the headline 45% figure, 31% of AI answers had serious sourcing problems such as missing or misleading attributions, and 20% contained major accuracy issues including hallucinated details and outdated information. One assistant, Google's Gemini, showed significant issues in 76% of its responses. This matters more every year because, per the Reuters Institute, 7% of online news consumers now use AI assistants for news, rising to 15% among those under 25.
The enterprise consequences are already concrete. Air Canada was held liable when its chatbot misrepresented the airline's bereavement-fare policy, and lawyers in multiple jurisdictions have been sanctioned for filing briefs citing court cases that the AI simply invented. Hallucination is not a quirk to be tolerated; it is a liability to be governed.
Key Takeaway
Any generative output that reaches a customer, a regulator, or a court needs a verification layer between the model and the real world. The 45% misrepresentation rate is not a reason to abandon the technology; it is the reason human review and retrieval-grounded answers are non-negotiable for high-stakes use cases.
2. Deepfakes and Synthetic Fraud
Generative models can now clone a voice or a face convincingly enough to defeat humans and many security systems. According to a Gartner survey from September 2025, 62% of organizations experienced a deepfake attack in the prior 12 months. The fraud volume is exploding: Pindrop measured a rise of more than 1,300% in deepfake fraud attempt frequency in contact centers, and in one notorious case a finance worker at the engineering firm Arup was tricked into paying out $25 million after a video call in which every other participant was a deepfake. Deloitte projects generative-AI-enabled fraud in the United States could reach between $12.3 billion and $40 billion by 2027.
The defensive gap is stark. Gartner found that only 10% of security leaders prioritize deepfake recognition in their awareness programs, even as roughly a third of CISOs report encountering deepfakes combined with social engineering on video calls.
3. Bias and Discrimination
Generative models learn from human-created data, and they faithfully reproduce, and often amplify, the biases in that data. FICO's chief analytics officer Scott Zoldi points to two mechanisms that are easy to overlook: a lack of sufficiently representative training data, and cognitive bias baked into the prompt engineering itself. The result is discriminatory outcomes that disproportionately affect minorities and marginalized groups, in exactly the high-stakes domains, hiring, lending, healthcare, where fairness matters most. This is the same governance challenge that surfaces when companies hand decisions to software, a theme we explored in our look at how predictive analytics is shifting from forecasts to autonomous decisions.
4. Data Privacy and Leakage
Large language models trained on vast scraped datasets can memorize personally identifiable information and, under the right prompt, reproduce it. The risk runs in both directions: models can leak training data, and employees can leak sensitive corporate data into third-party models. Restrictive controls often backfire by pushing staff toward unmonitored "shadow AI" tools, while the privacy controls themselves create new sensitive databases that must be secured. Privacy is not solved by a single policy; it requires continuous attention to what data flows into and out of every model.
5. Copyright and Intellectual Property
Generative tools are trained on enormous bodies of internet content whose provenance is frequently unknown and whose use was rarely consented to. That leaves enterprises exposed to intellectual-property infringement when they build on AI output, and to unresolved legal questions about who owns AI-generated work in the first place. Zoldi's warning on data provenance is blunt: model accuracy and legal safety both depend on understanding where the training data came from, and most vendors cannot fully answer that question.
Key Takeaway
Treat data provenance as a first-class procurement question. Before deploying a model, ask the vendor what it was trained on, what indemnification they offer against IP claims, and how they handle data that should never have been in the training set. If they cannot answer, that is itself the answer.
6. Accountability and the Autonomy Problem
As generative systems gain the ability to take actions, not just produce text, the question of who is responsible when something goes wrong becomes urgent. This is the natural next step from the autonomous AI agents now running production workloads across the enterprise, and it sharpens further as frontier models become capable of operating for long stretches without human supervision. When an agent makes a consequential decision, organizations need a clear chain of accountability and, as one researcher proposes, a "structured disagreement register" that documents the reasoning whenever a human and an AI diverge. Ungoverned autonomy is not a technology problem; it is a liability waiting for an owner.
7. Workforce Disruption and Environmental Cost
Two slower-burning risks round out the list. On jobs, the International Monetary Fund estimates that around 300 million jobs globally could be affected by AI, and some forecasters warn that a large share of entry-level white-collar roles could be displaced within a few years. The ethical obligation is to manage that transition deliberately rather than let it happen by accident, a question we examined in our piece on whether automation forces a rethink of how income and work are linked. On the environment, training and running large models consume enormous amounts of electricity and water, making the carbon and resource footprint of generative AI an ethical line item that belongs in any honest accounting of its costs.
How Enterprise Leaders Should Respond
The throughline across all seven risks is that generative AI's dangers are governance problems, not reasons to retreat. The organizations getting this right are not the ones that banned the technology, nor the ones that deployed it with no guardrails. They are the ones treating responsible AI as an operating discipline: verification layers on high-stakes outputs, deepfake awareness built into security training, bias testing before deployment, clear data-provenance requirements in procurement, and an explicit chain of accountability for every system allowed to act on its own. The 45% misrepresentation rate and the 62% deepfake-attack rate are not arguments against using generative AI. They are the reason using it responsibly has become a leadership requirement rather than a compliance afterthought.